ANALYTIC MODELS OF SUPERCOMPUTER PERFORMANCE IN MULTIPROGRAMMING ENVIRONMENTS

Supercomputers run multiprogrammed time-sharing operating systems, so their facilities can be shared by many local and remote users. Therefore, it is important to be able to assess the performance of supercomputers in multiprogrammed environments. Analytic models based on Queueing Networks (QNs) and Stochastic Petri Nets (SPNs) are used in this paper with two purposes: to evaluate the performance of supercomputers in multi-programmed environments, and to compare, perfor-mance-wise, conventional supercomputer architectures with a novel architecture proposed here. It is shown, with the aid of the analytic models, that the proposed architecture is preferable performance-wise over the existing conventional supercomputer architectures. A three-level workload characterization model for super-computers is presented. Input data for the numerical ex-amples discussed here are extracted from the well-known Los Alamos benchmark, and the results are vali-dated by simulation