Is Data Migration Evil in the NVM File System?

Abstract

The NVM file system often exhibits unstable I/O performance in a NUMA server environment due to frequent remote memory accesses when threads and data are exclusively placed on different NUMA nodes. Further, multiple threads may use all of the available bandwidth of the Integrated Memory Controller (iMC), causing an iMC bottleneck. NThread partly addresses the problems above by maximizing local memory accesses via migrating threads to data resident CPU node. However, NThread cannot benefit in cases when iM C is overloaded. Therefore, we propose Dragonfly, an approach that migrates data to the memory module of the CPU node where the thread is located when iM C is overloaded. The proposed approach inherently balances the load among iM Cs, thus offering a fair load-balancing among iMCs. Specifically, Dragonfly implements a Migration Trigger Policy (MTP) to migrate data between CPU nodes on an opportunistic basis, minimizing the performance overhead caused by unnecessary data migration. We implement and evaluate NThread and Dragonfly in the NOVA file system deployed on an Intel Optane DC PM server for different application scenarios via Filebench workloads. The evaluation confirms that Dragonfly outperforms on an average 3.26x higher throughput than NThread.