Theory and Operation of the Warwick Multiprocessor Scheduling (MS) System
S. C. Perry, J. S. Harper, Darren J. Kerbyson, G.R. Nudd · Warwick Research Archive Portal (University of Warwick) · 1999
This paper is concerned with the application of performance prediction techniques to the optimisation of parallel systems, and, in particular, the use of these techniques on-the-fly for optimising performance at run-time. In contrast to other performance tools, performance prediction results are made available very rapidly, which allows their use in real-time environments. When applied to program optimisation, this allows consideration of run-time variables such as input data and resource availability that are not, in general, available during the traditional (ahead-of-time) performance tuning stage. The main contribution of this work is the application of predictive performance data to the scheduling of a number of parallel tasks across a large heterogeneous distributed computing system. This is achieved through use of just-in-time performance prediction coupled with iterative heuristic algorithms for optimisation of the meta-schedule. The paper describes the main theoretical considerations for development of such a scheduling system, and then describes a prototype implementation, the MS scheduling system, together with some results obtained from this system when operated over a medium-sized (campus-wide) distributed computing network.