Exploiting approximate value locality for data synchronization on multi-core processors

Abstract

This paper shows that for a variety of parallel “soft computing” programs that use optimistic synchronization, the approximate nature of the values produced during execution can be exploited to improve performance significantly. Specifically, through mechanisms for imprecise sharing of values between threads, the amount of contention in these programs can be reduced thereby avoiding expensive aborts and improving parallel performance while keeping the results produced by the program within the bounds of an acceptable approximation. This is made possible due to our observation that for many such programs, a large fraction of the values produced during execution exhibit a substantial amount of value locality. We describe how this locality can be exploited using extensions to C/C++ language types that allow specification of limits on the precision and accuracy required and a novel value-aware conflict detection scheme that minimizes the number of conflicts while respecting these limits. Our experiments indicate that for the programs studied substantial speedups can be achieved - upto 5.7x over the original program for the same number of threads. We also present experimental evidence that for the programs studied, the amount of error introduced often grows relatively slowly.