Improving Data Locality in Applications using Message Passing

Abstract

This thesis presents a systematic study of two modes of program execution: synchronous and asynchronous. In synchronous mode, program components are tightly coupled. Traditional procedure call represents the synchronous execution mode. In asynchronous mode, program components execute independently of each other. Asynchronous message passing represents the asynchronous execution mode. The asynchronous mode of execution introduces communication overhead in the execution of program components. However it improves the temporal locality of data in a program by facilitating temporal and spatial reorganization of program components. Temporal reorganization refers to the batched execution of program components. Spatial reorganization refers to the scheduling of components on different processors in order to avoid the over-subscription of cache memory. Synchronous execution avoids the communication overhead. The goal of this study is to systematically understand the trade-offs associated with each execution mode and the effect of each mode on the throughput and the resource utilization of applications. The findings of this study help derive application designs for achieving high throughput in current and future multicore hardware.