Taskgraph: A Low Contention OpenMP Tasking Framework

OpenMP is the de-facto standard for shared memory systems in High-Performance Computing (HPC). It includes a tasking model that offers high-level of abstraction to effectively exploit structured (loop-based) and highly dynamic unstructured (task-based) parallelism an easy flexible way. Unfortunately, run-time overheads introduced manage tasks are (very) high most common frameworks (e.g., GCC, LLVM), which defeats potential benefits model, makes it suitable coarse-grained only. This paper presents taskgraph , framework uses task dependency graph (TDG) represent region code implemented with order reduce associated management tasks, i.e., contention parallel orchestration, including creation synchronization. The TDG avoids related resolution dependencies greatly reduces those deriving from accesses resources. Moreover, taskgraph introduces record-and-replay execution accelerates its second execution. Overall, multiple optimizations presented this allow exploiting fine-grained cope trend current applications pointing leverage massive on-node parallelism, scheduling paradigms. on LLVM 15.0. Results show implementation outperforms vanilla system terms performance scalability, all considering coarse fine grained tasks. Furthermore, proposed competitive alternative thread cases.