1
1
openmpi/ompi/mca/btl/openib/btl_openib_endpoint.c

925 строки
32 KiB
C
Исходник Обычный вид История

/* -*- Mode: C; c-basic-offset:4 ; -*- */
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2007 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
* Copyright (c) 2006-2007 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2006-2007 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2006-2007 Voltaire All rights reserved.
* Copyright (c) 2006-2007 Mellanox Technologies, Inc. All rights reserved.
*
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include <sys/time.h>
#include <time.h>
#include <errno.h>
#include <string.h>
#include "orte/mca/oob/base/base.h"
#include "orte/mca/rml/rml.h"
#include "orte/mca/errmgr/errmgr.h"
#include "ompi/types.h"
#include "ompi/mca/pml/base/pml_base_sendreq.h"
#include "ompi/class/ompi_free_list.h"
#include "btl_openib_endpoint.h"
#include "btl_openib_proc.h"
#include "btl_openib_xrc.h"
#include "btl_openib_async.h"
#include "connect/connect.h"
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
static void mca_btl_openib_endpoint_construct(mca_btl_base_endpoint_t* endpoint);
static void mca_btl_openib_endpoint_destruct(mca_btl_base_endpoint_t* endpoint);
static int post_send(mca_btl_openib_endpoint_t *ep,
mca_btl_openib_send_frag_t *frag, const bool rdma)
{
mca_btl_openib_module_t *openib_btl = ep->endpoint_btl;
mca_btl_base_segment_t *seg = &to_base_frag(frag)->segment;
struct ibv_sge *sg = &to_com_frag(frag)->sg_entry;
struct ibv_send_wr *sr_desc = &to_out_frag(frag)->sr_desc;
struct ibv_send_wr *bad_wr;
int qp = to_base_frag(frag)->base.order;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
sg->length = seg->seg_len + sizeof(mca_btl_openib_header_t) +
(rdma ? sizeof(mca_btl_openib_footer_t) : 0) + frag->coalesced_length;
sr_desc->send_flags = ib_send_flags(sg->length, ep);
if(ep->nbo)
BTL_OPENIB_HEADER_HTON(*frag->hdr);
if(rdma) {
int32_t head;
mca_btl_openib_footer_t* ftr =
(mca_btl_openib_footer_t*)(((char*)frag->hdr) + sg->length -
sizeof(mca_btl_openib_footer_t));
sr_desc->opcode = IBV_WR_RDMA_WRITE;
MCA_BTL_OPENIB_RDMA_FRAG_SET_SIZE(ftr, sg->length);
MCA_BTL_OPENIB_RDMA_MAKE_LOCAL(ftr);
#if OMPI_ENABLE_DEBUG
ftr->seq = ep->eager_rdma_remote.seq++;
#endif
if(ep->nbo)
BTL_OPENIB_FOOTER_HTON(*ftr);
sr_desc->wr.rdma.rkey = ep->eager_rdma_remote.rkey;
MCA_BTL_OPENIB_RDMA_MOVE_INDEX(ep->eager_rdma_remote.head, head);
sr_desc->wr.rdma.remote_addr =
ep->eager_rdma_remote.base.lval +
head * openib_btl->eager_rdma_frag_size +
sizeof(mca_btl_openib_header_t) +
mca_btl_openib_component.eager_limit +
sizeof(mca_btl_openib_footer_t);
sr_desc->wr.rdma.remote_addr -= sg->length;
} else {
if(BTL_OPENIB_QP_TYPE_PP(qp)) {
sr_desc->opcode = IBV_WR_SEND;
} else {
sr_desc->opcode = IBV_WR_SEND_WITH_IMM;
sr_desc->imm_data = ep->rem_info.rem_index;
}
}
#if HAVE_XRC
if(BTL_OPENIB_QP_TYPE_XRC(qp))
sr_desc->xrc_remote_srq_num = ep->rem_info.rem_srqs[qp].rem_srq_num;
#endif
assert(sg->addr == (uint64_t)(uintptr_t)frag->hdr);
return ibv_post_send(ep->qps[qp].qp->lcl_qp, sr_desc, &bad_wr);
}
static inline int acruire_wqe(mca_btl_openib_endpoint_t *ep,
mca_btl_openib_send_frag_t *frag)
{
int qp = to_base_frag(frag)->base.order;
int prio = !(to_base_frag(frag)->base.des_flags & MCA_BTL_DES_FLAGS_PRIORITY);
if(qp_get_wqe(ep, qp) < 0) {
qp_put_wqe(ep, qp);
OPAL_THREAD_LOCK(&ep->qps[qp].qp->lock);
opal_list_append(&ep->qps[qp].qp->pending_frags[prio],
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
(opal_list_item_t *)frag);
OPAL_THREAD_UNLOCK(&ep->qps[qp].qp->lock);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
return OMPI_ERR_OUT_OF_RESOURCE;
}
return OMPI_SUCCESS;
}
static inline int
acquire_eager_rdma_send_credit(mca_btl_openib_endpoint_t *endpoint)
{
if(OPAL_THREAD_ADD32(&endpoint->eager_rdma_remote.tokens, -1) < 0) {
OPAL_THREAD_ADD32(&endpoint->eager_rdma_remote.tokens, 1);
return OMPI_ERR_OUT_OF_RESOURCE;
}
return OMPI_SUCCESS;
}
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
static int acquire_send_credit(mca_btl_openib_endpoint_t *endpoint,
mca_btl_openib_send_frag_t *frag)
{
mca_btl_openib_module_t *openib_btl = endpoint->endpoint_btl;
int qp = to_base_frag(frag)->base.order;
int prio = !(to_base_frag(frag)->base.des_flags & MCA_BTL_DES_FLAGS_PRIORITY);
if(BTL_OPENIB_QP_TYPE_PP(qp)) {
if(OPAL_THREAD_ADD32(&endpoint->qps[qp].u.pp_qp.sd_credits, -1) < 0) {
OPAL_THREAD_ADD32(&endpoint->qps[qp].u.pp_qp.sd_credits, 1);
opal_list_append(&endpoint->qps[qp].pending_frags[prio],
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
(opal_list_item_t *)frag);
return OMPI_ERR_OUT_OF_RESOURCE;
}
} else {
if(OPAL_THREAD_ADD32(&openib_btl->qps[qp].u.srq_qp.sd_credits, -1) < 0)
{
OPAL_THREAD_ADD32(&openib_btl->qps[qp].u.srq_qp.sd_credits, 1);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
OPAL_THREAD_LOCK(&openib_btl->ib_lock);
opal_list_append(&openib_btl->qps[qp].u.srq_qp.pending_frags[prio],
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
(opal_list_item_t *)frag);
OPAL_THREAD_UNLOCK(&openib_btl->ib_lock);
return OMPI_ERR_OUT_OF_RESOURCE;
}
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
}
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
return OMPI_SUCCESS;
}
#define GET_CREDITS(FROM, TO) \
do { \
TO = FROM; \
} while(0 == OPAL_ATOMIC_CMPSET_32(&FROM, TO, 0))
/* this function is called with endpoint->endpoint_lock held */
int mca_btl_openib_endpoint_post_send(mca_btl_openib_endpoint_t *endpoint,
mca_btl_openib_send_frag_t *frag)
{
mca_btl_openib_header_t *hdr = frag->hdr;
mca_btl_base_descriptor_t *des = &to_base_frag(frag)->base;
int qp, ib_rc;
int32_t cm_return;
bool do_rdma = false;
size_t eager_limit;
if(OPAL_LIKELY(des->order == MCA_BTL_NO_ORDER))
des->order = frag->qp_idx;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
qp = des->order;
if(acruire_wqe(endpoint, frag) != OMPI_SUCCESS)
return OMPI_ERR_RESOURCE_BUSY;
eager_limit = mca_btl_openib_component.eager_limit +
sizeof(mca_btl_openib_header_coalesced_t) +
sizeof(mca_btl_openib_control_header_t);
if(des->des_src->seg_len + frag->coalesced_length <= eager_limit &&
(des->des_flags & MCA_BTL_DES_FLAGS_PRIORITY)) {
/* High priority frag. Try to send over eager RDMA */
if(acquire_eager_rdma_send_credit(endpoint) == OMPI_SUCCESS)
do_rdma = true;
}
if(!do_rdma && acquire_send_credit(endpoint, frag) != OMPI_SUCCESS) {
qp_put_wqe(endpoint, qp);
return OMPI_ERR_RESOURCE_BUSY;
}
GET_CREDITS(endpoint->eager_rdma_local.credits, hdr->credits);
if(hdr->credits)
hdr->credits |= BTL_OPENIB_RDMA_CREDITS_FLAG;
if(!do_rdma) {
if(BTL_OPENIB_QP_TYPE_PP(qp) && 0 == hdr->credits) {
GET_CREDITS(endpoint->qps[qp].u.pp_qp.rd_credits, hdr->credits);
}
} else {
hdr->credits |= (qp << 11);
}
GET_CREDITS(endpoint->qps[qp].u.pp_qp.cm_return, cm_return);
/* cm_seen is only 8 bytes, but cm_return is 32 bytes */
if(cm_return > 255) {
hdr->cm_seen = 255;
cm_return -= 255;
OPAL_THREAD_ADD32(&endpoint->qps[qp].u.pp_qp.cm_return, cm_return);
} else {
hdr->cm_seen = cm_return;
}
ib_rc = post_send(endpoint, frag, do_rdma);
if(!ib_rc)
return OMPI_SUCCESS;
if(endpoint->nbo)
BTL_OPENIB_HEADER_NTOH(*hdr);
if(BTL_OPENIB_IS_RDMA_CREDITS(hdr->credits)) {
OPAL_THREAD_ADD32(&endpoint->eager_rdma_local.credits,
BTL_OPENIB_CREDITS(hdr->credits));
}
qp_put_wqe(endpoint, qp);
if(do_rdma) {
OPAL_THREAD_ADD32(&endpoint->eager_rdma_remote.tokens, 1);
} else {
if(BTL_OPENIB_QP_TYPE_PP(qp)) {
OPAL_THREAD_ADD32(&endpoint->qps[qp].u.pp_qp.rd_credits,
hdr->credits);
OPAL_THREAD_ADD32(&endpoint->qps[qp].u.pp_qp.sd_credits, 1);
} else if BTL_OPENIB_QP_TYPE_SRQ(qp){
mca_btl_openib_module_t *openib_btl = endpoint->endpoint_btl;
OPAL_THREAD_ADD32(&openib_btl->qps[qp].u.srq_qp.sd_credits, 1);
}
}
BTL_ERROR(("error posting send request error %d: %s\n",
ib_rc, strerror(ib_rc)));
return OMPI_ERROR;
}
OBJ_CLASS_INSTANCE(mca_btl_openib_endpoint_t,
opal_object_t, mca_btl_openib_endpoint_construct,
mca_btl_openib_endpoint_destruct);
/*
* Initialize state of the endpoint instance.
*
*/
static mca_btl_openib_qp_t *endpoint_alloc_qp(void)
{
mca_btl_openib_qp_t *qp = calloc(1, sizeof(mca_btl_openib_qp_t));
if(!qp) {
BTL_ERROR(("Failed to allocate memory for qp"));
return NULL;
}
OBJ_CONSTRUCT(&qp->pending_frags[0], opal_list_t);
OBJ_CONSTRUCT(&qp->pending_frags[1], opal_list_t);
OBJ_CONSTRUCT(&qp->lock, opal_mutex_t);
return qp;
}
static void
endpoint_init_qp_pp(mca_btl_openib_endpoint_qp_t *ep_qp, const int qp)
{
mca_btl_openib_qp_info_t *qp_info = &mca_btl_openib_component.qp_infos[qp];
ep_qp->qp = endpoint_alloc_qp();
ep_qp->qp->users++;
/* local credits are set here such that on initial posting
* of the receive buffers we end up with zero credits to return
* to our peer. The peer initializes his sd_credits to reflect this
* below. Note that this may be a problem for iWARP as the sender
* now has credits even if the receive buffers are not yet posted
*/
ep_qp->u.pp_qp.rd_credits = -qp_info->rd_num;
ep_qp->u.pp_qp.rd_posted = 0;
ep_qp->u.pp_qp.cm_sent = 0;
ep_qp->u.pp_qp.cm_return = -qp_info->u.pp_qp.rd_rsv;
ep_qp->u.pp_qp.cm_received = qp_info->u.pp_qp.rd_rsv;
/* initialize the local view of credits */
ep_qp->u.pp_qp.sd_credits = qp_info->rd_num;
/* number of available send WQEs */
ep_qp->qp->sd_wqe = qp_info->rd_num;
}
static void
endpoint_init_qp_srq(mca_btl_openib_endpoint_qp_t *ep_qp, const int qp)
{
ep_qp->qp = endpoint_alloc_qp();
ep_qp->qp->users++;
/* number of available send WQEs */
ep_qp->qp->sd_wqe = mca_btl_openib_component.qp_infos[qp].u.srq_qp.sd_max;
}
static void
endpoint_init_qp_xrc(mca_btl_base_endpoint_t *ep, const int qp)
{
int max = ep->endpoint_btl->hca->ib_dev_attr.max_qp_wr -
(mca_btl_openib_component.use_eager_rdma ?
mca_btl_openib_component.max_eager_rdma : 0);
mca_btl_openib_endpoint_qp_t *ep_qp = &ep->qps[qp];
ep_qp->qp = ep->ib_addr->qp;
ep_qp->qp->sd_wqe += mca_btl_openib_component.qp_infos[qp].u.srq_qp.sd_max;
/* make sure that we don't overrun maximum supported by hca */
if (ep_qp->qp->sd_wqe > max)
ep_qp->qp->sd_wqe = max;
ep_qp->qp->users++;
}
static void endpoint_init_qp(mca_btl_base_endpoint_t *ep, const int qp)
{
mca_btl_openib_endpoint_qp_t *ep_qp = &ep->qps[qp];
ep_qp->rd_credit_send_lock = 0;
ep_qp->credit_frag = NULL;
OBJ_CONSTRUCT(&ep_qp->pending_frags[0], opal_list_t);
OBJ_CONSTRUCT(&ep_qp->pending_frags[1], opal_list_t);
switch(BTL_OPENIB_QP_TYPE(qp)) {
case MCA_BTL_OPENIB_PP_QP:
endpoint_init_qp_pp(ep_qp, qp);
break;
case MCA_BTL_OPENIB_SRQ_QP:
endpoint_init_qp_srq(ep_qp, qp);
break;
case MCA_BTL_OPENIB_XRC_QP:
if (NULL == ep->ib_addr->qp) {
ep->ib_addr->qp = endpoint_alloc_qp();
}
endpoint_init_qp_xrc(ep, qp);
break;
default:
BTL_ERROR(("Wrong QP type"));
break;
}
}
void mca_btl_openib_endpoint_init(mca_btl_openib_module_t *btl,
mca_btl_base_endpoint_t *ep,
ompi_btl_openib_connect_base_module_t *local_cpc,
mca_btl_openib_proc_modex_t *remote_proc_info,
ompi_btl_openib_connect_base_module_data_t *remote_cpc_data)
{
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
int qp;
ep->endpoint_btl = btl;
ep->use_eager_rdma = btl->hca->use_eager_rdma &
mca_btl_openib_component.use_eager_rdma;
ep->subnet_id = btl->port_info.subnet_id;
ep->endpoint_local_cpc = local_cpc;
ep->endpoint_remote_cpc_data = remote_cpc_data;
ep->rem_info.rem_lid = remote_proc_info->pm_port_info.lid;
ep->rem_info.rem_subnet_id = remote_proc_info->pm_port_info.subnet_id;
ep->rem_info.rem_mtu = remote_proc_info->pm_port_info.mtu;
This commit represents a bunch of work on a Mercurial side branch. As such, the commit message back to the master SVN repository is fairly long. = ORTE Job-Level Output Messages = Add two new interfaces that should be used for all new code throughout the ORTE and OMPI layers (we already make the search-and-replace on the existing ORTE / OMPI layers): * orte_output(): (and corresponding friends ORTE_OUTPUT, orte_output_verbose, etc.) This function sends the output directly to the HNP for processing as part of a job-specific output channel. It supports all the same outputs as opal_output() (syslog, file, stdout, stderr), but for stdout/stderr, the output is sent to the HNP for processing and output. More on this below. * orte_show_help(): This function is a drop-in-replacement for opal_show_help(), with two differences in functionality: 1. the rendered text help message output is sent to the HNP for display (rather than outputting directly into the process' stderr stream) 1. the HNP detects duplicate help messages and does not display them (so that you don't see the same error message N times, once from each of your N MPI processes); instead, it counts "new" instances of the help message and displays a message every ~5 seconds when there are new ones ("I got X new copies of the help message...") opal_show_help and opal_output still exist, but they only output in the current process. The intent for the new orte_* functions is that they can apply job-level intelligence to the output. As such, we recommend that all new ORTE and OMPI code use the new orte_* functions, not thei opal_* functions. === New code === For ORTE and OMPI programmers, here's what you need to do differently in new code: * Do not include opal/util/show_help.h or opal/util/output.h. Instead, include orte/util/output.h (this one header file has declarations for both the orte_output() series of functions and orte_show_help()). * Effectively s/opal_output/orte_output/gi throughout your code. Note that orte_output_open() takes a slightly different argument list (as a way to pass data to the filtering stream -- see below), so you if explicitly call opal_output_open(), you'll need to slightly adapt to the new signature of orte_output_open(). * Literally s/opal_show_help/orte_show_help/. The function signature is identical. === Notes === * orte_output'ing to stream 0 will do similar to what opal_output'ing did, so leaving a hard-coded "0" as the first argument is safe. * For systems that do not use ORTE's RML or the HNP, the effect of orte_output_* and orte_show_help will be identical to their opal counterparts (the additional information passed to orte_output_open() will be lost!). Indeed, the orte_* functions simply become trivial wrappers to their opal_* counterparts. Note that we have not tested this; the code is simple but it is quite possible that we mucked something up. = Filter Framework = Messages sent view the new orte_* functions described above and messages output via the IOF on the HNP will now optionally be passed through a new "filter" framework before being output to stdout/stderr. The "filter" OPAL MCA framework is intended to allow preprocessing to messages before they are sent to their final destinations. The first component that was written in the filter framework was to create an XML stream, segregating all the messages into different XML tags, etc. This will allow 3rd party tools to read the stdout/stderr from the HNP and be able to know exactly what each text message is (e.g., a help message, another OMPI infrastructure message, stdout from the user process, stderr from the user process, etc.). Filtering is not active by default. Filter components must be specifically requested, such as: {{{ $ mpirun --mca filter xml ... }}} There can only be one filter component active. = New MCA Parameters = The new functionality described above introduces two new MCA parameters: * '''orte_base_help_aggregate''': Defaults to 1 (true), meaning that help messages will be aggregated, as described above. If set to 0, all help messages will be displayed, even if they are duplicates (i.e., the original behavior). * '''orte_base_show_output_recursions''': An MCA parameter to help debug one of the known issues, described below. It is likely that this MCA parameter will disappear before v1.3 final. = Known Issues = * The XML filter component is not complete. The current output from this component is preliminary and not real XML. A bit more work needs to be done to configure.m4 search for an appropriate XML library/link it in/use it at run time. * There are possible recursion loops in the orte_output() and orte_show_help() functions -- e.g., if RML send calls orte_output() or orte_show_help(). We have some ideas how to fix these, but figured that it was ok to commit before feature freeze with known issues. The code currently contains sub-optimal workarounds so that this will not be a problem, but it would be good to actually solve the problem rather than have hackish workarounds before v1.3 final. This commit was SVN r18434.
2008-05-14 00:00:55 +04:00
orte_output(-1, "Got remote LID, subnet, MTU: %d, %lx, %d",
ep->rem_info.rem_lid,
ep->rem_info.rem_subnet_id,
ep->rem_info.rem_mtu);
for (qp = 0; qp < mca_btl_openib_component.num_qps; qp++) {
endpoint_init_qp(ep, qp);
}
}
static void mca_btl_openib_endpoint_construct(mca_btl_base_endpoint_t* endpoint)
{
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
/* setup qp structures */
endpoint->qps = (mca_btl_openib_endpoint_qp_t*)
calloc(mca_btl_openib_component.num_qps,
sizeof(mca_btl_openib_endpoint_qp_t));
if (MCA_BTL_XRC_ENABLED) {
endpoint->rem_info.rem_qps = (mca_btl_openib_rem_qp_info_t*)
calloc(1, sizeof(mca_btl_openib_rem_qp_info_t));
endpoint->rem_info.rem_srqs = (mca_btl_openib_rem_srq_info_t*)
calloc(mca_btl_openib_component.num_xrc_qps,
sizeof(mca_btl_openib_rem_srq_info_t));
} else {
endpoint->rem_info.rem_qps = (mca_btl_openib_rem_qp_info_t*)
calloc(mca_btl_openib_component.num_qps,
sizeof(mca_btl_openib_rem_qp_info_t));
endpoint->rem_info.rem_srqs = NULL;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
}
endpoint->ib_addr = NULL;
endpoint->xrc_recv_qp_num = 0;
endpoint->endpoint_btl = 0;
endpoint->endpoint_proc = 0;
endpoint->endpoint_local_cpc = NULL;
endpoint->endpoint_remote_cpc_data = NULL;
endpoint->endpoint_tstamp = 0.0;
endpoint->endpoint_state = MCA_BTL_IB_CLOSED;
endpoint->endpoint_retries = 0;
OBJ_CONSTRUCT(&endpoint->endpoint_lock, opal_mutex_t);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
OBJ_CONSTRUCT(&endpoint->pending_lazy_frags, opal_list_t);
OBJ_CONSTRUCT(&endpoint->pending_get_frags, opal_list_t);
OBJ_CONSTRUCT(&endpoint->pending_put_frags, opal_list_t);
endpoint->get_tokens = mca_btl_openib_component.ib_qp_ous_rd_atom;
/* initialize RDMA eager related parts */
endpoint->eager_recv_count = 0;
memset(&endpoint->eager_rdma_remote, 0,
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
sizeof(mca_btl_openib_eager_rdma_remote_t));
memset(&endpoint->eager_rdma_local, 0,
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
sizeof(mca_btl_openib_eager_rdma_local_t));
OBJ_CONSTRUCT(&endpoint->eager_rdma_local.lock, opal_mutex_t);
endpoint->rem_info.rem_lid = 0;
endpoint->rem_info.rem_subnet_id = 0;
Bring over all the work from the /tmp/ib-hw-detect branch. In addition to my design and testing, it was conceptually approved by Gil, Gleb, Pasha, Brad, and Galen. Functionally [probably somewhat lightly] tested by Galen. We may still have to shake out some bugs during the next few months, but it seems to be working for all the cases that I can throw at it. Here's a summary of the changes from that branch: * Move MCA parameter registration to a new file (btl_openib_mca.c): * Properly check the retun status of registering MCA params * Check for valid values of MCA parameters * Make help strings better * Otherwise, the only default value of an MCA param that was changed was max_btls; it went from 4 to -1 (meaning: use all available) * Properly prototyped internal functions in _component.c * Made a bunch of functions static that didn't need to be public * Renamed to remove "mca_" prefix from static functions * Call new MCA param registration function * Call new INI file read/lookup/finalize functions * Updated a bunch of macros to be "BTL_" instead of "ORTE_" * Be a little more consistent with return values * Handle -1 for the max_btls MCA param * Fixed a free() that should have been an OBJ_RELEASE() * Some re-indenting * Added INI-file parsing * New flex file: btl_openib_ini.l * New default HCA params .ini file (probably to be expanded over time by other HCA vendors) * Added more show_help messages for parsing problems * Read in INI files and cache the values for later lookup * When component opens an HCA, lookup to see if any corresponding values were found in the INI files (ID'ed by the HCA vendor_id and vendor_part_id) * Added btl_openib_verbose MCA param that shows what the INI-file stuff does (e.g., shows which MTU your HCA ends up using) * Added btl_openib_hca_param_files as a colon-delimited list of INI files to check for values during startup (in order, left-to-right, just like the MCA base directory param). * MTU is currently the only value supported in this framework. * It is not a fatal error if we don't find params for the HCA in the INI file(s). Instead, just print a warning. New MCA param btl_openib_warn_no_hca_params_found can be used to disable printing the warning. * Add MTU to peer negotiation when making a connection * Exchange maximum MTU; select the lesser of the two This commit was SVN r11182.
2006-08-14 23:30:37 +04:00
endpoint->rem_info.rem_mtu = 0;
endpoint->nbo = false;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
endpoint->use_eager_rdma = false;
endpoint->eager_rdma_remote.tokens = 0;
endpoint->eager_rdma_local.credits = 0;
}
/*
* Destroy a endpoint
*
*/
static void mca_btl_openib_endpoint_destruct(mca_btl_base_endpoint_t* endpoint)
{
bool pval_clean = false;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
int qp;
/* If the CPC has an endpoint_finalize function, call it */
if (NULL != endpoint->endpoint_local_cpc->cbm_endpoint_finalize) {
endpoint->endpoint_local_cpc->cbm_endpoint_finalize(endpoint);
}
/* Release memory resources */
do {
/* Make sure that mca_btl_openib_endpoint_connect_eager_rdma ()
* was not in "connect" or "bad" flow (failed to allocate memory)
* and changed the pointer back to NULL
*/
if(!opal_atomic_cmpset_ptr(&endpoint->eager_rdma_local.base.pval, NULL,
(void*)1)) {
if ((void*)1 != endpoint->eager_rdma_local.base.pval &&
NULL != endpoint->eager_rdma_local.base.pval) {
endpoint->endpoint_btl->super.btl_mpool->mpool_free(endpoint->endpoint_btl->super.btl_mpool,
endpoint->eager_rdma_local.base.pval,
(mca_mpool_base_registration_t*)endpoint->eager_rdma_local.reg);
pval_clean=true;
}
} else {
pval_clean=true;
}
} while (!pval_clean);
/* Close opened QPs if we have them*/
for(qp = 0; qp < mca_btl_openib_component.num_qps; qp++) {
MCA_BTL_OPENIB_CLEAN_PENDING_FRAGS(&endpoint->qps[qp].pending_frags[0]);
MCA_BTL_OPENIB_CLEAN_PENDING_FRAGS(&endpoint->qps[qp].pending_frags[1]);
OBJ_DESTRUCT(&endpoint->qps[qp].pending_frags[0]);
OBJ_DESTRUCT(&endpoint->qps[qp].pending_frags[1]);
if(--endpoint->qps[qp].qp->users != 0)
continue;
MCA_BTL_OPENIB_CLEAN_PENDING_FRAGS(
&endpoint->qps[qp].qp->pending_frags[0]);
MCA_BTL_OPENIB_CLEAN_PENDING_FRAGS(
&endpoint->qps[qp].qp->pending_frags[1]);
OBJ_DESTRUCT(&endpoint->qps[qp].qp->pending_frags[0]);
OBJ_DESTRUCT(&endpoint->qps[qp].qp->pending_frags[1]);
if(endpoint->qps[qp].qp->lcl_qp != NULL)
if(ibv_destroy_qp(endpoint->qps[qp].qp->lcl_qp))
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
BTL_ERROR(("Failed to destroy QP:%d\n", qp));
free(endpoint->qps[qp].qp);
}
/* free the qps */
free(endpoint->qps);
endpoint->qps = NULL;
/* unregister xrc recv qp */
#if HAVE_XRC
if (0 != endpoint->xrc_recv_qp_num) {
if(ibv_unreg_xrc_rcv_qp(endpoint->endpoint_btl->hca->xrc_domain,
endpoint->xrc_recv_qp_num)) {
BTL_ERROR(("Failed to unregister XRC recv QP:%d\n", endpoint->xrc_recv_qp_num));
}
}
#endif
OBJ_DESTRUCT(&endpoint->endpoint_lock);
/* Clean pending lists */
MCA_BTL_OPENIB_CLEAN_PENDING_FRAGS(&endpoint->pending_lazy_frags);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
OBJ_DESTRUCT(&endpoint->pending_lazy_frags);
MCA_BTL_OPENIB_CLEAN_PENDING_FRAGS(&endpoint->pending_get_frags);
OBJ_DESTRUCT(&endpoint->pending_get_frags);
MCA_BTL_OPENIB_CLEAN_PENDING_FRAGS(&endpoint->pending_put_frags);
OBJ_DESTRUCT(&endpoint->pending_put_frags);
}
/*
* call when the connect module has created all the qp's on an
* endpoint and needs to have some receive buffers posted
*/
int mca_btl_openib_endpoint_post_recvs(mca_btl_openib_endpoint_t *endpoint)
{
int qp;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
for (qp = 0; qp < mca_btl_openib_component.num_qps; ++qp) {
if (BTL_OPENIB_QP_TYPE_PP(qp)) {
mca_btl_openib_endpoint_post_rr_nolock(endpoint, qp);
} else {
mca_btl_openib_post_srr(endpoint->endpoint_btl, qp);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
}
}
return OMPI_SUCCESS;
}
/*
* called when the connect module has completed setup of an endpoint
*/
void mca_btl_openib_endpoint_connected(mca_btl_openib_endpoint_t *endpoint)
{
opal_list_item_t *frag_item, *ep_item;
mca_btl_openib_send_frag_t *frag;
mca_btl_openib_endpoint_t *ep;
bool master = false;
if (MCA_BTL_XRC_ENABLED) {
OPAL_THREAD_LOCK(&endpoint->ib_addr->addr_lock);
if (MCA_BTL_IB_ADDR_CONNECTED == endpoint->ib_addr->status) {
/* We are not xrc master */
/* set our qp pointer to master qp */
master = false;
} else {
/* I'm master of XRC */
endpoint->ib_addr->status = MCA_BTL_IB_ADDR_CONNECTED;
master = true;
}
}
/* Run over all qps and load alternative path */
#if OMPI_HAVE_THREAD_SUPPORT
if (APM_ENABLED) {
int i;
if (MCA_BTL_XRC_ENABLED) {
if (master) {
mca_btl_openib_load_apm(endpoint->ib_addr->qp->lcl_qp, endpoint);
}
} else {
for(i = 0; i < mca_btl_openib_component.num_qps; i++) {
mca_btl_openib_load_apm(endpoint->qps[i].qp->lcl_qp, endpoint);
}
}
}
#endif
endpoint->endpoint_state = MCA_BTL_IB_CONNECTED;
endpoint->endpoint_btl->hca->non_eager_rdma_endpoints++;
/* The connection is correctly setup. Now we can decrease the
event trigger. */
opal_progress_event_users_decrement();
/* While there are frags in the list, process them */
while (!opal_list_is_empty(&(endpoint->pending_lazy_frags))) {
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
frag_item = opal_list_remove_first(&(endpoint->pending_lazy_frags));
frag = to_send_frag(frag_item);
/* We need to post this one */
if(OMPI_ERROR == mca_btl_openib_endpoint_post_send(endpoint, frag))
BTL_ERROR(("Error posting send"));
}
/* if upper layer called put or get before connection moved to connected
* state then we restart them here */
mca_btl_openib_frag_progress_pending_put_get(endpoint,
mca_btl_openib_component.rdma_qp);
if(MCA_BTL_XRC_ENABLED) {
while(master && !opal_list_is_empty(&endpoint->ib_addr->pending_ep)) {
ep_item = opal_list_remove_first(&endpoint->ib_addr->pending_ep);
ep = (mca_btl_openib_endpoint_t *)ep_item;
if (OMPI_SUCCESS !=
endpoint->endpoint_local_cpc->cbm_start_connect(endpoint->endpoint_local_cpc,
ep)) {
BTL_ERROR(("Failed to connect pending endpoint\n"));
}
}
OPAL_THREAD_UNLOCK(&endpoint->ib_addr->addr_lock);
}
}
/*
* Attempt to send a fragment using a given endpoint. If the endpoint is not
* connected, queue the fragment and start the connection as required.
*/
int mca_btl_openib_endpoint_send(mca_btl_base_endpoint_t* ep,
mca_btl_openib_send_frag_t* frag)
{
int rc;
OPAL_THREAD_LOCK(&ep->endpoint_lock);
rc = check_endpoint_state(ep, &to_base_frag(frag)->base,
&ep->pending_lazy_frags);
if(OPAL_LIKELY(rc == OMPI_SUCCESS))
rc = mca_btl_openib_endpoint_post_send(ep, frag);
OPAL_THREAD_UNLOCK(&ep->endpoint_lock);
return rc;
}
/**
* Return control fragment.
*/
static void mca_btl_openib_endpoint_credits(
mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* des,
int status)
{
int qp;
mca_btl_openib_send_control_frag_t *frag = to_send_control_frag(des);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
qp = frag->qp_idx;
/* we don't acquire a WQE for credit message - so decrement.
* Note: doing it for QP used for credit management */
qp_get_wqe(ep, des->order);
if(check_send_credits(ep, qp) || check_eager_rdma_credits(ep))
mca_btl_openib_endpoint_send_credits(ep, qp);
else {
BTL_OPENIB_CREDITS_SEND_UNLOCK(ep, qp);
/* check one more time if credits are available after unlock */
send_credits(ep, qp);
}
}
/**
* Return credits to peer
*/
void mca_btl_openib_endpoint_send_credits(mca_btl_openib_endpoint_t* endpoint,
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
const int qp)
{
mca_btl_openib_module_t* openib_btl = endpoint->endpoint_btl;
mca_btl_openib_send_control_frag_t* frag;
mca_btl_openib_rdma_credits_header_t *credits_hdr;
int rc;
bool do_rdma = false;
int32_t cm_return;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
frag = endpoint->qps[qp].credit_frag;
if(OPAL_UNLIKELY(NULL == frag)) {
frag = alloc_credit_frag(openib_btl);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
frag->qp_idx = qp;
endpoint->qps[qp].credit_frag = frag;
/* set those once and forever */
to_base_frag(frag)->base.order = mca_btl_openib_component.credits_qp;
to_base_frag(frag)->base.des_cbfunc = mca_btl_openib_endpoint_credits;
to_base_frag(frag)->base.des_cbdata = NULL;
to_com_frag(frag)->endpoint = endpoint;
frag->hdr->tag = MCA_BTL_TAG_BTL;
to_base_frag(frag)->segment.seg_len =
sizeof(mca_btl_openib_rdma_credits_header_t);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
}
assert(frag->qp_idx == qp);
credits_hdr = (mca_btl_openib_rdma_credits_header_t*)
to_base_frag(frag)->segment.seg_addr.pval;
if(acquire_eager_rdma_send_credit(endpoint) == MPI_SUCCESS) {
do_rdma = true;
} else {
if(OPAL_THREAD_ADD32(&endpoint->qps[qp].u.pp_qp.cm_sent, 1) >
(mca_btl_openib_component.qp_infos[qp].u.pp_qp.rd_rsv - 1)) {
OPAL_THREAD_ADD32(&endpoint->qps[qp].u.pp_qp.cm_sent, -1);
BTL_OPENIB_CREDITS_SEND_UNLOCK(endpoint, qp);
return;
}
}
GET_CREDITS(endpoint->qps[qp].u.pp_qp.rd_credits, frag->hdr->credits);
frag->hdr->cm_seen = 0;
GET_CREDITS(endpoint->qps[qp].u.pp_qp.cm_return, cm_return);
if(cm_return > 255) {
frag->hdr->cm_seen = 255;
cm_return -= 255;
OPAL_THREAD_ADD32(&endpoint->qps[qp].u.pp_qp.cm_return, cm_return);
} else {
frag->hdr->cm_seen = cm_return;
}
GET_CREDITS(endpoint->eager_rdma_local.credits, credits_hdr->rdma_credits);
credits_hdr->qpn = qp;
credits_hdr->control.type = MCA_BTL_OPENIB_CONTROL_CREDITS;
if(endpoint->nbo)
BTL_OPENIB_RDMA_CREDITS_HEADER_HTON(*credits_hdr);
if((rc = post_send(endpoint, frag, do_rdma)) == 0)
return;
if(endpoint->nbo) {
BTL_OPENIB_HEADER_NTOH(*frag->hdr);
BTL_OPENIB_RDMA_CREDITS_HEADER_NTOH(*credits_hdr);
}
BTL_OPENIB_CREDITS_SEND_UNLOCK(endpoint, qp);
OPAL_THREAD_ADD32(&endpoint->qps[qp].u.pp_qp.rd_credits,
frag->hdr->credits);
OPAL_THREAD_ADD32(&endpoint->eager_rdma_local.credits,
credits_hdr->rdma_credits);
if(do_rdma)
OPAL_THREAD_ADD32(&endpoint->eager_rdma_remote.tokens, 1);
else
OPAL_THREAD_ADD32(&endpoint->qps[qp].u.pp_qp.cm_sent, -1);
BTL_ERROR(("error posting send request errno %d says %s", rc,
strerror(errno)));
}
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
/* local callback function for completion of eager rdma connect */
static void mca_btl_openib_endpoint_eager_rdma_connect_cb(
mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* endpoint,
struct mca_btl_base_descriptor_t* descriptor,
int status)
{
mca_btl_openib_hca_t *hca = endpoint->endpoint_btl->hca;
OPAL_THREAD_ADD32(&hca->non_eager_rdma_endpoints, -1);
assert(hca->non_eager_rdma_endpoints >= 0);
MCA_BTL_IB_FRAG_RETURN(descriptor);
}
/* send the eager rdma connect message to the remote endpoint */
static int mca_btl_openib_endpoint_send_eager_rdma(
mca_btl_base_endpoint_t* endpoint)
{
mca_btl_openib_module_t* openib_btl = endpoint->endpoint_btl;
mca_btl_openib_eager_rdma_header_t *rdma_hdr;
mca_btl_openib_send_control_frag_t* frag;
int rc;
frag = alloc_credit_frag(openib_btl);
if(NULL == frag) {
return -1;
}
to_base_frag(frag)->base.des_cbfunc =
mca_btl_openib_endpoint_eager_rdma_connect_cb;
to_base_frag(frag)->base.des_cbdata = NULL;
to_base_frag(frag)->base.des_flags |= MCA_BTL_DES_FLAGS_PRIORITY;
to_base_frag(frag)->base.order = mca_btl_openib_component.credits_qp;
to_base_frag(frag)->segment.seg_len =
sizeof(mca_btl_openib_eager_rdma_header_t);
to_com_frag(frag)->endpoint = endpoint;
frag->hdr->tag = MCA_BTL_TAG_BTL;
rdma_hdr = (mca_btl_openib_eager_rdma_header_t*)to_base_frag(frag)->segment.seg_addr.pval;
rdma_hdr->control.type = MCA_BTL_OPENIB_CONTROL_RDMA;
rdma_hdr->rkey = endpoint->eager_rdma_local.reg->mr->rkey;
rdma_hdr->rdma_start.lval = ompi_ptr_ptol(endpoint->eager_rdma_local.base.pval);
BTL_VERBOSE(("sending rkey %lu, rdma_start.lval %llu, pval %p, ival %u type %d and sizeof(rdma_hdr) %d\n",
rdma_hdr->rkey,
rdma_hdr->rdma_start.lval,
rdma_hdr->rdma_start.pval,
rdma_hdr->rdma_start.ival,
rdma_hdr->control.type,
sizeof(mca_btl_openib_eager_rdma_header_t)
));
if(endpoint->nbo) {
BTL_OPENIB_EAGER_RDMA_CONTROL_HEADER_HTON((*rdma_hdr));
BTL_VERBOSE(("after HTON: sending rkey %lu, rdma_start.lval %llu, pval %p, ival %u\n",
rdma_hdr->rkey,
rdma_hdr->rdma_start.lval,
rdma_hdr->rdma_start.pval,
rdma_hdr->rdma_start.ival
));
}
rc = mca_btl_openib_endpoint_send(endpoint, frag);
if (OMPI_SUCCESS == rc ||OMPI_ERR_RESOURCE_BUSY == rc)
return OMPI_SUCCESS;
MCA_BTL_IB_FRAG_RETURN(frag);
BTL_ERROR(("Error sending RDMA buffer", strerror(errno)));
return rc;
}
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
/* Setup eager RDMA buffers and notify the remote endpoint*/
void mca_btl_openib_endpoint_connect_eager_rdma(
mca_btl_openib_endpoint_t* endpoint)
{
mca_btl_openib_module_t* openib_btl = endpoint->endpoint_btl;
char *buf;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
mca_btl_openib_recv_frag_t *headers_buf;
int i;
/* Set local rdma pointer to 1 temporarily so other threads will not try
* to enter the function */
if(!opal_atomic_cmpset_ptr(&endpoint->eager_rdma_local.base.pval, NULL,
(void*)1))
return;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
headers_buf = (mca_btl_openib_recv_frag_t*)
malloc(sizeof(mca_btl_openib_recv_frag_t) *
mca_btl_openib_component.eager_rdma_num);
if(NULL == headers_buf)
goto unlock_rdma_local;
buf = openib_btl->super.btl_mpool->mpool_alloc(openib_btl->super.btl_mpool,
openib_btl->eager_rdma_frag_size *
mca_btl_openib_component.eager_rdma_num,
mca_btl_openib_component.buffer_alignment,
MCA_MPOOL_FLAGS_CACHE_BYPASS,
(mca_mpool_base_registration_t**)&endpoint->eager_rdma_local.reg);
if(!buf)
goto free_headers_buf;
buf = buf + openib_btl->eager_rdma_frag_size -
sizeof(mca_btl_openib_footer_t) - openib_btl->super.btl_eager_limit -
sizeof(mca_btl_openib_header_t);
for(i = 0; i < mca_btl_openib_component.eager_rdma_num; i++) {
ompi_free_list_item_t *item;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
mca_btl_openib_recv_frag_t * frag;
mca_btl_openib_frag_init_data_t init_data;
item = (ompi_free_list_item_t*)&headers_buf[i];
item->registration = (void*)endpoint->eager_rdma_local.reg;
item->ptr = buf + i * openib_btl->eager_rdma_frag_size;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
OBJ_CONSTRUCT(item, mca_btl_openib_recv_frag_t);
init_data.order = MCA_BTL_NO_ORDER;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
init_data.list = NULL;
mca_btl_openib_frag_init(item, &init_data);
frag = to_recv_frag(item);
to_base_frag(frag)->type = MCA_BTL_OPENIB_FRAG_EAGER_RDMA;
to_com_frag(frag)->endpoint = endpoint;
frag->ftr = (mca_btl_openib_footer_t*)
((char*)to_base_frag(frag)->segment.seg_addr.pval +
mca_btl_openib_component.eager_limit);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
MCA_BTL_OPENIB_RDMA_MAKE_REMOTE(frag->ftr);
}
endpoint->eager_rdma_local.frags = headers_buf;
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
endpoint->eager_rdma_local.rd_win =
mca_btl_openib_component.eager_rdma_num >> 2;
endpoint->eager_rdma_local.rd_win =
endpoint->eager_rdma_local.rd_win?endpoint->eager_rdma_local.rd_win:1;
/* set local rdma pointer to real value */
opal_atomic_cmpset_ptr(&endpoint->eager_rdma_local.base.pval, (void*)1,
buf);
if(mca_btl_openib_endpoint_send_eager_rdma(endpoint) == OMPI_SUCCESS) {
mca_btl_openib_hca_t *hca = endpoint->endpoint_btl->hca;
mca_btl_openib_endpoint_t **p;
OBJ_RETAIN(endpoint);
This commit brings in two major things: 1. Galen's fine-grain control of queue pair resources in the openib BTL. 1. Pasha's new implementation of asychronous HCA event handling. Pasha's new implementation doesn't take much explanation, but the new "multifrag" stuff does. Note that "svn merge" was not used to bring this new code from the /tmp/ib_multifrag branch -- something Bad happened in the periodic trunk pulls on that branch making an actual merge back to the trunk effectively impossible (i.e., lots and lots of arbitrary conflicts and artifical changes). :-( == Fine-grain control of queue pair resources == Galen's fine-grain control of queue pair resources to the OpenIB BTL (thanks to Gleb for fixing broken code and providing additional functionality, Pasha for finding broken code, and Jeff for doing all the svn work and regression testing). Prior to this commit, the OpenIB BTL created two queue pairs: one for eager size fragments and one for max send size fragments. When the use of the shared receive queue (SRQ) was specified (via "-mca btl_openib_use_srq 1"), these QPs would use a shared receive queue for receive buffers instead of the default per-peer (PP) receive queues and buffers. One consequence of this design is that receive buffer utilization (the size of the data received as a percentage of the receive buffer used for the data) was quite poor for a number of applications. The new design allows multiple QPs to be specified at runtime. Each QP can be setup to use PP or SRQ receive buffers as well as giving fine-grained control over receive buffer size, number of receive buffers to post, when to replenish the receive queue (low water mark) and for SRQ QPs, the number of outstanding sends can also be specified. The following is an example of the syntax to describe QPs to the OpenIB BTL using the new MCA parameter btl_openib_receive_queues: {{{ -mca btl_openib_receive_queues \ "P,128,16,4;S,1024,256,128,32;S,4096,256,128,32;S,65536,256,128,32" }}} Each QP description is delimited by ";" (semicolon) with individual fields of the QP description delimited by "," (comma). The above example therefore describes 4 QPs. The first QP is: P,128,16,4 Meaning: per-peer receive buffer QPs are indicated by a starting field of "P"; the first QP (shown above) is therefore a per-peer based QP. The second field indicates the size of the receive buffer in bytes (128 bytes). The third field indicates the number of receive buffers to allocate to the QP (16). The fourth field indicates the low watermark for receive buffers at which time the BTL will repost receive buffers to the QP (4). The second QP is: S,1024,256,128,32 Shared receive queue based QPs are indicated by a starting field of "S"; the second QP (shown above) is therefore a shared receive queue based QP. The second, third and fourth fields are the same as in the per-peer based QP. The fifth field is the number of outstanding sends that are allowed at a given time on the QP (32). This provides a "good enough" mechanism of flow control for some regular communication patterns. QPs MUST be specified in ascending receive buffer size order. This requirement may be removed prior to 1.3 release. This commit was SVN r15474.
2007-07-18 05:15:59 +04:00
assert(((opal_object_t*)endpoint)->obj_reference_count == 2);
do {
p = &hca->eager_rdma_buffers[hca->eager_rdma_buffers_count];
} while(!opal_atomic_cmpset_ptr(p, NULL, endpoint));
OPAL_THREAD_ADD32(&openib_btl->eager_rdma_channels, 1);
/* from this point progress function starts to poll new buffer */
OPAL_THREAD_ADD32(&hca->eager_rdma_buffers_count, 1);
return;
}
openib_btl->super.btl_mpool->mpool_free(openib_btl->super.btl_mpool,
buf, (mca_mpool_base_registration_t*)endpoint->eager_rdma_local.reg);
free_headers_buf:
free(headers_buf);
unlock_rdma_local:
/* set local rdma pointer back to zero. Will retry later */
opal_atomic_cmpset_ptr(&endpoint->eager_rdma_local.base.pval,
endpoint->eager_rdma_local.base.pval, NULL);
endpoint->eager_rdma_local.frags = NULL;
}