Put send completions to low prio CQ. Receive is more important.

This commit was SVN r16817.

ompi/mca/btl/openib/btl_openib.c

@@ -287,12 +287,10 @@ static int create_srq(mca_btl_openib_module_t *openib_btl)
             openib_btl->qps[qp].u.srq_qp.rd_posted = 0;
 #if HAVE_XRC
             if(BTL_OPENIB_QP_TYPE_XRC(qp)) {
-                int prio = qp_cq_prio(qp);
                 openib_btl->qps[qp].u.srq_qp.srq =
                     ibv_create_xrc_srq(openib_btl->hca->ib_pd,
                             openib_btl->hca->xrc_domain,
-                            openib_btl->hca->ib_cq[prio], &attr);
+                            openib_btl->hca->ib_cq[qp_cq_prio(qp)], &attr);
-                openib_btl->hca->cq_users[prio]++;
             } else
 #endif
             {
@@ -310,8 +308,20 @@ static int create_srq(mca_btl_openib_module_t *openib_btl)
        
     return OMPI_SUCCESS;
 }
-static int adjust_cq(mca_btl_openib_hca_t *hca, const int cq_size, const int cq)
+
+static int adjust_cq(mca_btl_openib_hca_t *hca, const int cq)
 {
+    uint32_t cq_size = hca->cq_size[cq];
+
+    /* make sure we don't exceed the maximum CQ size and that we 
+     * don't size the queue smaller than otherwise requested 
+     */
+     if(cq_size < mca_btl_openib_component.ib_cq_size[cq])
+        cq_size = mca_btl_openib_component.ib_cq_size[cq];
+
+    if(cq_size > (uint32_t)hca->ib_dev_attr.max_cq)
+        cq_size = hca->ib_dev_attr.max_cq;
+
     if(NULL == hca->ib_cq[cq]) {
         hca->ib_cq[cq] = ibv_create_cq_compat(hca->ib_dev_context, cq_size,
 #if OMPI_ENABLE_PROGRESS_THREADS == 1
@@ -347,7 +357,7 @@ static int adjust_cq(mca_btl_openib_hca_t *hca, const int cq_size, const int cq)
 #endif
     }
 #ifdef HAVE_IBV_RESIZE_CQ
-    else {
+    else if (cq_size > mca_btl_openib_component.ib_cq_size[cq]){
         int rc;
         rc = ibv_resize_cq(hca->ib_cq[cq], cq_size);
         /* For ConnectX the resize CQ is not implemented and verbs returns -ENOSYS 
@@ -358,66 +368,37 @@ static int adjust_cq(mca_btl_openib_hca_t *hca, const int cq_size, const int cq)
         }
     }
 #endif
-    hca->cq_size[cq] = cq_size;
 
     return OMPI_SUCCESS;
 }
 
-static int mca_btl_openib_size_queues( struct mca_btl_openib_module_t* openib_btl, size_t nprocs) 
+static int mca_btl_openib_size_queues(struct mca_btl_openib_module_t* openib_btl, size_t nprocs) 
 {
-    uint32_t min_hp_cq_size = openib_btl->hca->cq_size[BTL_OPENIB_HP_CQ],
+    uint32_t send_cqes, recv_cqes;
-             min_lp_cq_size = openib_btl->hca->cq_size[BTL_OPENIB_HP_CQ],
-             cq_size;
     int rc = OMPI_SUCCESS, qp;
     mca_btl_openib_hca_t *hca = openib_btl->hca;
 
     /* figure out reasonable sizes for completion queues */
     for(qp = 0; qp < mca_btl_openib_component.num_qps; qp++) { 
         if(BTL_OPENIB_QP_TYPE_SRQ(qp)) {
-            cq_size = mca_btl_openib_component.qp_infos[qp].rd_num +
+            send_cqes = mca_btl_openib_component.qp_infos[qp].u.srq_qp.sd_max;
-                mca_btl_openib_component.qp_infos[qp].u.srq_qp.sd_max;
+            recv_cqes = mca_btl_openib_component.qp_infos[qp].rd_num;
-            if(mca_btl_openib_component.qp_infos[qp].size <=
-                    mca_btl_openib_component.eager_limit) {
-                min_hp_cq_size += cq_size;
-            } else {
-                min_lp_cq_size += cq_size;
-            }
         } else {
-            cq_size = (mca_btl_openib_component.qp_infos[qp].rd_num +
+            send_cqes = (mca_btl_openib_component.qp_infos[qp].rd_num +
-                    mca_btl_openib_component.qp_infos[qp].u.pp_qp.rd_rsv) *
+                mca_btl_openib_component.qp_infos[qp].u.pp_qp.rd_rsv) * nprocs;
-                2 * nprocs;
+            recv_cqes = send_cqes;
-            if(mca_btl_openib_component.qp_infos[qp].size <=
-                    mca_btl_openib_component.eager_limit) {
-                min_hp_cq_size += cq_size;
-            } else {
-                min_lp_cq_size += cq_size;
-            }
         }
+        openib_btl->hca->cq_size[qp_cq_prio(qp)] += recv_cqes;
+        openib_btl->hca->cq_size[BTL_OPENIB_LP_CQ] += send_cqes;
     }
    
-    /* make sure we don't exceed the maximum CQ size and that we 
+    rc = adjust_cq(hca, BTL_OPENIB_HP_CQ);
-     * don't size the queue smaller than otherwise requested 
+    if(rc != OMPI_SUCCESS)
-     */
+        goto out;
-     if(min_lp_cq_size < mca_btl_openib_component.ib_lp_cq_size)
-        min_lp_cq_size =  mca_btl_openib_component.ib_lp_cq_size;
-    if(min_lp_cq_size > (uint32_t)openib_btl->hca->ib_dev_attr.max_cq)
-        min_lp_cq_size = openib_btl->hca->ib_dev_attr.max_cq;
 
-    if(min_hp_cq_size < mca_btl_openib_component.ib_hp_cq_size)
+    rc = adjust_cq(hca, BTL_OPENIB_LP_CQ);
-        min_hp_cq_size =  mca_btl_openib_component.ib_hp_cq_size;
+    if(rc != OMPI_SUCCESS)
-    if(min_hp_cq_size > (uint32_t)openib_btl->hca->ib_dev_attr.max_cq)
+        goto out;
-        min_hp_cq_size = openib_btl->hca->ib_dev_attr.max_cq;
-
-    if(min_hp_cq_size != hca->cq_size[BTL_OPENIB_HP_CQ]) {
-        rc = adjust_cq(hca, min_hp_cq_size, BTL_OPENIB_HP_CQ);
-        if(rc != OMPI_SUCCESS)
-            goto out;
-    }
-    if(min_lp_cq_size != hca->cq_size[BTL_OPENIB_LP_CQ]) {
-        rc = adjust_cq(hca, min_lp_cq_size, BTL_OPENIB_LP_CQ);
-        if(rc != OMPI_SUCCESS)
-            goto out;
-    }
 
     if(0 == openib_btl->num_peers)
        rc = create_srq(openib_btl); 

ompi/mca/btl/openib/btl_openib.h

@@ -144,8 +144,7 @@ struct mca_btl_openib_component_t {
     uint32_t reg_mru_len;    /**< Length of the registration cache most recently used list */
     uint32_t use_srq;        /**< Use the Shared Receive Queue (SRQ mode) */ 
     
-    uint32_t ib_lp_cq_size;     /**< Max outstanding CQE on the CQ */  
+    uint32_t ib_cq_size[2];  /**< Max outstanding CQE on the CQ */  
-    uint32_t ib_hp_cq_size;     /**< Max outstanding CQE on the CQ */  
     
     uint32_t ib_sg_list_size; /**< Max scatter/gather descriptor entries on the WQ */ 
     uint32_t ib_pkey_ix;     /**< InfiniBand pkey index */
@@ -253,7 +252,6 @@ struct mca_btl_openib_hca_t {
     struct ibv_device_attr ib_dev_attr;
     struct ibv_pd *ib_pd;
     struct ibv_cq *ib_cq[2];
-    uint32_t cq_users[2];
     uint32_t cq_size[2];
     mca_mpool_base_module_t *mpool;
     /* MTU for this HCA */

ompi/mca/btl/openib/btl_openib_component.c

@@ -495,8 +495,6 @@ static int init_one_hca(opal_list_t *btl_list, struct ibv_device* ib_dev)
     hca->btls = 0;
     hca->ib_cq[BTL_OPENIB_HP_CQ] = NULL;
     hca->ib_cq[BTL_OPENIB_LP_CQ] = NULL;
-    hca->cq_users[BTL_OPENIB_HP_CQ] = 0;
-    hca->cq_users[BTL_OPENIB_LP_CQ] = 0;
     hca->cq_size[BTL_OPENIB_HP_CQ] = 0;
     hca->cq_size[BTL_OPENIB_LP_CQ] = 0;
     OBJ_CONSTRUCT(&hca->hca_lock, opal_mutex_t); 
@@ -1550,21 +1548,26 @@ static int btl_openib_module_progress(mca_btl_openib_hca_t* hca)
 {
     static char *cq_name[] = {"HP CQ", "LP CQ"};
     int cq, qp;
-    int count = 0,ne = 0;
+    int count = 0, ne = 0;
     mca_btl_openib_com_frag_t* frag;
     mca_btl_base_descriptor_t *des;
     mca_btl_openib_endpoint_t* endpoint; 
     mca_btl_openib_module_t *openib_btl = NULL;
     struct ibv_wc wc; 
 
-    for(cq = 0; cq < 2; cq++) {
+    for(cq = 0; cq < 2;) {
-        if(0 == hca->cq_users[cq])
-            continue;
         ne = ibv_poll_cq(hca->ib_cq[cq], 1, &wc);
-        if(0 == ne) 
+        if(0 == ne) {
+            /* don't check low prio cq if there was something in high prio cq */
+            if(count)
+                break;
+            cq++;
             continue;
+        }
         if(ne < 0)
             goto error;
+
+        count++;
        
         des = (mca_btl_base_descriptor_t*)(uintptr_t)wc.wr_id;
         frag = to_com_frag(des);
@@ -1604,8 +1607,6 @@ static int btl_openib_module_progress(mca_btl_openib_hca_t* hca)
                 /* 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);
-
-                count++;
                 break;
             case IBV_WC_RECV:
                 if(wc.wc_flags & IBV_WC_WITH_IMM) {
@@ -1623,8 +1624,6 @@ static int btl_openib_module_progress(mca_btl_openib_hca_t* hca)
                     return 0;
                 }
 
-                count++; 
-
                 /* decide if it is time to setup an eager rdma channel */
                 if (!endpoint->eager_rdma_local.base.pval &&
                         endpoint->use_eager_rdma &&
@@ -1640,7 +1639,8 @@ static int btl_openib_module_progress(mca_btl_openib_hca_t* hca)
                 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);
+                    openib_btl->error_cb(&openib_btl->super,
+                            MCA_BTL_ERROR_FLAGS_FATAL);
                 break;
         }
     }

ompi/mca/btl/openib/btl_openib_mca.c

@@ -214,8 +214,8 @@ int btl_openib_register_mca_params(void)
                   "queue (will automatically be set to a minimum of "
                   "(2 * number_of_peers * btl_openib_rd_num))",
                   1000, &ival, REGINT_GE_ONE));
-    mca_btl_openib_component.ib_lp_cq_size =
+    mca_btl_openib_component.ib_cq_size[BTL_OPENIB_LP_CQ] =
-        mca_btl_openib_component.ib_hp_cq_size = (uint32_t) ival;
+        mca_btl_openib_component.ib_cq_size[BTL_OPENIB_HP_CQ] = (uint32_t) ival;
     
     CHECK(reg_int("ib_sg_list_size", "Size of IB segment list "
                   "(must be >= 1)", 

ompi/mca/btl/openib/connect/btl_openib_connect_oob.c

@@ -330,14 +330,13 @@ static int qp_create_one(mca_btl_base_endpoint_t* endpoint, int qp,
     struct ibv_qp *my_qp;
     struct ibv_qp_init_attr init_attr;
     struct ibv_qp_attr attr;
-    int prio = qp_cq_prio(qp);
 
     memset(&init_attr, 0, sizeof(init_attr));
     memset(&attr, 0, sizeof(attr));
 
     init_attr.qp_type = IBV_QPT_RC;
-    init_attr.send_cq = openib_btl->hca->ib_cq[prio];
+    init_attr.send_cq = openib_btl->hca->ib_cq[BTL_OPENIB_LP_CQ];
-    init_attr.recv_cq = openib_btl->hca->ib_cq[prio];
+    init_attr.recv_cq = openib_btl->hca->ib_cq[qp_cq_prio(qp)];
     init_attr.srq     = srq;
     init_attr.cap.max_send_sge = mca_btl_openib_component.ib_sg_list_size;
     init_attr.cap.max_recv_sge = mca_btl_openib_component.ib_sg_list_size;
@@ -371,7 +370,6 @@ static int qp_create_one(mca_btl_base_endpoint_t* endpoint, int qp,
     /* Setup meta data on the endpoint */
     endpoint->qps[qp].qp->lcl_psn = lrand48() & 0xffffff;
     endpoint->qps[qp].credit_frag = NULL;
-    openib_btl->hca->cq_users[prio]++;
 
     return OMPI_SUCCESS;
 }

ompi/mca/btl/openib/connect/btl_openib_connect_xoob.c

@@ -250,8 +250,6 @@ static int xoob_qp_create(mca_btl_base_endpoint_t* endpoint, xoob_qp_type type)
 
     qp_init_attr.send_cq = qp_init_attr.recv_cq = openib_btl->hca->ib_cq[prio];
 
-    openib_btl->hca->cq_users[prio]++;
-
     /* no need recv queue; receives are posted to srq */
     qp_init_attr.cap.max_recv_wr = 0;
     qp_init_attr.cap.max_send_wr = send_wr;