Integrating Asynchronous Task Parallelism with OpenSHMEM

Graph 500 on OpenSHMEM: Using a Practical Survey of Past Work to Motivate Novel Algorithmic Developments

“…Hence, while previous application studies with AsyncSHMEM have focused on the use of asynchronous task parallelism for hybrid applications [13], in G500 we focus more on concurrency and programmability. That is, using a single runtime thread per PE with computation and communication multiplexed cooperatively on it by the AsyncSHMEM runtime.…”

“…The changes made to create the AsyncSHMEM versions of Concurrent-Fence and Concurrent-Hash are similar. In particular, we eliminate the need to periodically poll for new incoming messages by using the novel shmem_async_when family of APIs described in previous work [13]: In Concurrent-Fence and Concurrent-Hash, we use the shmem_async_when API to trigger tasks when new incoming messages are detected by either triggering them on the special header value in Concurrent-Fence, or on a change in the value stored at the location where we expect a header checksum to appear in Concurrent-Hash.…”

“…This report summarizes investigation into implementing the Graph500 benchmark on OpenSHMEM, and focuses on first building a strong and practical understanding of the strengths and limitations of past work before proposing and developing novel extensions. While our initial investigation explores implementations in pure OpenSHMEM, we also describe work implementing Graph500 using the AsyncSHMEM framework described in [13]. AsynSHMEM is a hybrid programming model and runtime system that composes OpenSHMEM and asynchronous task parallelism in order to enable novel APIs and functionality.…”

“…AsynSHMEM is a hybrid programming model and runtime system that composes OpenSHMEM and asynchronous task parallelism in order to enable novel APIs and functionality. For more details on AsyncSHMEM, please refer to the original publication [13] or the source code available at [1].…”

“…Section 3 describes the most relevant/successful OpenSHMEM-based implementations we have explored to date. Section 4 discusses how we use the AsyncSHMEM framework [13] to improve the developed reference OpenSHMEM implementations described in Section 3. Section 5 compares the performance of various OpenSHMEM-based and reference implementations.…”