Understanding the Behavior of Shared Memory Applications Using the SMiLE Monitoring Framework

Data locality is a key factor for the performance of parallel systems. In a Distributed Shared Memory (DSM) system, however, it is difficult for the users to maintain a high data locality as it is usually a priori unknown how the data is distributed among the nodes. In this article we introduce a monitoring framework that allows users to understand the memory behavior of parallel applications. The information offered by the monitor system enables data to be allocated or redistributed more rationally among memories. This reduces remote memory accesses and further improves parallel performance. Keywords—SCI, monitoring, data locality, performance optimization.