Knowledge-Based Modeling Approach for Performance Measurement of Parallel Systems

Parallel systems are important computing platforms because they offer tremendous potential to solve inherentlyparallel and computation intensive applications. Performance is always a key consideration in determining the success of suchsystems. Evaluating and analyzing parallel system is difficult due to the complex interaction between applicationcharacteristics and architectural features. Traditional performance methodologies like experimental measurement, theoretical/analytical modeling and simulation naturally apply to the performance evaluation of parallel systems. Experimental measurement uses real or synthetic workloads, usually known as benchmarks, to evaluate and analyze theirperformance on actual hardware. Theoretical/analytical models try to abstract details of a parallel system. Simulation andother performance monitoring/visualization tools are extremely popular because they can capture the dynamic nature of theinteraction between applications and architectures. Each of them has several types. For example, experimental measurement hassoftware, hardware, and hybrid. Theoretical/analytical modeling has queueing network, petri net, etc. and simulation hasdiscrete event, trace/execution driven, Monte Carlo. All of them have their own advantages and disadvantages. The first partof this paper will concentrate on identifying parameters for carrying out a comparative survey on these techniques and secondpart will justify the need for some kind of modelling approach which combines the advantages of all the three performanceevaluation techniques and lastly paper will be focusing on an integrated model combining all the three techniques and usingknowledge-based systems to evaluate the performance of parallel systems. This paper also discusses certain issues likeselecting an appropriate metric for evaluating parallel systems; need to select proper workload and workloadcharacterization