An Evaluation of Memory Contention in Global Memory Systems

Queueing system model has been used to study the performance of Global Memory Management Systems. We extended the AMVA solution presented in 1] to model a multi-class system where applications can have diierent access patterns, i.e., diierent page fault rates. We believe this model is more accurate since it can capture sudden changes in application behavior when predicting the response time of a system. We validate our multiclass AMVA solution against the original single-class model. We run experiments for a set of six diierent applications to assess the impact on the average response time of the presence of a small number of clients with high fault rate in the system. We also run experiments considering load imbalance across the servers. The results show how load imbalance can increase the already signiicant degradation on performance when high fault rate customers are added to the system.