November 14th, 2012
Communication performance is of paramount importance to high performance computing (HPC) applications. MPI is widely used in HPC due to its portability across platforms. However, the performance of MPI implementations on large-scale supercomputers is significantly impacted by factors including its inherent buffering, type checking, and other control overheads.
At the same time, we are witnessing the advent of intelligent network interfaces and system interconnects together with support for paradigms such as active messages. The Parallel Active Message Interface (PAMI) library for the IBM Blue Gene/Q system and the User Generic Network Interface (uGNI) for the Cray XE6 system are communication libraries enabling one to fully exploit the features of the underlying network interconnects. In this poster, we evaluate the performance of MPI with PAMI and uGNI and demonstrate that these libraries achieve significant improvements in communication performance over MPI.
We are working towards creating abstractions that hide the complexities found in current system interconnects, and expose a simple API to enable applications and middleware developers fully exploit the underlying features.
Bui, H., Vishwanath, V., Leigh, J., Papka, M.E., Evaluating Communication Performance of BlueGene/Q and Cray XE6 Supercomputers, (poster) SC12, Salt Lake City, UT, November 14th, 2012.