Storage Workload Characterization and Consolidation in Virtualized Environments

Server virtualization is accelerating the already existing need for shared storage area network (SAN) infrastructure for the associated benefits of migration, flexible management and economies of sharing. Understanding the characteristics of a workload in terms of access locality, IO sizes, read write ratio etc. is crucial for effective placement and consolidation. The lack of visibility into workload details and placement often leads to design inefficiencies and overprovisioning. This paper presents workload characterization study of three top-tier enterprise applications using VMware ESX server hypervisor. We further separate out different components (for example data, index and redo log in a database) of these workloads to understand their behavior in isolation. We find that most workloads show highly random access patterns. Next, we study the impact of storage consolidation on workloads (both random and sequential) and their burstiness. Our experiments show that its beneficial to put random workloads together because it improves utilization and reduces latency without hurting performance. In case of consolidating random and sequential workloads, we see performance degradation for the sequential workload. Burst analysis reveals that most workloads are very bursty and the rate can change by as much as 50% within a few seconds. This makes it even more important to have the ability to absorb bursts and exploit benefits of statistical multiplexing by doing consolidation of workloads, whenever possible.