1
1
openmpi/opal/mca/btl/ugni/btl_ugni_smsg.c
Nathan Hjelm d5cdeb81d0 btl/ugni: improve multi-threaded performance
This commit updates the ugni btl to make use of multiple device
contexts to improve the multi-threaded RMA performance. This commit
contains the following:

 - Cleanup the endpoint structure by removing unnecessary field. The
   structure now also contains all the fields originally handled by the
   common/ugni endpoint.

 - Clean up the fragment allocation code to remove the need to
   initialize the my_list member of the fragment structure. This
   member is not initialized by the free list initializer function.

 - Remove the (now unused) common/ugni component. btl/ugni no longer
   need the component. common/ugni was originally split out of
   btl/ugni to support bcol/ugni. As that component exists there is no
   reason to keep this component.

 - Create wrappers for the ugni functionality required by
   btl/ugni. This was done to ease supporting multiple device
   contexts. The wrappers are thread safe and currently use a spin
   lock instead of a mutex. This produces better performance when
   using multiple threads spread over multiple cores. In the future
   this lock may be replaced by another serialization mechanism. The
   wrappers are located in a new file: btl_ugni_device.h.

 - Remove unnecessary device locking from serial parts of the ugni
   btl. This includes the first add-procs and module finalize.

 - Clean up fragment wait list code by moving enqueue into common
   function.

 - Expose the communication domain flags as an MCA variable. The
   defaults have been updated to reflect the recommended setting for
   knl and haswell.

 - Avoid allocating fragments for communication with already
   overloaded peers.

 - Allocate RDMA endpoints dyncamically. This is needed to support
   spreading RMA operations accross multiple contexts.

 - Add support for spreading RMA communication over multiple ugni
   device contexts. This should greatly improve the threading
   performance when communicating with multiple peers. By default the
   number of virtual devices depends on 1) whether
   opal_using_threads() is set, 2) how many local processes are in the
   job, and 3) how many bits are available in the pid. The last is
   used to ensure that each CDM is created with a unique id.

Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2017-03-13 14:46:06 -06:00

243 строки
8.0 KiB
C

/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2011-2017 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2011 UT-Battelle, LLC. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "btl_ugni_smsg.h"
#include "btl_ugni_rdma.h"
static void mca_btl_ugni_smsg_mbox_construct (mca_btl_ugni_smsg_mbox_t *mbox) {
struct mca_btl_ugni_reg_t *ugni_reg =
(struct mca_btl_ugni_reg_t *) mbox->super.registration;
mca_rcache_base_registration_t *base_reg =
(mca_rcache_base_registration_t *) ugni_reg;
/* initialize mailbox attributes */
mbox->attr.smsg_attr.msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
mbox->attr.smsg_attr.msg_maxsize = mca_btl_ugni_component.ugni_smsg_limit;
mbox->attr.smsg_attr.mbox_maxcredit = mca_btl_ugni_component.smsg_max_credits;
mbox->attr.smsg_attr.mbox_offset = (uintptr_t) mbox->super.ptr - (uintptr_t) base_reg->base;
mbox->attr.smsg_attr.msg_buffer = base_reg->base;
mbox->attr.smsg_attr.buff_size = mca_btl_ugni_component.smsg_mbox_size;
mbox->attr.smsg_attr.mem_hndl = ugni_reg->handle.gni_handle;
mbox->attr.proc_name = OPAL_PROC_MY_NAME;
mbox->attr.rmt_irq_mem_hndl = mca_btl_ugni_component.modules[0].devices[0].smsg_irq_mhndl;
}
OBJ_CLASS_INSTANCE(mca_btl_ugni_smsg_mbox_t, opal_free_list_item_t,
mca_btl_ugni_smsg_mbox_construct, NULL);
int mca_btl_ugni_smsg_init (mca_btl_ugni_module_t *ugni_module)
{
gni_return_t rc;
for (int i = 0 ; i < mca_btl_ugni_component.virtual_device_count ; ++i) {
rc = GNI_SmsgSetMaxRetrans (ugni_module->devices[i].dev_handle,
mca_btl_ugni_component.smsg_max_retries);
if (GNI_RC_SUCCESS != rc) {
BTL_ERROR(("error setting maximum SMSG retries %s",gni_err_str[rc]));
return mca_btl_rc_ugni_to_opal (rc);
}
}
return OPAL_SUCCESS;
}
/* progress */
int mca_btl_ugni_smsg_process (mca_btl_base_endpoint_t *ep)
{
mca_btl_ugni_module_t *ugni_module = mca_btl_ugni_ep_btl (ep);
mca_btl_active_message_callback_t *reg;
mca_btl_ugni_base_frag_t frag;
mca_btl_base_segment_t seg;
bool disconnect = false;
uintptr_t data_ptr;
gni_return_t rc;
uint32_t len;
int count = 0;
if (!opal_atomic_cmpset_32 (&ep->smsg_progressing, 0, 1)) {
/* already progressing (we can't support reentry here) */
return 0;
}
/* per uGNI documentation we loop until the mailbox is empty */
do {
uint8_t tag = GNI_SMSG_ANY_TAG;
rc = mca_btl_ugni_smsg_get_next_wtag (ep->smsg_ep_handle, &data_ptr, &tag);
if (GNI_RC_SUCCESS != rc) {
if (OPAL_LIKELY(GNI_RC_NOT_DONE == rc)) {
BTL_VERBOSE(("no smsg message waiting. rc = %s", gni_err_str[rc]));
ep->smsg_progressing = 0;
return count;
}
BTL_ERROR(("unhandled GNI_SmsgGetNextWTag error"));
return OPAL_ERROR;
}
assert (0 != data_ptr);
count++;
BTL_VERBOSE(("got smsg fragment. tag = %d\n", tag));
switch (tag) {
case MCA_BTL_UGNI_TAG_SEND:
frag.hdr.send = ((mca_btl_ugni_send_frag_hdr_t *) data_ptr)[0];
tag = frag.hdr.send.lag >> 24;
len = frag.hdr.send.lag & 0x00ffffff;
BTL_VERBOSE(("received smsg fragment. hdr = {len = %u, tag = %d}", len, tag));
reg = mca_btl_base_active_message_trigger + tag;
frag.base.des_segments = &seg;
frag.base.des_segment_count = 1;
seg.seg_addr.pval = (void *)((uintptr_t)data_ptr + sizeof (mca_btl_ugni_send_frag_hdr_t));
seg.seg_len = len;
assert (NULL != reg->cbfunc);
reg->cbfunc(&ugni_module->super, tag, &(frag.base), reg->cbdata);
break;
case MCA_BTL_UGNI_TAG_GET_INIT:
frag.hdr.eager_ex = ((mca_btl_ugni_eager_ex_frag_hdr_t *) data_ptr)[0];
mca_btl_ugni_start_eager_get (ep, frag.hdr.eager_ex, NULL);
break;
case MCA_BTL_UGNI_TAG_RDMA_COMPLETE:
frag.hdr.rdma = ((mca_btl_ugni_rdma_frag_hdr_t *) data_ptr)[0];
if (((mca_btl_ugni_base_frag_t *)frag.hdr.rdma.ctx)->flags & MCA_BTL_UGNI_FRAG_SMSG_COMPLETE) {
mca_btl_ugni_frag_complete (frag.hdr.rdma.ctx, OPAL_SUCCESS);
} else {
/* let the local smsg completion finish this frag */
((mca_btl_ugni_base_frag_t *)frag.hdr.rdma.ctx)->flags &= ~MCA_BTL_UGNI_FRAG_IGNORE;
}
break;
case MCA_BTL_UGNI_TAG_DISCONNECT:
/* remote endpoint has disconnected */
disconnect = true;
break;
default:
BTL_ERROR(("unknown tag %d\n", tag));
break;
}
rc = mca_btl_ugni_smsg_release (ep->smsg_ep_handle);
if (OPAL_UNLIKELY(GNI_RC_SUCCESS != rc)) {
BTL_ERROR(("Smsg release failed! rc = %d", rc));
return OPAL_ERROR;
}
} while (!disconnect);
ep->smsg_progressing = 0;
/* disconnect if we get here */
opal_mutex_lock (&ep->lock);
mca_btl_ugni_ep_disconnect (ep, false);
opal_mutex_unlock (&ep->lock);
return count;
}
static inline int
mca_btl_ugni_handle_remote_smsg_overrun (mca_btl_ugni_module_t *btl)
{
size_t endpoint_count;
unsigned int ep_index;
int count, rc;
BTL_VERBOSE(("btl/ugni_component detected SMSG CQ overrun. "
"processing message backlog..."));
/* we don't know which endpoint lost an smsg completion. clear the
smsg remote cq and check all mailboxes */
/* clear out remote cq */
mca_btl_ugni_cq_clear (btl->devices, btl->smsg_remote_cq);
endpoint_count = opal_pointer_array_get_size (&btl->endpoints);
for (ep_index = 0, count = 0 ; ep_index < endpoint_count ; ++ep_index) {
mca_btl_base_endpoint_t *ep;
ep = (mca_btl_base_endpoint_t *) opal_pointer_array_get_item (&btl->endpoints,
ep_index);
if (NULL == ep || MCA_BTL_UGNI_EP_STATE_CONNECTED != ep->state) {
continue;
}
/* clear out smsg mailbox */
rc = mca_btl_ugni_smsg_process (ep);
if (OPAL_LIKELY(rc >= 0)) {
count += rc;
}
}
return count;
}
int mca_btl_ugni_progress_remote_smsg (mca_btl_ugni_module_t *btl)
{
mca_btl_base_endpoint_t *ep;
gni_cq_entry_t event_data;
gni_return_t grc;
uint64_t inst_id;
grc = mca_btl_ugni_gni_cq_get_event (btl->devices, btl->smsg_remote_cq, &event_data);
if (GNI_RC_NOT_DONE == grc) {
return 0;
}
if (OPAL_UNLIKELY(GNI_RC_SUCCESS != grc || !GNI_CQ_STATUS_OK(event_data) ||
GNI_CQ_OVERRUN(event_data))) {
if (GNI_RC_ERROR_RESOURCE == grc ||
(GNI_RC_SUCCESS == grc && GNI_CQ_OVERRUN(event_data))) {
/* recover from smsg cq overrun */
return mca_btl_ugni_handle_remote_smsg_overrun (btl);
}
BTL_ERROR(("unhandled error in GNI_CqGetEvent"));
/* unhandled error: crash */
assert (0);
return mca_btl_rc_ugni_to_opal (grc);
}
BTL_VERBOSE(("REMOTE CQ: Got event 0x%" PRIx64 ". msg id = %" PRIu64
". ok = %d, type = %" PRIu64, (uint64_t) event_data,
GNI_CQ_GET_INST_ID(event_data), GNI_CQ_STATUS_OK(event_data),
GNI_CQ_GET_TYPE(event_data)));
inst_id = GNI_CQ_GET_INST_ID(event_data);
ep = (mca_btl_base_endpoint_t *) opal_pointer_array_get_item (&btl->endpoints, inst_id);
if (OPAL_UNLIKELY(MCA_BTL_UGNI_EP_STATE_CONNECTED != ep->state)) {
/* due to the nature of datagrams we may get a smsg completion before
we get mailbox info from the peer */
BTL_VERBOSE(("event occurred on an unconnected endpoint! ep state = %d", ep->state));
return 0;
}
return mca_btl_ugni_smsg_process (ep);
}