Taming Performance Hotspots in Cloud Storage with Dynamic Load Redistribution

Abstract

Cloud storage services are associated with high latency variance, and degraded throughput which is problematic when users are fetching and storing content for interactive applications. This can be attributed to performance hotspots created by slow nodes in a storage cluster, and performance interference caused by multi-tenancy, and background tasks such as data scrubbing, backfilling, recovery, etc. In this paper, we present DLR, a system that improves the performance of cloud storage services in the presence of hardware heterogeneity, and performance interference through a dynamic load redistribution technique. We designed DLR to dynamically adjust the load serving ratio of storage servers based on the system-level performance measurements from the storage cluster. We implemented DLR using Ceph, a popular distributed object storage system, and evaluated its performance on NSFCloud's Chameleon testbed using Ceph's Rados benchmark. Experimental results show that DLR improves the average throughput and latency of Ceph storage by up to 65%, and 41% respectively compared to the default case. Compared to Ceph's in-built load balancing technique, DLR improves the throughput by up to 98%, and latency by 96%.