Characterizing the dynamic behavior of workload execution in SVM systems

The overhead associated with software management of shared virtual memory (SVM) systems can seriously impact overall system performance. One way to remedy this situation is to design more efficient SVM consistency protocols. In this paper we study a number of parallel workload characteristics that can negatively impact the performance of SVM systems. We attempt to quantify the sources of performance loss in some parallel workloads. Our goal is to better understand these characteristics, enabling us to develop SVM protocols that can adjust to dynamics in workload behavior. This paper has three main contributions: i) we measure the contention for synchronization resources, showing how applications exhibit distinct phases during their execution, ii) we quantify the relationship between page size and fragmentation/false sharing while varying the sharing unit size, and iii) we study the synergies between the contention for synchronization resources and fragmentation/false sharing, providing hints for developing improved protocols.