A FLOPs based model for performance analysis and scheduling of applications for single and multiple CPUs

摘要

Optimal resource utilization ever remains a critical problem in any computing environment. Resource utilization merely depends upon an effective scheduling scheme. One approach is to have effective scheduling on the bases of application's performance on different resources. Once we know, the execution time of an application prior to execution, we can schedule it on a particular node in given computing environment, to gain maximum resource utilization. In the high performance computing environment, it is hard to estimate execution time accurately for CPU-oriented applications. So many approaches have been proposed to estimate performance of an application, such as source instrumentation, dynamic instrumentation, binary instrumentation, etc., especially for high computing - scientific applications. Dynamic approach, to model a performance of a scientific application, is based on its execution time itself for a given execution platform. But when the platform itself changed, a model may fail to estimate execution time accurately and hence it is not a platform independent modeling. This poor estimation leads to bad scheduling and ultimately poor resource utilization. We have found a better approach to estimate and model performance of scientific and high-computing applications based on floating point operations (FLOPs). We have targeted basically all high performance computing (HPC) and scientific applications, which need enormous computing power to produce results in time. In this paper, we explain design architecture of a system, implemented based on proposed idea. This system works for single-CPU and multiple-CPUs applications. In the same flow, we have also integrated usefulness of application modeling with application profiling to schedule on given HPC platform (PC, cluster or GRID) to achieve maximum resource utilization.