| Hi-index | 0.00 |
In a distributed shared memory system, each memory page isassociated with a home node that maintains the directory state forcache lines within that page. Memory access patterns and homenode locations have a strong influence on performance, especiallyif remote communication is costly. Since access patterns are difficultto predict and may change dynamically, it is useful to dynamicallymigrate home nodes to reduce the amount of remote communication.This paper presents a new and efficient algorithmfor migrating home nodes in distributed shared memory systems.Unlike previous page migration algorithms, our algorithm avoidsglobal coordination, allowing the system to be more responsive tochanging workloads. We verify the algorithm's correctness withthe Mur\varphi protocol verification tool. We explore several policiesfor deciding when and where to migrate home nodes. Trace-drivensimulations of several SPLASH-2 benchmarks show that our homemigration algorithm and policies can reduce the amount of remotecommunication by 50%. The results also emphasize the importanceof minimizing the cost of migration.