Speaker
Description
The architecture of supercomputers is evolving to expose massive parallelism by
considerably increasing the number of compute units per node. HPC users must
adapt their applications to remain efficient on current and ultimately on
future hardware. Open Multi-Processing (OpenMP) and the Message Passing
Interface (MPI) are two HPC programming standards widely used and both aim at
performant and portable codes but work on different parallelism levels: the
shared and the distributed memory levels.
OpenMP proposes a task-based programming model which composability shall enable
seemless hybridization with other asynchronous programming models such as MPI.
Composing the two standards thus appears as a well-suited solution for
performant and portable codes.
This poster presents the level of performance to expect from an optimized
MPI+(task-based OpenMP) software stack. Our methodology consists in modeling
hybrid applications to a unified task graph scheduling problem on which some
metrics are defined to analyze application performances. Proxy-applications
are ported, analyzed, and improved through both user-code and runtime
optimizations wherever it is the most suitable to preserve real-world code
representativeness. Performance results show 1.9 speedup weak-scaled to 16,000
cores from our task-based over the for-loop parallel version on LULESH. We
show the highest performances on fine dependant tasks of 100 us. in average
with about 80% communication overlap.
