Exploring Memory Management Strategies in Catamount

In this paper, we describe how the mapping of virtual to physical memory that is set up by the operating system can have a significant and unexpected impact on the performance of STREAM [6] and a sparse solver “mini-application” that exhibits STREAM-like behavior. The Cray Performance Analysis Toolkit (CrayPAT) [2] is utilized to identify the cause of the performance degradation to be row buffer conflicts in the memory system. This is shown to be due to the internal architecture of modern commodity DRAM chips. To mitigate the performance impact of row buffer conflicts, we explore alternative strategies for constructing a process’s virtual to physical memory mapping. Ultimately, this work points out another level of locality in the memory system, i.e., DRAM row buffers, that few system software and application developers are aware of and that can have a significant impact on performance.