Grappa : A Latency-Tolerant Runtime for Large-Scale Irregular Applications

Grappa is a runtime system for commodity clusters of multicore computers that presents a massively parallel, single address space abstraction to applications. Grappa’s purpose is to enable scalable performance of irregular parallel applications, such as branch and bound optimization, SPICE circuit simulation, and graph processing. Poor data locality, imbalanced parallel work and complex communication patterns make scaling these applications difficult. Grappa serves both as a C++ user library and as a foundation for higher level languages. Grappa tolerates delays to remote memory by multiplexing thousands of lightweight workers to each processor core, balances load via fine-grained distributed work-stealing, increases communication throughput by aggregating smaller data requests into large ones, and provides efficient synchronization and remote operations. We present a detailed description of the Grappa system and performance comparisons on several irregular benchmarks to hand-optimized MPI code and to the Cray XMT, a custom system used to target the real-time graph-analytics market. We find Grappa to be 9X faster than MPI on a random access microbenchmark, between 3.5X and 5.4X slower than MPI on applications, and between 2.6X faster and 4.4X slower than the XMT.