ofi: add inject path to ofi btl

We are working with a provider where some apps may do better using the OB1+OFI BTL than the OFI MTL.
To help some with performance, particularly for small message injection rate (e.g. osu_mbw_mr)
this patch adds an fi_inject option to the OFI BTL send method for messages that are
less than or equal to a provider's maximum inject message size.

Its important to be able to fall back to normal fi_send as several providers return -FI_EAGAIN regularly.

This patch improves the 8-byte message performance (as measured by osu_mbw_mr) by about 30%
when using the OB1 PML and OFI BTL and the PSM2 provider.

A new MCA parameter is added to be able to disable the inject path - btl_ofi_disable_inject.

Signed-off-by: Howard Pritchard <[email protected]>

Author: hppritcha

opal/mca/btl/ofi/btl_ofi.h

@@ -162,6 +162,10 @@ struct mca_btl_ofi_component_t {
 
     size_t namelen;
 
+    /** Maximum inject size */
+    size_t max_inject_size;
+    bool disable_inject;
+
     /** All BTL OFI modules (1 per tl) */
     mca_btl_ofi_module_t *modules[MCA_BTL_OFI_MAX_MODULES];
 };

opal/mca/btl/ofi/btl_ofi_component.c

@@ -189,6 +189,13 @@ static int mca_btl_ofi_component_register(void)
                                            MCA_BASE_VAR_SCOPE_READONLY,
                                            &mca_btl_ofi_component.rd_num);
 
+    mca_btl_ofi_component.disable_inject = false;
+    (void) mca_base_component_var_register(&mca_btl_ofi_component.super.btl_version, "disable_inject",
+                                           "disable use of fi_inject for short messages.",
+                                           MCA_BASE_VAR_TYPE_BOOL, NULL, 0, 0, OPAL_INFO_LVL_5,
+                                           MCA_BASE_VAR_SCOPE_READONLY,
+                                           &mca_btl_ofi_component.disable_inject);
+
     /* for now we want this component to lose to the MTL. */
     module->super.btl_exclusivity = MCA_BTL_EXCLUSIVITY_HIGH - 50;
 
@@ -477,6 +484,11 @@ static int mca_btl_ofi_init_device(struct fi_info *info)
         goto fail;
     }
 
+    /**
+     * Save the maximum sizes.
+     */
+    mca_btl_ofi_component.max_inject_size = ofi_info->tx_attr->inject_size;
+
     /* AV */
     av_attr.type = FI_AV_MAP;
     rc = fi_av_open(domain, &av_attr, &av, NULL);

opal/mca/btl/ofi/btl_ofi_frag.c

@@ -89,6 +89,7 @@ int mca_btl_ofi_send(mca_btl_base_module_t *btl, mca_btl_base_endpoint_t *endpoi
                      mca_btl_base_descriptor_t *descriptor, mca_btl_base_tag_t tag)
 {
     int rc = 0;
+    size_t msg_sz;
     mca_btl_ofi_context_t *context;
     mca_btl_ofi_module_t *ofi_btl = (mca_btl_ofi_module_t *) btl;
     mca_btl_ofi_endpoint_t *ofi_ep = (mca_btl_ofi_endpoint_t *) endpoint;
@@ -100,9 +101,35 @@ int mca_btl_ofi_send(mca_btl_base_module_t *btl, mca_btl_base_endpoint_t *endpoi
     /* This tag is the active message tag for the remote side */
     frag->hdr.tag = tag;
 
-    /* create completion context */
     context = get_ofi_context(ofi_btl);
-    comp = mca_btl_ofi_frag_completion_alloc(btl, context, frag, MCA_BTL_OFI_TYPE_SEND);
+    msg_sz = sizeof(mca_btl_ofi_header_t) + frag->hdr.len;
+
+    /*
+     * If supported by the provider and not disabled by
+     * the btl_ofi_disable_inject  MCA parameter,
+     * try to use the inject path for short messages.
+     * fi_inject can return -FI_EAGAIN if the provider is
+     * unable to buffer the message, so one needs to fall
+     * back to the fi_send path when that error code is
+     * returned. 
+     */
+    if (msg_sz <= mca_btl_ofi_component.max_inject_size &&
+        false == mca_btl_ofi_component.disable_inject) {
+        rc = fi_inject(context->tx_ctx, &frag->hdr, msg_sz, ofi_ep->peer_addr);
+        if (FI_SUCCESS == rc) {
+            mca_btl_ofi_frag_complete(frag, OPAL_SUCCESS);
+            return OPAL_SUCCESS;
+        }
+        /*
+         * -FI_EAGAIN is okay but any other error is a problem
+         */
+        if (-FI_EAGAIN != rc) {
+            return OPAL_ERROR;
+        }
+    }
+
+    /* create completion context */
+     comp = mca_btl_ofi_frag_completion_alloc(btl, context, frag, MCA_BTL_OFI_TYPE_SEND);
 
     /* send the frag. Note that we start sending from BTL header + payload
      * because we need the other side to have this header information. */