Past Event: Babuška Forum

Integrative Parallel Programming in HPC

Victor Eijkhout, Research Scientist, HPC Software Tools Group, TACC, UT Austin

Abstract

With the ongoing proliferation of architecture types (distributed memory clusters, shared memory, co-processors) comes a proliferation of programming modes (message passing, active messages, loop-based and task-based parallelism, new SIMD variants), complicating the life of a scientific programmer. To extract all possible performance, a code often has to rely on using more than one mode of parallelism, making it hard to maintain, hard to port, and far from 'future-proof'. In this talk I will present a new framework for parallel programming, called the 'Integrative Model for Parallelism (IMP)'. It is based on an abstract model of parallel computing that unifies existing models, yet is detailed enough to give performance comparable to hand-written codes in these models. I will give a basic overview of the IMP concepts, show a prototype implementation, and indicate future directions of research. Bio Victor joined TACC in 2005 as a Research Scientist in the High Performance Computing group. Before coming to TACC, he held positions at the University of Illinois, the University of California at Los Angeles, and the University of Tennessee in Knoxville. Research areas: Numerical Linear Algebra, Parallel processing, Performance optimization, Machine learning, and Component software. http://www.tacc.utexas.edu/~eijkhout/