make torchstore api concurrent safe #55
Draft
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Outdated suggestions cannot be applied.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.
This pull request introduces a robust concurrency control mechanism for RDMA operations in TorchStore by adding an asynchronous executor and refactoring the codebase to ensure safe, sequential execution of RDMA requests. The changes also update client and storage interfaces to support this executor, improve error handling, and clarify environment variable defaults.
Concurrency and RDMA execution improvements:
OnceCellandSequentialExecutorclasses intorchstore/_async_utils.pyto provide async initialization and sequential execution for RDMA tasks, ensuring only one RDMA operation runs at a time.LocalClientand storage classes to accept and useSequentialExecutorfor RDMA operations, including passing the executor through all relevant layers and enforcing concurrency limits via a semaphore and method decorator. [1] [2] [3] [4] [5] [6] [7] [8] [9]API and initialization changes:
torchstore/api.pyto useOnceCell[LocalClient]for lazy, single initialization and refactored related functions to support this pattern. [1] [2] [3]Transport buffer interface updates:
TransportBufferandRDMATransportBufferto require an executor for RDMA read/write operations, enforcing sequential execution and updating method signatures. [1] [2] [3] [4]Configuration and minor fixes:
TORCHSTORE_RDMA_ENABLEDenvironment variable to"1", enabling RDMA by default.These changes collectively ensure safe and efficient RDMA usage in TorchStore, improve initialization and resource management, and lay groundwork for future scalability.