1
1
openmpi/opal/mca/btl/ugni/btl_ugni_module.c
Nathan Hjelm b0ac6276a6 btl/ugni: improve multi-threaded RDMA performance
This commit improves the injection rate and latency for RDMA
operations. This is done by the following improvements:

 - If C11's _Thread_local keyword is available then always use the
   same virtual device index for the same thread when using RDMA. If
   the keyword is not available then attempt to use any device that
   isn't already in use. The binding support is enabled by default but
   can be disabled via the btl_ugni_bind_devices MCA variable.

 - When posting FMA and RDMA operations always attempt to reap
   completions after posting the operation. This allows us to
   better balance the work of reaping completions across all
   application threads.

 - Limit the total number of outstanding BTE transactions. This
   fixes a performance bug when using many threads.

 - Split out RDMA and local SMSG completion queue sizes. The RDMA
   queue size is better tuned for performance with RMA-MT.

 - Split out put and get FMA limits. The old btl_ugni_fma_limit MCA
   variable is deprecated. The new variable names are:
   btl_ugni_fma_put_limit and btl_ugni_fma_get_limit.

 - Change how post descriptors are handled. They are no longer
   allocated seperately from the RDMA endpoints.

 - Some cleanup to move error code out of the critical path.

 - Disable the FMA sharing flag on the CDM when we detect that there
   should be enough FMA descriptors for the number of virtual devices
   we plan will create. If the user sets this flag we will not unset
   it. This change should improve the small-message RMA performance by
   ~ 10%.

Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
2018-06-26 11:31:35 -06:00

363 строки
13 KiB
C

/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2011-2018 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2011 UT-Battelle, LLC. All rights reserved.
* Copyright (c) 2014-2016 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* Copyright (c) 2017 Intel, Inc. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "opal_config.h"
#include "btl_ugni.h"
#include "btl_ugni_frag.h"
#include "btl_ugni_endpoint.h"
#include "btl_ugni_prepare.h"
#include "btl_ugni_smsg.h"
static int
mca_btl_ugni_free (struct mca_btl_base_module_t *btl,
mca_btl_base_descriptor_t *des);
static int
mca_btl_ugni_module_finalize (struct mca_btl_base_module_t* btl);
static struct mca_btl_base_descriptor_t *
mca_btl_ugni_prepare_src (struct mca_btl_base_module_t *btl,
struct mca_btl_base_endpoint_t *endpoint,
struct opal_convertor_t *convertor,
uint8_t order, size_t reserve, size_t *size,
uint32_t flags);
static mca_btl_base_registration_handle_t *
mca_btl_ugni_register_mem (mca_btl_base_module_t *btl, mca_btl_base_endpoint_t *endpoint, void *base,
size_t size, uint32_t flags);
static int mca_btl_ugni_deregister_mem (mca_btl_base_module_t *btl, mca_btl_base_registration_handle_t *handle);
mca_btl_ugni_module_t mca_btl_ugni_module = {
.super = {
.btl_component = &mca_btl_ugni_component.super,
.btl_add_procs = mca_btl_ugni_add_procs,
.btl_del_procs = mca_btl_ugni_del_procs,
.btl_finalize = mca_btl_ugni_module_finalize,
.btl_alloc = mca_btl_ugni_alloc,
.btl_free = mca_btl_ugni_free,
.btl_prepare_src = mca_btl_ugni_prepare_src,
.btl_send = mca_btl_ugni_send,
.btl_sendi = mca_btl_ugni_sendi,
.btl_put = mca_btl_ugni_put,
.btl_get = mca_btl_ugni_get,
.btl_register_mem = mca_btl_ugni_register_mem,
.btl_deregister_mem = mca_btl_ugni_deregister_mem,
.btl_atomic_op = mca_btl_ugni_aop,
.btl_atomic_fop = mca_btl_ugni_afop,
.btl_atomic_cswap = mca_btl_ugni_acswap,
.btl_flush = mca_btl_ugni_flush,
}
};
static void mca_btl_ugni_datagram_event (int foo, short bar, void *arg)
{
mca_btl_ugni_module_t *ugni_module = (mca_btl_ugni_module_t *) arg;
mca_btl_ugni_device_t *device = ugni_module->devices;
struct timeval tv = {.tv_sec = 0, .tv_usec = MCA_BTL_UGNI_CONNECT_USEC};
mca_btl_ugni_progress_datagram (device);
opal_event_evtimer_add (&ugni_module->connection_event, &tv);
}
int
mca_btl_ugni_module_init (mca_btl_ugni_module_t *ugni_module)
{
int rc;
BTL_VERBOSE(("binding module %p to device 0", (void *) ugni_module));
/* copy module defaults (and function pointers) */
memmove (ugni_module, &mca_btl_ugni_module, sizeof (mca_btl_ugni_module));
ugni_module->initialized = false;
ugni_module->nlocal_procs = 0;
ugni_module->active_datagrams = 0;
ugni_module->active_rdma_count = 0;
opal_event_evtimer_set (opal_sync_event_base, &ugni_module->connection_event,
mca_btl_ugni_datagram_event, ugni_module);
OBJ_CONSTRUCT(&ugni_module->failed_frags, opal_list_t);
OBJ_CONSTRUCT(&ugni_module->failed_frags_lock, opal_mutex_t);
OBJ_CONSTRUCT(&ugni_module->eager_get_pending, opal_list_t);
OBJ_CONSTRUCT(&ugni_module->eager_get_pending_lock,opal_mutex_t);
for (int i = 0 ; i < MCA_BTL_UGNI_LIST_MAX ; ++i) {
OBJ_CONSTRUCT(ugni_module->frags_lists + i, opal_free_list_t);
}
OBJ_CONSTRUCT(&ugni_module->pending_smsg_frags_bb, opal_pointer_array_t);
OBJ_CONSTRUCT(&ugni_module->ep_wait_list_lock,opal_mutex_t);
OBJ_CONSTRUCT(&ugni_module->ep_wait_list, opal_list_t);
OBJ_CONSTRUCT(&ugni_module->endpoint_lock, opal_mutex_t);
OBJ_CONSTRUCT(&ugni_module->endpoints, opal_pointer_array_t);
OBJ_CONSTRUCT(&ugni_module->id_to_endpoint, opal_hash_table_t);
OBJ_CONSTRUCT(&ugni_module->smsg_mboxes, opal_free_list_t);
OBJ_CONSTRUCT(&ugni_module->eager_get_pending, opal_list_t);
/* set up virtual device handles */
for (int i = 0 ; i < mca_btl_ugni_component.virtual_device_count ; ++i) {
rc = mca_btl_ugni_device_init (ugni_module->devices + i, i);
if (OPAL_UNLIKELY(OPAL_SUCCESS != rc)) {
BTL_VERBOSE(("error initializing uGNI device handle"));
return rc;
}
}
/* create wildcard endpoint on first device to listen for connections.
* there is no need to bind this endpoint. We are single threaded
* here so there is no need for a device lock. */
rc = GNI_EpCreate (ugni_module->devices[0].dev_handle, NULL,
&ugni_module->wildcard_ep);
if (OPAL_UNLIKELY(OPAL_SUCCESS != rc)) {
BTL_ERROR(("error creating wildcard ugni endpoint"));
return mca_btl_rc_ugni_to_opal (rc);
}
/* post wildcard datagram */
rc = mca_btl_ugni_wildcard_ep_post (ugni_module);
if (OPAL_UNLIKELY(OPAL_SUCCESS != rc)) {
BTL_ERROR(("error posting wildcard datagram"));
return rc;
}
return OPAL_SUCCESS;
}
static int
mca_btl_ugni_module_finalize (struct mca_btl_base_module_t *btl)
{
mca_btl_ugni_module_t *ugni_module = (mca_btl_ugni_module_t *)btl;
mca_btl_base_endpoint_t *ep;
uint64_t key;
int rc;
if (ugni_module->initialized) {
/* close all open connections and release endpoints */
OPAL_HASH_TABLE_FOREACH(key, uint64, ep, &ugni_module->id_to_endpoint) {
if (NULL != ep) {
mca_btl_ugni_release_ep (ep);
}
}
if (mca_btl_ugni_component.progress_thread_enabled) {
mca_btl_ugni_kill_progress_thread();
}
/* destroy all cqs */
rc = GNI_CqDestroy (ugni_module->smsg_remote_cq);
if (GNI_RC_SUCCESS != rc) {
BTL_ERROR(("error tearing down RX SMSG CQ - %s",gni_err_str[rc]));
}
if (mca_btl_ugni_component.progress_thread_enabled) {
rc = GNI_CqDestroy (ugni_module->smsg_remote_irq_cq);
if (GNI_RC_SUCCESS != rc) {
BTL_ERROR(("error tearing down remote SMSG CQ - %s",gni_err_str[rc]));
}
}
/* cancel wildcard post */
rc = GNI_EpPostDataCancelById (ugni_module->wildcard_ep,
MCA_BTL_UGNI_CONNECT_WILDCARD_ID |
OPAL_PROC_MY_NAME.vpid);
if (GNI_RC_SUCCESS != rc) {
BTL_VERBOSE(("btl/ugni error cancelling wildcard post"));
}
/* tear down wildcard endpoint */
rc = GNI_EpDestroy (ugni_module->wildcard_ep);
if (GNI_RC_SUCCESS != rc) {
BTL_VERBOSE(("btl/ugni error destroying endpoint - %s",gni_err_str[rc]));
}
opal_event_del (&ugni_module->connection_event);
}
for (int i = 0 ; i < MCA_BTL_UGNI_LIST_MAX ; ++i) {
OBJ_DESTRUCT(ugni_module->frags_lists + i);
}
OBJ_DESTRUCT(&ugni_module->ep_wait_list);
OBJ_DESTRUCT(&ugni_module->smsg_mboxes);
OBJ_DESTRUCT(&ugni_module->pending_smsg_frags_bb);
OBJ_DESTRUCT(&ugni_module->id_to_endpoint);
OBJ_DESTRUCT(&ugni_module->endpoint_lock);
OBJ_DESTRUCT(&ugni_module->endpoints);
OBJ_DESTRUCT(&ugni_module->eager_get_pending);
OBJ_DESTRUCT(&ugni_module->eager_get_pending_lock);
if (ugni_module->rcache) {
mca_rcache_base_module_destroy (ugni_module->rcache);
}
for (int i = 0 ; i < mca_btl_ugni_component.virtual_device_count ; ++i) {
mca_btl_ugni_device_fini (ugni_module->devices + i);
}
ugni_module->initialized = false;
return OPAL_SUCCESS;
}
mca_btl_base_descriptor_t *
mca_btl_ugni_alloc(struct mca_btl_base_module_t *btl,
struct mca_btl_base_endpoint_t *endpoint,
uint8_t order, size_t size, uint32_t flags)
{
mca_btl_ugni_base_frag_t *frag = NULL;
/* do not allocate a fragment unless the wait list is relatively small. this
* reduces the potential for resource exhaustion. note the wait list only exists
* because we have no way to notify the sender that credits are available. */
if (OPAL_UNLIKELY(opal_list_get_size (&endpoint->frag_wait_list) > 32)) {
return NULL;
}
if (size <= mca_btl_ugni_component.smsg_max_data) {
frag = mca_btl_ugni_frag_alloc_smsg (endpoint);
} else if (size <= btl->btl_eager_limit) {
frag = mca_btl_ugni_frag_alloc_eager_send (endpoint);
}
if (OPAL_UNLIKELY(NULL == frag)) {
return NULL;
}
BTL_VERBOSE(("btl/ugni_module allocated frag of size: %u, flags: %x. frag = %p",
(unsigned int)size, flags, (void *) frag));
frag->base.des_flags = flags;
frag->base.order = order;
frag->base.des_segments = &frag->segments[1];
frag->base.des_segment_count = 1;
frag->segments[0].seg_addr.pval = NULL;
frag->segments[0].seg_len = 0;
frag->segments[1].seg_addr.pval = frag->base.super.ptr;
frag->segments[1].seg_len = size;
frag->flags = MCA_BTL_UGNI_FRAG_BUFFERED;
if (size > mca_btl_ugni_component.smsg_max_data) {
mca_btl_ugni_reg_t *registration;
frag->hdr_size = sizeof (frag->hdr.eager);
frag->flags |= MCA_BTL_UGNI_FRAG_EAGER | MCA_BTL_UGNI_FRAG_IGNORE;
registration = (mca_btl_ugni_reg_t *) frag->base.super.registration;
frag->hdr.eager.memory_handle = registration->handle;
} else {
frag->hdr_size = sizeof (frag->hdr.send);
}
return &frag->base;
}
static int
mca_btl_ugni_free (struct mca_btl_base_module_t *btl,
mca_btl_base_descriptor_t *des)
{
return mca_btl_ugni_frag_return ((mca_btl_ugni_base_frag_t *) des);
}
static struct mca_btl_base_descriptor_t *
mca_btl_ugni_prepare_src (struct mca_btl_base_module_t *btl,
mca_btl_base_endpoint_t *endpoint,
struct opal_convertor_t *convertor,
uint8_t order, size_t reserve, size_t *size,
uint32_t flags)
{
/* do not allocate a fragment unless the wait list is relatively small. this
* reduces the potential for resource exhaustion. note the wait list only exists
* because we have no way to notify the sender that credits are available. */
if (OPAL_UNLIKELY(opal_list_get_size (&endpoint->frag_wait_list) > 32)) {
return NULL;
}
return mca_btl_ugni_prepare_src_send (btl, endpoint, convertor,
order, reserve, size, flags);
}
static mca_btl_base_registration_handle_t *
mca_btl_ugni_register_mem (mca_btl_base_module_t *btl, mca_btl_base_endpoint_t *endpoint, void *base,
size_t size, uint32_t flags)
{
mca_btl_ugni_module_t *ugni_module = (mca_btl_ugni_module_t *) btl;
mca_btl_ugni_reg_t *reg;
int access_flags = flags & MCA_BTL_REG_FLAG_ACCESS_ANY;
int rc;
rc = ugni_module->rcache->rcache_register (ugni_module->rcache, base, size, 0, access_flags,
(mca_rcache_base_registration_t **) &reg);
if (OPAL_UNLIKELY(OPAL_SUCCESS != rc)) {
return NULL;
}
return &reg->handle;
}
static int mca_btl_ugni_deregister_mem (mca_btl_base_module_t *btl, mca_btl_base_registration_handle_t *handle)
{
mca_btl_ugni_module_t *ugni_module = (mca_btl_ugni_module_t *) btl;
mca_btl_ugni_reg_t *reg =
(mca_btl_ugni_reg_t *)((intptr_t) handle - offsetof (mca_btl_ugni_reg_t, handle));
(void) ugni_module->rcache->rcache_deregister (ugni_module->rcache, &reg->base);
return OPAL_SUCCESS;
}
int mca_btl_ugni_event_fatal_error (gni_return_t grc, gni_cq_entry_t event_data)
{
/* combined error check for get event and get completed. we might miss exactly
* what happened but it is unrecoverable anyway. fwiw, this error path has
* never been seen in production. */
if (GNI_CQ_OVERRUN(event_data)) {
/* TODO -- need to handle overrun -- how do we do this without an event?
will the event eventually come back? Ask Cray */
BTL_ERROR(("CQ overrun detected in RDMA event data. can not recover"));
} else {
BTL_ERROR(("Error in GNI_GetComplete %s", gni_err_str[grc]));
}
return mca_btl_rc_ugni_to_opal (grc);
}
int mca_btl_ugni_device_handle_event_error (mca_btl_ugni_rdma_desc_t *rdma_desc, gni_cq_entry_t event_data)
{
mca_btl_ugni_device_t *device = rdma_desc->device;
uint32_t recoverable = 1;
(void) GNI_CqErrorRecoverable (event_data, &recoverable);
if (OPAL_UNLIKELY(++rdma_desc->tries >= mca_btl_ugni_component.rdma_max_retries || !recoverable)) {
char char_buffer[1024];
GNI_CqErrorStr (event_data, char_buffer, sizeof (char_buffer));
BTL_ERROR(("giving up on desciptor %p, recoverable %d: %s", rdma_desc, recoverable, char_buffer));
return OPAL_ERROR;
}
return _mca_btl_ugni_repost_rdma_desc_device (device, rdma_desc);
}