Reducing communication overhead under parallel list processing in multicore clusters

Greedily using GPU capacity for data list processing in multicore-GPU platforms

“…Multicore DLML, or MC-DLML, was designed for this purpose. MC-DLML main design issues were [5], [6], [10]: i) locking overhead for MPI calls to be thread safe, ii) cache locality and memory consumption, and iii) intra-node synchronisation cost. Figure 1 shows a typical application in MC-DLML.…”

“…Obviously, stealing data items requires synchronisation between application threads, and there are many ways to organise it. We have designed a few synchronisation algorithms for this context [5], [6], [10]. The overall organisation of DLML_get() with those algorithms is that shown in Figure 2.…”

“…The first version of DLML was based on multiprocess parallelism and communication based on MPI message passing only. MultiCore (MC) DLML [5], [6] is based on multithreaded parallelism and shared memory communication for intra-node parallelism and message passing for inter-node parallelism. The first design of Multicore-GPU (MG) DLML [1] extended MC-DLML with both: i) communication between the host multicore platform and the attached GPU, and ii) processing of new dynamically generated data within the GPU.…”