Block Migration in Broadcast-based Multiprocessor Architectures

This paper presents techniques that improve the performance of parallel programs on distributed shared memory NUMA multiprocessors by implementing dynamic memory block and page migration. Our techniques address the latencies caused by the contention within the network and attempt to enhance data locality by migrating pages to reduce remote references. We analyze the behavior of eight multiprocessor applications, which exhibit a wide range of network traffic patterns. Results show that several applications that encounter hot spots and network congestion see a reduction of run time by more than a factor of ten