ITR: Opportunistic Parallel Computation

Wayne State University, Detroit MI

Investigators

Abstract

This project proposes runtime and compiler support that will enable programs to harvest idle SMPs and/or single processor workstations to perform parallel computations. The unique feature of this system is the ability to adapt parallel programs to the dynamic availability of processors while exploiting the locality within an SMP. The project integrates several goals, namely: 1) Extend the Distributed Shared Memory, Strings, to support thread migration, incorporate adaptation to the changing number of available processors at runtime and propose techniques to balance data locality and the parallelism used when the number of processors changes at runtime. 2) Study the impact of eviction time on remapping strategies and constraints. 3) Develop compile-time support for parallel programs which can be executed in an environment where the number and the availability of the processors can change. 4) Develop analytic models and extensively evaluate the above compiler and runtime techniques using several "real programs". 5) Integrate the utilization of idle cycles for parallel computing on cluster of SMP workstations into the existing parallelization environment.