70 likes | 130 Views
Proposed ideas for consideration under AUS. Revisit Mixed MPI-OpenMP Programming model?. Is it time to revive mixed MPI-OpenMP programming model? People have been looking at it for a while Not very successful so far Not enough cores/processor to justify Not enough processors/node to justify
E N D
Revisit Mixed MPI-OpenMP Programming model? • Is it time to revive mixed MPI-OpenMP programming model? • People have been looking at it for a while • Not very successful so far • Not enough cores/processor to justify • Not enough processors/node to justify • Halfhearted attempts? • Have things changed? • 8-16 cores per node in current T2 systems • Are we at the tipping point? • Or, where is the tipping point?
Hybrid OpenMP-MPI Benchmark • Simple benchmark code • Permits systematic evaluation • Vary compute-comm ratio • Vary comm message sizes • Vary MPI – OpenMP balance • Should we expect better performance? • Is this a worthwhile approach for real applications? • Hopefully provides some limits in the idealized case
Real application with Mixed MPI-OpenMP • WRF • ENZO • POPS • Other user codes?
FFT benchmarks on T2 systems • Not the HPCC one • More realistic dimensions • 256^3 to 4086^3 • 2D processor decomposition already implemented (PK’s code) • How would this compare with a slab decomposition with Mixed MPI-OpenMP? • Does one exist? • Also can result in tuning suggestions for users
Hybrid Architectures (Hardware) • Emerging architecture • There is a lot to be learned • Benchmark existing applications? • NAMD • WRF
IO Benchmarking? • Experiment with models actually used by users • Create "Best Practices“ for IO?