Compile-support software pipelining for DSMs

Despite its merits, software DSM has a serious drawback, that is, over-simplified data distribution schemes. Most software DSM systems are page-based and often fail to exactly satisfy the data partition required by the programmer and cause a serious communication overhead. In order to reduce communication cost incurred by the limitation of the underlying DSM system, we present two communicative primitives to easier utilize software pipelining, giving good legibility to paralellized code. We performed software pipelining on appbt, a program in the NAS benchmark, on a software DSM called JIAJIA. Experiments show that the speed-up is greatly improved.