ch4/wait: Optimize waitall when all requests are local #7620

raffenet · 2025-10-07T01:24:42Z

Pull Request Description

Optimize progress for a batch of local communication requests by skipping netmod progress. Shows a 5-10% improvement in ch4/shm bandwidth measurements on a single node of Cascade Lake.

Author Checklist

Provide Description
Particularly focus on why, not what. Reference background, issues, test failures, xfail entries, etc.
Commits Follow Good Practice
Commits are self-contained and do not do two things at once.
Commit message is of the form: module: short description
Commit message explains what's in the commit.
Passes All Tests
Whitespace checker. Warnings test. Additional tests via comments.
Contribution Agreement
For non-Argonne authors, check contribution agreement.
If necessary, request an explicit comment from your companies PR approval manager.

raffenet · 2025-10-07T01:25:06Z

test:mpich/ch4/most

raffenet · 2025-10-07T15:21:53Z

test:mpich/ch4/most

hzhou · 2025-10-07T20:32:19Z

src/mpid/ch4/src/ch4_wait.h

    state->vci_count = idx;
+
+    if (local_only) {
+        state->flag &= ~MPIDI_PROGRESS_NM;


We should still perform netmod progress during "global" progress.

mpich/src/mpid/ch4/src/ch4_progress.h

Line 78 in f47908f

if (state->flag & MPIDI_PROGRESS_NM && (is_global || !made_progress)) { \

Make it

if (is_global || (state->flag & MPIDI_PROGRESS_NM && !made_progress)) { \

raffenet · 2025-10-08T01:29:20Z

test:mpich/ch4/most

raffenet · 2025-10-09T18:06:09Z

test:mpich/ch4/most

raffenet · 2025-10-09T19:03:13Z

test:mpich/ch4/most

raffenet · 2025-10-09T19:17:08Z

test:mpich/ch4/most

raffenet · 2025-10-09T21:10:04Z

test:mpich/ch4/most

raffenet · 2025-10-10T14:59:06Z

test:mpich/ch4/most

raffenet · 2025-10-13T14:47:33Z

test:mpich/ch4/most

raffenet · 2025-10-13T16:47:52Z

ch4/ofi vci build crashed when the node it was running on went offline. this is ready for review.

hzhou · 2025-10-13T17:16:31Z

src/mpid/ch4/src/mpidig_rma.h

    int vci_target = MPIDI_WIN_TARGET_VCI(win, target_rank);
    sreq = MPIDIG_request_create(MPIR_REQUEST_KIND__RMA, 2, vci, vci_target);
    MPIR_ERR_CHKANDSTMT(sreq == NULL, mpi_errno, MPIX_ERR_NOREQ, goto fn_fail, "**nomemreq");
+    MPIDI_REQUEST_SET_LOCAL(sreq, MPIDI_rank_is_local(target_rank, win->comm_ptr), NULL);


Two separate comments:

We should add is_local to MPIDIG_request_create to ensure request created in the ch4-layer always have locality set

What about requests created in the MPIR-layer or requests that straddles between shm/netmod such as collectives or anysource receive? We may need expand MPIDI_REQUEST(rreq, is_local) to enum (or more like optional).

Two separate comments:

We should add is_local to MPIDIG_request_create to ensure request created in the ch4-layer always have locality set

Can do.

What about requests created in the MPIR-layer or requests that straddles between shm/netmod such as collectives or anysource receive? We may need expand MPIDI_REQUEST(rreq, is_local) to enum (or more like optional).

Actually this already happens for anysrc_partner via MPID_Request_create_hook. We can also initialize is_local=0 there and then rely on the shm layer to set it to true when appropriate. That would work since the optimization is to skip NM progress.

We can also initialize is_local=0 there and then rely on the shm layer to set it to true when appropriate. That would work since the optimization is to skip NM progress.

Yeah, we can do that and refactor later.

raffenet · 2025-10-13T19:56:09Z

test:mpich/ch4/most
test:mpich/ch4/gpu/most

raffenet · 2025-10-13T19:56:34Z

test:mpich/ch4/xpmem

When global progress is triggered we should ignore the progress state flags and just call every progress function. We also need to trigger global progress in the single vci case because we plan to make the progress loop locality aware.

When a request is created, initialize is_local to false. Avoids accidentally reading an uninitialized value. Other layers will update the value to true where appropriate, such as the shared memory or generic active message code.

IPC protocol is local by definition, so make sure to reflect that in any requests created in IPC code.

Skip netmod progress when the request(s) being waited on are determined to be local only.

raffenet · 2025-10-16T18:30:33Z

test:mpich/ch4/most
test:mpich/ch4/gpu/ofi

raffenet force-pushed the shm-local-only branch from 2f33a6b to f7218b9 Compare October 7, 2025 01:24

raffenet force-pushed the shm-local-only branch from f7218b9 to c825241 Compare October 7, 2025 15:21

hzhou reviewed Oct 7, 2025

View reviewed changes

raffenet force-pushed the shm-local-only branch from c825241 to 3a8cf7e Compare October 8, 2025 01:28

raffenet force-pushed the shm-local-only branch 2 times, most recently from 95989b8 to c405c76 Compare October 9, 2025 18:05

raffenet force-pushed the shm-local-only branch from c405c76 to 62ce4f7 Compare October 9, 2025 19:03

raffenet force-pushed the shm-local-only branch from 62ce4f7 to af516b9 Compare October 9, 2025 19:16

raffenet force-pushed the shm-local-only branch from af516b9 to 5f1827a Compare October 9, 2025 21:09

raffenet force-pushed the shm-local-only branch from 5f1827a to 98c5984 Compare October 10, 2025 14:49

raffenet force-pushed the shm-local-only branch from 98c5984 to 2c95be5 Compare October 13, 2025 14:47

raffenet requested a review from hzhou October 13, 2025 16:47

hzhou reviewed Oct 13, 2025

View reviewed changes

raffenet force-pushed the shm-local-only branch from 2c95be5 to df64e86 Compare October 13, 2025 19:55

raffenet added 4 commits October 16, 2025 13:28

ch4/progress: Improvements to global progress

8b4b5c9

When global progress is triggered we should ignore the progress state flags and just call every progress function. We also need to trigger global progress in the single vci case because we plan to make the progress loop locality aware.

ch4: Initialize request locality at creation

527a225

When a request is created, initialize is_local to false. Avoids accidentally reading an uninitialized value. Other layers will update the value to true where appropriate, such as the shared memory or generic active message code.

ch4/ipc: Add missing locality setting

d1f41d8

IPC protocol is local by definition, so make sure to reflect that in any requests created in IPC code.

ch4/wait: Optimize progress for local requests

67241fe

Skip netmod progress when the request(s) being waited on are determined to be local only.

raffenet force-pushed the shm-local-only branch from df64e86 to 67241fe Compare October 16, 2025 18:29

ch4/wait: Optimize waitall when all requests are local #7620

Are you sure you want to change the base?

ch4/wait: Optimize waitall when all requests are local #7620

Uh oh!

Conversation

raffenet commented Oct 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Pull Request Description

Author Checklist

Uh oh!

raffenet commented Oct 7, 2025

Uh oh!

raffenet commented Oct 7, 2025

Uh oh!

hzhou Oct 7, 2025 • 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!

raffenet commented Oct 8, 2025

Uh oh!

raffenet commented Oct 9, 2025

Uh oh!

raffenet commented Oct 9, 2025

Uh oh!

raffenet commented Oct 9, 2025

Uh oh!

raffenet commented Oct 9, 2025

Uh oh!

raffenet commented Oct 10, 2025

Uh oh!

raffenet commented Oct 13, 2025

Uh oh!

raffenet commented Oct 13, 2025

Uh oh!

hzhou Oct 13, 2025

Choose a reason for hiding this comment

Uh oh!

raffenet Oct 13, 2025

Choose a reason for hiding this comment

Uh oh!

hzhou Oct 13, 2025

Choose a reason for hiding this comment

Uh oh!

raffenet commented Oct 13, 2025

Uh oh!

raffenet commented Oct 13, 2025

Uh oh!

raffenet commented Oct 16, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

raffenet commented Oct 7, 2025 •

edited

Loading

hzhou Oct 7, 2025 •

edited

Loading