Skip to content

Conversation

@JuliaRS
Copy link
Contributor

@JuliaRS JuliaRS commented Jan 28, 2026

Added proper MPI_Win_fence calls to ensure all MPI_Accumulate operations occur within valid RMA epochs (per-iteration for NON-AGGREGATE mode, single epoch for AGGREGATE mode).
Initialized target buffer before the first RMA operation to prevent validation errors on sample #0 when CHECK is enabled.
Added MPI_Barrier for synchronization after initialization.
Updated comments to English for clarity.
This resolves issues with "Got invalid buffer" during validation and "Wrong synchronization of RMA calls" runtime errors.

JuliaRS added 3 commits January 28, 2026 09:33
- Added proper MPI_Win_fence calls to ensure all MPI_Accumulate operations
  occur within valid RMA epochs (per-iteration for NON-AGGREGATE mode,
  single epoch for AGGREGATE mode).
- Initialized target buffer before the first RMA operation to prevent
  validation errors on sample #0 when CHECK is enabled.
- Added MPI_Barrier for synchronization after initialization.
- Updated comments to English for clarity.

This resolves issues with "Got invalid buffer" during validation and
"Wrong synchronization of RMA calls" runtime errors.
- Measure only fence → Accumulate → fence (exclude init/validation/barriers).
- Move target initialization and CHK_DIFF outside the timed region.
- Normalize AGGREGATE time by r_cache_iter when CHECK is enabled.

Improves accuracy of AGGREGATE and NON-AGGREGATE results with -DCHECK.
@JuliaRS JuliaRS force-pushed the imb_ext_accumulate branch from 3a406e7 to ddd6f5a Compare January 28, 2026 15:34
@JuliaRS JuliaRS self-assigned this Jan 28, 2026
@JuliaRS JuliaRS added the bug label Jan 28, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants