How Many Threads to Spawn during Program Multithreading?

Abstract

Thread-level program parallelization is key for exploiting the hardware parallelism of the emerging multi-core systems. Several techniques have been proposed for program multithreading. However, the existing techniques do not address the following key issues associated with multithread execution of a given program: (a) Whether multithreaded execution is faster than sequential execution; (b) How many threads to spawn during program multithreading. In this paper, we address the above limitations. Specifically, we propose a novel approach – T-OPT – to determine how many threads to spawn during multithreaded execution of a given program region. The latter helps to check under-subscribing and over-subscribing of the hardware resources. This in turn facilitates exploitation on higher level of thread-level parallelism (TLP) than what can be achieved using the state-of-the-art. We show that, from program dependence standpoint, use of larger number of threads than advocated by the proposed approach does not yield higher degree of TLP. We present a couple of case studies and results using kernels, extracted from open source codes, to demonstrate the efficacy of our techniques on a real machine.