prov/efa: Introduce PPS enhancement interface by shijin-aws · Pull Request #12225 · ofiwg/libfabric

shijin-aws · 2026-05-08T01:10:28Z

A series of commit that introduce an operation level interface to allow user enhance the packet per second (PPS) .

shijin-aws · 2026-05-12T03:20:17Z

@mrgolin can u take a look

YonatanNachum · 2026-05-15T17:36:40Z

 	}

+
+	if (efa_env.enable_high_pps && (flags & FI_EFA_WR_HIGH_PPS))


the efa_env.enable_high_pps will be dropped eventually ?

Yes, it is a temporary to gate this feature before the firmware deployment is fully completed

jiaxiyan · 2026-05-18T18:20:03Z

+
+@pytest.mark.fabric(params=["efa", "efa-direct"])
+@pytest.mark.message_sizes(default_efa=PERF_SIZES, default_efa_direct=DIRECT_RMA_SIZES,
+                           pr_ci_efa=PERF_PR_CI, pr_ci_efa_direct=DIRECT_RMA_SIZES)


We don't need pr_ci_efa=PERF_PR_CI, pr_ci_efa_direct=DIRECT_RMA_SIZES if not @pytest.mark.pr_ci

I see all the tests above have such decorator, even if most of them don't have pytest.mark.pr_ci. I do want it to get run within PR CI, what is the suggested decorators exactly? @charlesstoll

talavr-amazon · 2026-05-18T21:28:38Z


+
+/*
+ * EFA provider-specific operation flags (bits 60-63).


nit: a sentence directing to flags 0-59 could be useful

talavr-amazon · 2026-05-18T21:34:34Z

 	}

+
+	if (efa_env.enable_processing_hints)


would FI_EFA_WR_HIGH_PPS ever be set if enable_processing_hints is not?
maybe just check for the first here?

FI_EFA_WR_HIGH_PPS should always be exposed to applications. So we need the env check as a runtime gate. And I would add a comment that such env is only a temporary gate before the fw deployment and we should remove it after the deployment

I made the environment variable as a gate as a separate commit f8124ce, and it should be reverted finally after the deployment.

talavr-amazon · 2026-05-18T21:36:21Z

+{
+	if (flags & FI_EFA_WR_HIGH_PPS) {
+		uint8_t wqe_hints = 0;
+		wqe_hints |= EFA_IO_PROCESSING_HINT_BURST_PPS_SENSITIVE;


this seems like you're preparing this function for more potential hints in the future, which I'm not sure we need to at this point. But if you do, then wqe_hints declaration and EFA_SET should be outside the HIGH_PPS if blocjk

My old implementation was just calling an EFA_SET(...., 1) when (flags & FI_EFA_WR_HIGH_PPS). Then I saw rdma-core's implementation is assuming that can be extended so I followed.
I also thought about making that wqe_hint and EFA_SET outside the if block but I think that can cause them always called for all flags which is not necessary

Updated the code to only call the whole processing hint function when (flags & FI_EFA_WR_HIGH_PPS)

talavr-amazon · 2026-05-18T21:37:52Z

 	ibv_wr_set_ud_addr(qp->ibv_qp_ex, ah->ibv_ah, qpn, qkey);

+#if HAVE_EFADV_WR_PROCESSING_HINTS
+	if (efa_env.enable_processing_hints && (flags & FI_EFA_WR_HIGH_PPS))


how can enable_processing_hints be set if HAVE_EFADV_WR_PROCESSING_HINTS is not defined? maybe the macro check is redundant?

HAVE_EFADV_WR_PROCESSING_HINTS only protects rdma-core interface, but we don't need a new rdma-core interface to enable processing hints, for data path direct code path. So these two are orthogonal

Same above, I made the environment variable as a gate as a separate commit f8124ce, and it should be removed finally after the deployment.

sunkuamzn · 2026-05-19T05:55:49Z

+ * such as the FI_EFA_WR_HIGH_PPS flag. It currently supports write,
+ * writedata, and read operations.
+ *
+ * Unlike fi_rma_bw, this test uses a nonblocking benchmark loop that


Why not make this test fi_rma_bw_non_blocking?

@j-xiong do you have an opinion? Will this test be useful for other providers?

Yes, I think it is useful in general.

This test was implementing some efa specific flags, I can do some refactor to make it ported to common code for the non-blocking pattern. But it will require more work

Define FI_EFA_WR_HIGH_PPS as a provider-specific operation flag (bit 60) in fi_ext_efa.h. Applications pass this flag in fi_writemsg() to hint the device to optimize for higher message rate on RDMA write operations. Signed-off-by: Shi Jin <sjina@amazon.com>

In efa_data_path_direct_post_write(), set the processing hint bit in the TX WQE metadata ctrl3 field when both efa_env.enable_high_pps and FI_EFA_WR_HIGH_PPS are set. Add efa_send_wr_set_processing_hint_high_pps() helper and the EFA_IO_TX_META_DESC_PROCESSING_HINT_MASK definition for the new ctrl3 field in efa_io_defs.h. Signed-off-by: Shi Jin <sjina@amazon.com>

Implement the rdma-core data path integration for the high PPS processing hint on RDMA write operations. During QP creation, set EFADV_WR_EX_WITH_PROCESSING_HINT in wr_flags to enable the WQE-level hint setter. In efa_ibv_post_write(), call efadv_wr_set_processing_hint() when both the feature flag and FI_EFA_WR_HIGH_PPS are set. Add configure checks for efadv_qp_from_ibv_qp_ex, wr_flags, and EFADV_WR_PROCESSING_HINT_BURST_PPS_SENSITIVE. All changes are guarded by HAVE_EFADV_WR_PROCESSING_HINT for backward compatibility with older rdma-core versions. Signed-off-by: Shi Jin <sjina@amazon.com>

Add fi_efa_rma_bw, an EFA-specific RMA bandwidth test that supports write and writedata operations with EFA-specific features such as the FI_EFA_WR_HIGH_PPS flag. Unlike fi_rma_bw, this test uses a nonblocking benchmark loop that interleaves posting and completion polling to keep the pipeline full, similar to the approach used by rdma-core/perftest. This avoids blocking at window boundaries and maximizes throughput. Signed-off-by: Shi Jin <sjina@amazon.com>

Disable shm because it is a efa-only test. Signed-off-by: Shi Jin <sjina@amazon.com>

Add enable_high_pps field to efa_env struct, gated by the undocumented FI_EFA_ENABLE_HIGH_PPS environment variable. This allows controlled rollout of PPS optimization before firmware deployment is complete. Signed-off-by: Shi Jin <sjina@amazon.com>

shijin-aws force-pushed the pps branch from ad12f58 to f5333a9 Compare May 12, 2026 01:12

shijin-aws requested a review from a team May 12, 2026 01:53

shijin-aws force-pushed the pps branch from f5333a9 to 75d6655 Compare May 12, 2026 17:43

YonatanNachum reviewed May 15, 2026

View reviewed changes

jiaxiyan reviewed May 15, 2026

View reviewed changes

Comment thread prov/efa/src/efa_data_path_direct_internal.h Outdated

Comment thread prov/efa/test/efa_unit_test_data_path_direct.c

Comment thread prov/efa/src/efa_io_defs.h

Comment thread prov/efa/src/efa_base_ep.c

Comment thread prov/efa/test/efa_unit_test_rma.c Outdated

shijin-aws force-pushed the pps branch 3 times, most recently from 0fda097 to 94ff076 Compare May 16, 2026 01:43

shijin-aws requested review from YonatanNachum and jiaxiyan May 18, 2026 15:18

YonatanNachum approved these changes May 18, 2026

View reviewed changes

jiaxiyan reviewed May 18, 2026

View reviewed changes

talavr-amazon approved these changes May 18, 2026

View reviewed changes

talavr-amazon reviewed May 18, 2026

View reviewed changes

talavr-amazon approved these changes May 18, 2026

View reviewed changes

talavr-amazon previously approved these changes May 18, 2026

View reviewed changes

shijin-aws dismissed talavr-amazon’s stale review via 18de125 May 19, 2026 02:02

shijin-aws force-pushed the pps branch 2 times, most recently from 18de125 to 8f410a3 Compare May 19, 2026 02:06

sunkuamzn reviewed May 19, 2026

View reviewed changes

shijin-aws force-pushed the pps branch from 8f410a3 to ff12f4a Compare May 20, 2026 02:11

shijin-aws requested review from jiaxiyan, sunkuamzn and talavr-amazon May 20, 2026 04:55

shijin-aws added 5 commits May 20, 2026 15:20

fabtests/efa: Run high pps rma test

c594b95

Disable shm because it is a efa-only test. Signed-off-by: Shi Jin <sjina@amazon.com>

shijin-aws force-pushed the pps branch from ff12f4a to f8124ce Compare May 20, 2026 15:32

		}


		if (efa_env.enable_high_pps && (flags & FI_EFA_WR_HIGH_PPS))

Conversation

shijin-aws commented May 8, 2026

Uh oh!

shijin-aws commented May 12, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

shijin-aws May 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

shijin-aws May 19, 2026 •

edited

Loading