An OpenMP Runtime for Transparent Work Sharing across Cache-Incoherent Heterogeneous Nodes

In this work, we present libHetMP , an OpenMP runtime for automatically and transparently distributing parallel computation across heterogeneous nodes. targets platforms comprising CPUs with different instruction set architectures (ISA) coupled by a high-speed memory interconnect, where cross-ISA binary incompatibility non-coherent caches require application data be marshaled to shared CPUs. Because of this, work distribution decisions must take into account both relative compute performance asymmetric communication overheads. drives workload without programmer intervention measuring characteristics during cross-node execution. A novel HetProbe loop iteration scheduler decides if execution is beneficial either distributes according the when it or places all on homogeneous providing best not. We evaluate using kernels from several benchmark suites show geometric mean 41% speedup in time some workloads may showcase irregular behavior among iterations, extend second called HetProbe-I. The evaluation HetProbe-I shows can further improve computation, cases up 24%, triggering periodic decisions.