Sparse Matrix Indexing Benchmarks

Daniel Havey, Aram Melkoumian, Feng Nan

In this project we will investigate methods of measuring the speed at which the communication infrastructure operates on parallel computers.  The de-facto standard in supercomputer benchmarks is Jack Dongarra's top500 list.  This benchmark does a fine job of measuring the number of flops performed by a system but because it is based on PDGEMM it largely ignores the communication overhead incurred by a particular algorithm.  We intend to address this problem by taking the flops out of the measurement and simply counting the time spent in communication.  In order to do this we will develop a metric to measure the amount of communication performed.  Something like pointers passed per second.  As a lower bound we will use an embarrassingly parallel problem (PDGEMM) and as an upper bound we will use an all-to-all dominated problem (GUPS).  We expect our measurements for the Game of Life and Conjugate Gradient to fall somewhere in between these bounds.