Live Webcast 15th Annual Charm++ Workshop

Many computationally intensive parallel programs are also memory intensive. Even when the memory required to model a particular application is larger than the available memory, the virtual memory system permits the program to run. However, this comes at a substantial performance cost because disk accesses incurred by paging are substantially slower than memory accesses. Out-of-core techniques are often developed for each application separately to deal with this issue. We present a generic, application independent, technique that automatically improves paging performance, by using prefetching and multithreading. We take advantage of the virtualization provided by Charm++ objects in parallel programs. The data driven objects provide the prediction mechanism necessary for an effective prefetching scheme. An analytical model as well as experimental data is presented to demonstrate the advantages of the proposed scheme.