Improve CUDA resource management for MPI jobs#185
Open
vmitq wants to merge 2 commits intowavefunction91:masterfrom
Open
Improve CUDA resource management for MPI jobs#185vmitq wants to merge 2 commits intowavefunction91:masterfrom
vmitq wants to merge 2 commits intowavefunction91:masterfrom
Conversation
Split memory evenly between processes on one GPU
There was a problem hiding this comment.
Pull request overview
This PR updates device-backend initialization to become MPI-aware (when built with MPI), enabling per-node GPU assignment in a round-robin manner and adjusting reported available GPU memory for multi-process-per-GPU runs.
Changes:
- Extend
make_device_backendto accept an MPI communicator (when MPI is enabled) and propagate it from the device runtime environment. - Add an MPI-aware
CUDABackendconstructor that selects a GPU based on local shared-memory rank. - Adjust
CUDABackend::get_available_mem()to scale down available memory based on local process/device sharing.
Reviewed changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 6 comments.
Show a summary per file
| File | Description |
|---|---|
src/runtime_environment/device/hip/hip_backend.cxx |
Updates backend factory signature to accept an MPI communicator (currently unused). |
src/runtime_environment/device/device_runtime_environment_impl.hpp |
Passes the runtime communicator into make_device_backend (MPI builds). |
src/runtime_environment/device/device_backend.hpp |
Updates factory declaration to accept MPI communicator; adds include to access MPI types/macros. |
src/runtime_environment/device/cuda/cuda_backend.hpp |
Adds MPI-related state and an MPI-aware constructor to CUDABackend. |
src/runtime_environment/device/cuda/cuda_backend.cxx |
Implements MPI-aware CUDA init (local rank → GPU) and memory-splitting logic. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
- Free local_comm in CUDABackend destructor - Guard against ndev <= 0 - Move MPI_Barrier from get_available_mem() to the call site - Narrow include in device_backend.hpp - Suppress unused MPI_Comm parameter warning in HIP backend
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The code detects the number of local MPI processes and available CUDA devices and assigns a GPU ID to each process in a round-robin fashion. When determining the available memory, it is divided evenly among the processes sharing the same GPU.
That simplifies GPU resource management when running jobs with multiple GPUs per host or multiple processes per GPU.