Treffer: Enabling Inter-Machine Parallelism in High-Level Languages with SEJITS and MapReduce

Selective, embedded, just-in-time specialization (SEJITS) is a technique for optimizing embedded domain-specific languages through the use of specializers, or code modules developed by expert programmers that target particular accelerators such as multicore processors and GPUs via justin-time compilation. We extend SEJITS to exploit intermachine parallelism by targeting clusters of machines via MapReduce. Our work enables the development of specializers for large, data-parallel applications whose workflows can be cast as MapReduce operations. We present an implementation that targets Hadoop and we describe specializers for two applications. The first, a pure-Python protein docking application, requires a 1-line change to realize a 280x speedup on a cluster with 450 cores. The second, an audio processing application, demonstrates our approach’s ability to leverage clusters of GPU-equipped machines by composing parallel programming patterns. Results indicate that clusters are viable targets for specialization, and that pattern composition is a useful technique for managing multilevel parallelism.