janniklinde opened a new pull request, #2409: URL: https://github.com/apache/systemds/pull/2409
This PR introduces pipelining support for operators that produce at most one `MatrixBlock` per incoming block. As current out-of-core primitives are highly concurrent, which may lead to fan-outs (and consequential OOMs) for fast-producing source streams, this implementation aims to process downstream operations in the same thread. To reliably clean up referenced objects from caller methods (stack), we defer the downstream task using a thread-local context (if available). Because the deferred call is still executed in the same task of the caller, sequences of pipelining operators are completed before new tasks are taken from the task queue. Additionally, this PR adds new primitives required for general matrix multiplies and adds new operators relying on these new primitives. We also added (optional) messaging capabilities operators and streams, which may be required to communicate stream capabilities in future (e.g., targeted requests of tiles, cached, ...). While it is still uncertain to what degree these messages can be used, they are required for future experiments. As streams sometimes need size information, they may now hold the corresponding `CacheableData` object. Here I don't know if these references can create issues regarding memory management. Finally, this PR contains various bugfixes, safety checks, improved error handling and additional cache features. Sorry (again) for the large PR. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
