Performance Models for Legacy System Migration and Multi-core Computers - an MVA Approach

Abstract

Computer system architecture and design have witnessed continual change, with resulting challenges for performance modeling. In this thesis, we explore performance evaluation and modeling using queueing networks and Mean Value Analysis (MVA) approaches. We first investigate scenarios where legacy systems are migrated to newer platforms. We apply a novel MVA-based method - PELE - to address performance evaluation for the migrations. We extend the applicability of PELE with several case studies. Through the experiments, we demonstrate that the MVA algorithm is an efficient technique for computer system performance modeling. However, we find that it fails to address some features of multi-core platforms. To solve those problems, we present another MVA-based method - APEM. In APEM, we propose an approximation to estimate service demands, and use a flow-equivalent aggregation technique to approximate non-product-form networks. To validate the application of our method, we investigate three case studies of a closed, an open and a semi-open system, respectively. We show that our method achieves better accuracy compared with other commonly used MVA algorithms. In addition, we propose an approximate MVA algorithm - SMVA - to address the numerical instability of MVA that we encountered in previous case studies. Future research may use SMVA as a starting point to continue improvement of numerically stable MVA algorithms.