Analytical Modeling for Operating System Schedulers on NUMA Systems 1

Performance evaluation by benchmarking is one of the main approaches for measuring performance of a computer system. However, it is important to measure parts of a system before they are even implemented. This can be achieved through an analytical description of the system, allowing the analysis of the system performance. Additionally, the analytical model can be extended to consider also reliability issues. This paper presents a generic model for an Operating System (OS) scheduler using the Stochastic Automata Networks (SAN) formalism. SAN are used to describe processes and processors in the OS and their behavior when processes have to be migrated. Moreover, processor failures are also modeled in order to provide reliability indices. The proposed model uses actual benchmarks results obtained from a 4-processor Itanium2 SMP machine and a 12-processor Itanium2 NUMA machine.