Live Webcast 15th Annual Charm++ Workshop

Many of the computationally intensive parallel applications are also memory intensive. The Operating System's Virtual Memory System will let the application run even when the required memory is beyond the available physical memory. However this comes at a substantial cost of taking a page fault, whenever there is a memory access for the swapped out pages. Normally, an application does not have any control of this memory system. But since the application will have a better idea of what pages will be needed in near future, out-of-core techniques for each application are developed seperately to deal with this issue. We present a generic, application independent, technique as part of the application's runtime library to improve the paging performance of these memory-intensive parallel applications. We exploit the message-driven execution style of parallel programming along with the virtualization provided by Charm++ objects. The data driven objects provide the prediction mechanism necessary for an effective prefetching scheme. The implementaion of this automatic out-of-core execution technique inside the Charm++ runtime libray is described along with experimental data using a real-world application.