ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность

OpenIB BTL has three channels through which data can be received (eager rdma,

Просмотр исходного кода 
high prio QPs and low prio QPs) and because not all of them are polled each time
progrgess() is called (to save on latency) starvation is possible. The commit
fixes this. Now each channel is polled, but higher priority channels are polled
more often. Three new parameters are introduced that control polling ratios 
between different channels.

This commit was SVN r17024.
Этот коммит содержится в:
Gleb Natapov 2007-12-23 12:29:34 +00:00
родитель 7eca186568
Коммит b06d92bdab
5 изменённых файлов: 295 добавлений и 220 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

14
ompi/mca/btl/openib/btl_openib.c
Просмотреть файл

@ -952,7 +952,7 @@ static int mca_btl_finalize_hca(struct mca_btl_openib_hca_t *hca)
}
}
OBJ_DESTRUCT(&hca->hca_lock);
free(hca);
OBJ_RELEASE(hca);
return OMPI_SUCCESS;
}
@ -960,7 +960,7 @@ int mca_btl_openib_finalize(struct mca_btl_base_module_t* btl)
{
mca_btl_openib_module_t* openib_btl;
mca_btl_openib_endpoint_t* endpoint;
int ep_index, rdma_index, i;
int ep_index, i;
int qp, rc = OMPI_SUCCESS;
openib_btl = (mca_btl_openib_module_t*) btl;
@ -978,16 +978,6 @@ int mca_btl_openib_finalize(struct mca_btl_base_module_t* btl)
mca_btl_openib_component.ib_num_btls--;
/* Release eager RDMAs */
for(rdma_index=0;
rdma_index < opal_pointer_array_get_size(openib_btl->eager_rdma_buffers);
rdma_index++) {
endpoint=opal_pointer_array_get_item(openib_btl->eager_rdma_buffers,rdma_index);
if(!endpoint) {
continue;
}
OBJ_RELEASE(endpoint);
}
/* Release all QPs */
for(ep_index=0;
ep_index < opal_pointer_array_get_size(openib_btl->hca->endpoints);

21
ompi/mca/btl/openib/btl_openib.h
Просмотреть файл

@ -197,6 +197,9 @@ struct mca_btl_openib_component_t {
that they all exist) */
char **if_list;
bool use_message_coalescing;
uint32_t cq_poll_ratio;
uint32_t cq_poll_progress;
uint32_t eager_rdma_poll_ratio;
#ifdef HAVE_IBV_FORK_INIT
/** Whether we want fork support or not */
int want_fork_support;
@ -242,7 +245,10 @@ typedef struct mca_btl_openib_port_info_t mca_btl_openib_port_info_t;
MCA_BTL_OPENIB_LID_HTON(hdr); \
} while (0)
struct mca_btl_openib_hca_t {
struct mca_btl_base_endpoint_t;
typedef struct mca_btl_openib_hca_t {
opal_object_t super;
struct ibv_device *ib_dev; /* the ib device */
#if OMPI_ENABLE_PROGRESS_THREADS == 1
struct ibv_comp_channel *ib_channel; /* Channel event for the HCA */
@ -262,6 +268,9 @@ struct mca_btl_openib_hca_t {
uint8_t use_eager_rdma;
uint8_t btls; /** < number of btls using this HCA */
opal_pointer_array_t *endpoints;
uint16_t hp_cq_polls;
uint16_t eager_rdma_polls;
bool pollme;
#if OMPI_HAVE_THREADS
volatile bool got_fatal_event;
#endif
@ -269,8 +278,11 @@ struct mca_btl_openib_hca_t {
struct ibv_xrc_domain *xrc_domain;
int xrc_fd;
#endif
};
typedef struct mca_btl_openib_hca_t mca_btl_openib_hca_t;
uint32_t non_eager_rdma_endpoints;
int32_t eager_rdma_buffers_count;
struct mca_btl_base_endpoint_t **eager_rdma_buffers;
} mca_btl_openib_hca_t;
OBJ_CLASS_DECLARATION(mca_btl_openib_hca_t);
struct mca_btl_openib_module_pp_qp_t {
int32_t dummy;
@ -325,8 +337,7 @@ struct mca_btl_openib_module_t {
size_t ib_inline_max; /**< max size of inline send*/
size_t eager_rdma_frag_size; /**< length of eager frag */
opal_pointer_array_t *eager_rdma_buffers; /**< RDMA buffers to poll */
volatile int32_t eager_rdma_buffers_count; /**< number of RDMA buffers */
volatile int32_t eager_rdma_channels; /**< number of open RDMA channels */
mca_btl_base_module_error_cb_fn_t error_cb; /**< error handler */

448
ompi/mca/btl/openib/btl_openib_component.c
Просмотреть файл

@ -84,7 +84,7 @@ static void merge_values(ompi_btl_openib_ini_values_t *target,
ompi_btl_openib_ini_values_t *src);
static char* btl_openib_component_status_to_string(enum ibv_wc_status status);
static int btl_openib_component_progress(void);
static int btl_openib_module_progress(mca_btl_openib_hca_t *hca);
static int poll_hca(mca_btl_openib_hca_t *hca, int count);
static void progress_pending_frags_pp(mca_btl_base_endpoint_t *, const int);
static void progress_pending_frags_srq( mca_btl_openib_module_t* , const int);
static void progress_pending_eager_rdma(mca_btl_base_endpoint_t*);
@ -506,6 +506,37 @@ static int init_one_port(opal_list_t *btl_list, mca_btl_openib_hca_t *hca,
return OMPI_SUCCESS;
}
static void hca_construct(mca_btl_openib_hca_t *hca)
{
hca->ib_dev = NULL;
hca->ib_dev_context = NULL;
hca->btls = 0;
hca->ib_cq[BTL_OPENIB_HP_CQ] = NULL;
hca->ib_cq[BTL_OPENIB_LP_CQ] = NULL;
hca->cq_size[BTL_OPENIB_HP_CQ] = 0;
hca->cq_size[BTL_OPENIB_LP_CQ] = 0;
hca->non_eager_rdma_endpoints = 0;
hca->hp_cq_polls = mca_btl_openib_component.cq_poll_ratio;
hca->eager_rdma_polls = mca_btl_openib_component.eager_rdma_poll_ratio;
hca->pollme = true;
hca->eager_rdma_buffers_count = 0;
hca->eager_rdma_buffers = NULL;
}
static void hca_destruct(mca_btl_openib_hca_t *hca)
{
if(hca->eager_rdma_buffers) {
int i;
for(i = 0; i < hca->eager_rdma_buffers_count; i++)
if(hca->eager_rdma_buffers[i])
OBJ_RELEASE(hca->eager_rdma_buffers[i]);
free(hca->eager_rdma_buffers);
}
}
OBJ_CLASS_INSTANCE(mca_btl_openib_hca_t, opal_object_t, hca_construct,
hca_destruct);
static int init_one_hca(opal_list_t *btl_list, struct ibv_device* ib_dev)
{
struct mca_mpool_base_resources_t mpool_resources;
@ -515,7 +546,7 @@ static int init_one_hca(opal_list_t *btl_list, struct ibv_device* ib_dev)
ompi_btl_openib_ini_values_t values, default_values;
int *allowed_ports = NULL;
hca = malloc(sizeof(mca_btl_openib_hca_t));
hca = OBJ_NEW(mca_btl_openib_hca_t);
if(NULL == hca){
BTL_ERROR(("Failed malloc: %s:%d\n", __FILE__, __LINE__));
return OMPI_ERR_OUT_OF_RESOURCE;
@ -523,11 +554,7 @@ static int init_one_hca(opal_list_t *btl_list, struct ibv_device* ib_dev)
hca->ib_dev = ib_dev;
hca->ib_dev_context = ibv_open_device(ib_dev);
hca->btls = 0;
hca->ib_cq[BTL_OPENIB_HP_CQ] = NULL;
hca->ib_cq[BTL_OPENIB_LP_CQ] = NULL;
hca->cq_size[BTL_OPENIB_HP_CQ] = 0;
hca->cq_size[BTL_OPENIB_LP_CQ] = 0;
OBJ_CONSTRUCT(&hca->hca_lock, opal_mutex_t);
if(NULL == hca->ib_dev_context){
BTL_ERROR(("error obtaining device context for %s errno says %s\n",
@ -541,7 +568,7 @@ static int init_one_hca(opal_list_t *btl_list, struct ibv_device* ib_dev)
goto close_hca;
}
/* If mca_btl_if_include/exclude were specified, get usable ports */
allowed_ports = (int*) malloc(hca->ib_dev_attr.phys_port_cnt * sizeof(int));
allowed_ports = (int*)malloc(hca->ib_dev_attr.phys_port_cnt * sizeof(int));
port_cnt = get_port_list(hca, allowed_ports);
if(0 == port_cnt) {
ret = OMPI_SUCCESS;
@ -730,6 +757,17 @@ static int init_one_hca(opal_list_t *btl_list, struct ibv_device* ib_dev)
hca->endpoints = OBJ_NEW(opal_pointer_array_t);
opal_pointer_array_init(hca->endpoints, 10, INT_MAX, 10);
opal_pointer_array_add(&mca_btl_openib_component.hcas, hca);
if(mca_btl_openib_component.max_eager_rdma > 0 &&
mca_btl_openib_component.use_eager_rdma &&
hca->use_eager_rdma) {
hca->eager_rdma_buffers =
calloc(mca_btl_openib_component.max_eager_rdma * hca->btls,
sizeof(mca_btl_openib_endpoint_t*));
if(NULL == hca->eager_rdma_buffers) {
BTL_ERROR(("Memory allocation fails\n"));
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
mca_btl_openib_component.hcas_count++;
return OMPI_SUCCESS;
}
@ -756,7 +794,7 @@ close_hca:
free(allowed_ports);
}
free_hca:
free(hca);
OBJ_RELEASE(hca);
return ret;
}
@ -910,12 +948,7 @@ static int finish_btl_init(mca_btl_openib_module_t *openib_btl)
}
}
openib_btl->eager_rdma_buffers = OBJ_NEW(opal_pointer_array_t);
opal_pointer_array_init(openib_btl->eager_rdma_buffers,
mca_btl_openib_component.max_eager_rdma,
mca_btl_openib_component.max_eager_rdma,
0);
openib_btl->eager_rdma_buffers_count = 0;
openib_btl->eager_rdma_channels = 0;
openib_btl->eager_rdma_frag_size = OPAL_ALIGN(
sizeof(mca_btl_openib_header_t) +
@ -1500,91 +1533,99 @@ void* mca_btl_openib_progress_thread(opal_object_t* arg)
ibv_ack_cq_events(ev_cq, 1);
while(btl_openib_module_progress(hca));
while(poll_hca(hca, 0));
}
return PTHREAD_CANCELED;
}
#endif
static int progress_one_hca(mca_btl_openib_hca_t *hca)
{
int i, c, count = 0, ret;
mca_btl_openib_recv_frag_t* frag;
mca_btl_openib_endpoint_t* endpoint;
uint32_t non_eager_rdma_endpoints = 0;
c = hca->eager_rdma_buffers_count;
non_eager_rdma_endpoints += (hca->non_eager_rdma_endpoints + hca->pollme);
for(i = 0; i < c; i++) {
endpoint = hca->eager_rdma_buffers[i];
if(!endpoint)
continue;
OPAL_THREAD_LOCK(&endpoint->eager_rdma_local.lock);
frag = MCA_BTL_OPENIB_GET_LOCAL_RDMA_FRAG(endpoint,
endpoint->eager_rdma_local.head);
if(MCA_BTL_OPENIB_RDMA_FRAG_LOCAL(frag)) {
uint32_t size;
mca_btl_openib_module_t *btl = endpoint->endpoint_btl;
opal_atomic_rmb();
if(endpoint->nbo) {
BTL_OPENIB_FOOTER_NTOH(*frag->ftr);
}
size = MCA_BTL_OPENIB_RDMA_FRAG_GET_SIZE(frag->ftr);
#if OMPI_ENABLE_DEBUG
if (frag->ftr->seq != endpoint->eager_rdma_local.seq)
BTL_ERROR(("Eager RDMA wrong SEQ: received %d expected %d",
frag->ftr->seq,
endpoint->eager_rdma_local.seq));
endpoint->eager_rdma_local.seq++;
#endif
MCA_BTL_OPENIB_RDMA_NEXT_INDEX(endpoint->eager_rdma_local.head);
OPAL_THREAD_UNLOCK(&endpoint->eager_rdma_local.lock);
frag->hdr = (mca_btl_openib_header_t*)(((char*)frag->ftr) -
size + sizeof(mca_btl_openib_footer_t));
to_base_frag(frag)->segment.seg_addr.pval =
((unsigned char* )frag->hdr) + sizeof(mca_btl_openib_header_t);
ret = btl_openib_handle_incoming(btl, to_com_frag(frag)->endpoint,
frag, size - sizeof(mca_btl_openib_footer_t));
if (ret != MPI_SUCCESS) {
btl->error_cb(&btl->super, MCA_BTL_ERROR_FLAGS_FATAL);
return 0;
}
count++;
} else
OPAL_THREAD_UNLOCK(&endpoint->eager_rdma_local.lock);
}
hca->eager_rdma_polls--;
if(0 == count || non_eager_rdma_endpoints != 0 || !hca->eager_rdma_polls) {
count += poll_hca(hca, count);
hca->eager_rdma_polls = mca_btl_openib_component.eager_rdma_poll_ratio;
}
return count;
}
/*
* IB component progress.
*/
static int btl_openib_component_progress(void)
{
int i, j, c;
int count = 0, ret;
mca_btl_openib_endpoint_t* endpoint;
int i;
int count = 0;
#if OMPI_HAVE_THREADS
if(mca_btl_openib_component.use_async_event_thread &&
mca_btl_openib_component.fatal_counter) {
if(OPAL_UNLIKELY(mca_btl_openib_component.use_async_event_thread &&
mca_btl_openib_component.fatal_counter)) {
goto error;
}
#endif
/* Poll for RDMA completions - if any succeed, we don't process the slower
* queues.
*/
for(i = 0; i < mca_btl_openib_component.ib_num_btls; i++) {
mca_btl_openib_recv_frag_t* frag;
mca_btl_openib_module_t* openib_btl = mca_btl_openib_component.openib_btls[i];
c = openib_btl->eager_rdma_buffers_count;
for(j = 0; j < c; j++) {
endpoint =
opal_pointer_array_get_item(openib_btl->eager_rdma_buffers, j);
if(!endpoint) /* shouldn't happen */
continue;
OPAL_THREAD_LOCK(&endpoint->eager_rdma_local.lock);
frag = MCA_BTL_OPENIB_GET_LOCAL_RDMA_FRAG (endpoint,
endpoint->eager_rdma_local.head);
if(MCA_BTL_OPENIB_RDMA_FRAG_LOCAL(frag)) {
uint32_t size;
opal_atomic_rmb();
if(endpoint->nbo) {
BTL_OPENIB_FOOTER_NTOH(*frag->ftr);
}
size = MCA_BTL_OPENIB_RDMA_FRAG_GET_SIZE(frag->ftr);
#if OMPI_ENABLE_DEBUG
if (frag->ftr->seq != endpoint->eager_rdma_local.seq)
BTL_ERROR(("Eager RDMA wrong SEQ: received %d expected %d",
frag->ftr->seq,
endpoint->eager_rdma_local.seq));
endpoint->eager_rdma_local.seq++;
#endif
MCA_BTL_OPENIB_RDMA_NEXT_INDEX(endpoint->eager_rdma_local.head);
OPAL_THREAD_UNLOCK(&endpoint->eager_rdma_local.lock);
frag->hdr = (mca_btl_openib_header_t*)(((char*)frag->ftr) -
size + sizeof(mca_btl_openib_footer_t));
to_base_frag(frag)->segment.seg_addr.pval =
((unsigned char* )frag->hdr) +
sizeof(mca_btl_openib_header_t);
ret = btl_openib_handle_incoming(openib_btl,
to_com_frag(frag)->endpoint, frag,
size - sizeof(mca_btl_openib_footer_t));
if (ret != MPI_SUCCESS) {
openib_btl->error_cb(&openib_btl->super,
MCA_BTL_ERROR_FLAGS_FATAL);
return 0;
}
count++;
} else
OPAL_THREAD_UNLOCK(&endpoint->eager_rdma_local.lock);
}
}
if(count) return count;
for(i = 0; i < mca_btl_openib_component.hcas_count; i++) {
mca_btl_openib_hca_t *hca =
opal_pointer_array_get_item(&mca_btl_openib_component.hcas, i);
count += btl_openib_module_progress(hca);
count += progress_one_hca(hca);
}
return count;
@ -1595,150 +1636,163 @@ error:
mca_btl_openib_component.fatal_counter = 0;
/* Lets found all fatal events */
for(i = 0; i < mca_btl_openib_component.ib_num_btls; i++) {
mca_btl_openib_module_t* openib_btl = mca_btl_openib_component.openib_btls[i];
mca_btl_openib_module_t* openib_btl =
mca_btl_openib_component.openib_btls[i];
if(openib_btl->hca->got_fatal_event) {
openib_btl->error_cb(&openib_btl->super, MCA_BTL_ERROR_FLAGS_FATAL);
openib_btl->error_cb(&openib_btl->super, MCA_BTL_ERROR_FLAGS_FATAL);
}
}
return count;
#endif
}
static int btl_openib_module_progress(mca_btl_openib_hca_t* hca)
static char *cq_name[] = {"HP CQ", "LP CQ"};
static void handle_wc(mca_btl_openib_hca_t* hca, const uint32_t cq,
struct ibv_wc *wc)
{
static char *cq_name[] = {"HP CQ", "LP CQ"};
int cq, qp;
int count = 0, ne = 0;
static int flush_err_printed[] = {0, 0};
mca_btl_openib_com_frag_t* frag;
mca_btl_base_descriptor_t *des;
mca_btl_openib_endpoint_t* endpoint;
mca_btl_openib_endpoint_t* endpoint;
mca_btl_openib_module_t *openib_btl = NULL;
struct ibv_wc wc;
ompi_proc_t* remote_proc = NULL;
int qp;
for(cq = 0; cq < 2;) {
des = (mca_btl_base_descriptor_t*)(uintptr_t)wc->wr_id;
frag = to_com_frag(des);
/* For receive fragments "order" contains QP idx the fragment was posted
* to. For send fragments "order" contains QP idx the fragment was send
* through */
qp = des->order;
endpoint = frag->endpoint;
if(endpoint)
openib_btl = endpoint->endpoint_btl;
if(wc->status != IBV_WC_SUCCESS)
goto error;
/* Handle work completions */
switch(wc->opcode) {
case IBV_WC_RDMA_READ:
OPAL_THREAD_ADD32(&endpoint->get_tokens, 1);
/* fall through */
case IBV_WC_RDMA_WRITE:
case IBV_WC_SEND:
if(openib_frag_type(des) == MCA_BTL_OPENIB_FRAG_SEND) {
opal_list_item_t *i;
while((i = opal_list_remove_first(&to_send_frag(des)->coalesced_frags)))
to_base_frag(i)->base.des_cbfunc(&openib_btl->super, endpoint,
&to_base_frag(i)->base, OMPI_SUCCESS);
}
/* Process a completed send/put/get */
des->des_cbfunc(&openib_btl->super, endpoint, des,OMPI_SUCCESS);
/* return send wqe */
qp_put_wqe(endpoint, qp);
if(IBV_WC_SEND == wc->opcode && !BTL_OPENIB_QP_TYPE_PP(qp)) {
OPAL_THREAD_ADD32(&openib_btl->qps[qp].u.srq_qp.sd_credits, 1);
/* new SRQ credit available. Try to progress pending frags*/
progress_pending_frags_srq(openib_btl, qp);
}
/* new wqe or/and get token available. Try to progress pending frags */
progress_pending_frags_wqe(endpoint->qps[qp].qp);
mca_btl_openib_frag_progress_pending_put_get(endpoint, qp);
break;
case IBV_WC_RECV:
if(wc->wc_flags & IBV_WC_WITH_IMM) {
endpoint = (mca_btl_openib_endpoint_t*)
opal_pointer_array_get_item(hca->endpoints, wc->imm_data);
frag->endpoint = endpoint;
openib_btl = endpoint->endpoint_btl;
}
/* Process a RECV */
if(btl_openib_handle_incoming(openib_btl, endpoint, to_recv_frag(frag),
wc->byte_len) != OMPI_SUCCESS) {
openib_btl->error_cb(&openib_btl->super, MCA_BTL_ERROR_FLAGS_FATAL);
break;
}
/* decide if it is time to setup an eager rdma channel */
if(!endpoint->eager_rdma_local.base.pval && endpoint->use_eager_rdma &&
wc->byte_len < mca_btl_openib_component.eager_limit &&
openib_btl->eager_rdma_channels <
mca_btl_openib_component.max_eager_rdma &&
OPAL_THREAD_ADD32(&endpoint->eager_recv_count, 1) ==
mca_btl_openib_component.eager_rdma_threshold) {
mca_btl_openib_endpoint_connect_eager_rdma(endpoint);
}
break;
default:
BTL_ERROR(("Unhandled work completion opcode is %d", wc->opcode));
if(openib_btl)
openib_btl->error_cb(&openib_btl->super, MCA_BTL_ERROR_FLAGS_FATAL);
break;
}
return;
error:
if(endpoint && endpoint->endpoint_proc && endpoint->endpoint_proc->proc_ompi)
remote_proc = endpoint->endpoint_proc->proc_ompi;
if(wc->status != IBV_WC_WR_FLUSH_ERR || !flush_err_printed[cq]++) {
BTL_PEER_ERROR(remote_proc, ("error polling %s with status %s "
"status number %d for wr_id %llu opcode %d qp_idx %d",
cq_name[cq], btl_openib_component_status_to_string(wc->status),
wc->status, wc->wr_id, wc->opcode, qp));
}
if(IBV_WC_RETRY_EXC_ERR == wc->status)
opal_show_help("help-mpi-btl-openib.txt", "btl_openib:retry-exceeded", true);
if(openib_btl)
openib_btl->error_cb(&openib_btl->super, MCA_BTL_ERROR_FLAGS_FATAL);
}
static int poll_hca(mca_btl_openib_hca_t* hca, int count)
{
int ne = 0, cq;
uint32_t hp_iter = 0;
struct ibv_wc wc;
hca->pollme = false;
for(cq = 0; cq < 2 && hp_iter < mca_btl_openib_component.cq_poll_progress;)
{
ne = ibv_poll_cq(hca->ib_cq[cq], 1, &wc);
if(0 == ne) {
/* don't check low prio cq if there was something in high prio cq */
if(count)
/* don't check low prio cq if there was something in high prio cq,
* but for each cq_poll_ratio hp cq polls poll lp cq once */
if(count && hca->hp_cq_polls)
break;
cq++;
hca->hp_cq_polls = mca_btl_openib_component.cq_poll_ratio;
continue;
}
if(ne < 0)
goto error;
count++;
des = (mca_btl_base_descriptor_t*)(uintptr_t)wc.wr_id;
frag = to_com_frag(des);
/* For receive fragments "order" contains QP idx the fragment was posted
* to. For send fragments "order" contains QP idx the fragment was send
* through */
qp = des->order;
endpoint = frag->endpoint;
if(endpoint)
openib_btl = endpoint->endpoint_btl;
if(wc.status != IBV_WC_SUCCESS)
goto error;
/* Handle work completions */
switch(wc.opcode) {
case IBV_WC_RDMA_READ:
OPAL_THREAD_ADD32(&endpoint->get_tokens, 1);
/* fall through */
case IBV_WC_RDMA_WRITE:
case IBV_WC_SEND:
if(openib_frag_type(des) == MCA_BTL_OPENIB_FRAG_SEND) {
opal_list_item_t *i;
while((i = opal_list_remove_first(
&to_send_frag(des)->coalesced_frags))) {
to_base_frag(i)->base.des_cbfunc(&openib_btl->super,
endpoint, &to_base_frag(i)->base, OMPI_SUCCESS);
}
}
/* Process a completed send/put/get */
des->des_cbfunc(&openib_btl->super, endpoint, des,OMPI_SUCCESS);
/* return send wqe */
qp_put_wqe(endpoint, qp);
if(IBV_WC_SEND == wc.opcode && !BTL_OPENIB_QP_TYPE_PP(qp)) {
OPAL_THREAD_ADD32(&openib_btl->qps[qp].u.srq_qp.sd_credits, 1);
/* new SRQ credit available. Try to progress pending frags*/
progress_pending_frags_srq(openib_btl, qp);
}
/* new wqe or/and get token available. Try to progress pending frags */
progress_pending_frags_wqe(endpoint->qps[qp].qp);
mca_btl_openib_frag_progress_pending_put_get(endpoint, qp);
break;
case IBV_WC_RECV:
if(wc.wc_flags & IBV_WC_WITH_IMM) {
endpoint = (mca_btl_openib_endpoint_t*)
opal_pointer_array_get_item(hca->endpoints, wc.imm_data);
frag->endpoint = endpoint;
openib_btl = endpoint->endpoint_btl;
}
/* Process a RECV */
if(btl_openib_handle_incoming(openib_btl, endpoint,
to_recv_frag(frag), wc.byte_len) != OMPI_SUCCESS) {
openib_btl->error_cb(&openib_btl->super,
MCA_BTL_ERROR_FLAGS_FATAL);
return 0;
}
/* decide if it is time to setup an eager rdma channel */
if (!endpoint->eager_rdma_local.base.pval &&
endpoint->use_eager_rdma &&
wc.byte_len < mca_btl_openib_component.eager_limit &&
openib_btl->eager_rdma_buffers_count <
mca_btl_openib_component.max_eager_rdma &&
OPAL_THREAD_ADD32(&endpoint->eager_recv_count, 1) ==
mca_btl_openib_component.eager_rdma_threshold) {
mca_btl_openib_endpoint_connect_eager_rdma(endpoint);
}
break;
default:
BTL_ERROR(("Unhandled work completion opcode is %d",
wc.opcode));
if(openib_btl)
openib_btl->error_cb(&openib_btl->super,
MCA_BTL_ERROR_FLAGS_FATAL);
break;
if(BTL_OPENIB_HP_CQ == cq) {
hca->pollme = true;
hp_iter++;
hca->hp_cq_polls--;
}
handle_wc(hca, cq, &wc);
}
return count;
error:
if(ne < 0){
BTL_ERROR(("error polling %s with %d errno says %s\n",
cq_name[cq], ne, strerror(errno)));
} else {
static int flush_err_printed[] = {0, 0};
ompi_proc_t* remote_proc = NULL;
if(frag && endpoint && endpoint->endpoint_proc &&
endpoint->endpoint_proc->proc_ompi) {
remote_proc = endpoint->endpoint_proc->proc_ompi;
}
if(wc.status != IBV_WC_WR_FLUSH_ERR || !flush_err_printed[cq]++) {
BTL_PEER_ERROR(remote_proc, ("error polling %s with status %s "
"status number %d for wr_id %llu opcode %d qp_idx %d",
cq_name[cq],
btl_openib_component_status_to_string(wc.status),
wc.status, wc.wr_id, wc.opcode, qp));
}
if(wc.status == IBV_WC_RETRY_EXC_ERR) {
opal_show_help("help-mpi-btl-openib.txt",
"btl_openib:retry-exceeded", true);
}
}
if(openib_btl)
openib_btl->error_cb(&openib_btl->super, MCA_BTL_ERROR_FLAGS_FATAL);
BTL_ERROR(("error polling %s with %d errno says %s\n", cq_name[cq], ne,
strerror(errno)));
return count;
}

17
ompi/mca/btl/openib/btl_openib_endpoint.c
Просмотреть файл

@ -544,7 +544,8 @@ void mca_btl_openib_endpoint_connected(mca_btl_openib_endpoint_t *endpoint)
}
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();
@ -850,13 +851,19 @@ void mca_btl_openib_endpoint_connect_eager_rdma(
buf);
if(mca_btl_openib_endpoint_send_eager_rdma(endpoint) == OMPI_SUCCESS) {
/* This can never fail because max number of entries allocated
* at init time */
mca_btl_openib_hca_t *hca = endpoint->endpoint_btl->hca;
mca_btl_openib_endpoint_t **p;
OBJ_RETAIN(endpoint);
assert(((opal_object_t*)endpoint)->obj_reference_count == 2);
opal_pointer_array_add(openib_btl->eager_rdma_buffers, endpoint);
do {
p = &hca->eager_rdma_buffers[hca->eager_rdma_buffers_count];
} while(!opal_atomic_cmpset_ptr(p, NULL, endpoint));
OPAL_THREAD_ADD32(&hca->non_eager_rdma_endpoints, -1);
assert(hca->non_eager_rdma_endpoints >= 0);
OPAL_THREAD_ADD32(&openib_btl->eager_rdma_channels, 1);
/* from this point progress function starts to poll new buffer */
OPAL_THREAD_ADD32(&openib_btl->eager_rdma_buffers_count, 1);
OPAL_THREAD_ADD32(&hca->eager_rdma_buffers_count, 1);
return;
}

15
ompi/mca/btl/openib/btl_openib_mca.c
Просмотреть файл

@ -425,8 +425,21 @@ int btl_openib_register_mca_params(void)
"Use message coalescing", 1, &ival, 0));
mca_btl_openib_component.use_message_coalescing = (0 != ival);
/* Info only */
CHECK(reg_int("cq_poll_ratio",
"how often poll high priority CQ versus low priority CQ",
100, &ival, REGINT_GE_ONE));
mca_btl_openib_component.cq_poll_ratio = (uint32_t)ival;
CHECK(reg_int("eager_rdma_poll_ratio",
"how often poll eager RDMA channel versus CQ",
100, &ival, REGINT_GE_ONE));
mca_btl_openib_component.eager_rdma_poll_ratio = (uint32_t)ival;
CHECK(reg_int("hp_cq_poll_per_progress",
"max number of completion events to process for each call "
"of BTL progress engine",
10, &ival, REGINT_GE_ONE));
mca_btl_openib_component.cq_poll_progress = (uint32_t)ival;
/* Info only */
mca_base_param_reg_int(&mca_btl_openib_component.super.btl_version,
"have_fork_support",
"Whether the OpenFabrics stack supports applications that invoke the \"fork()\" system call or not (0 = no, 1 = yes). Note that this value does NOT indicate whether the system being run on supports \"fork()\" with OpenFabrics applications or not.",