Bruce L. Worthington

Characterization of storage workload traces from production Windows Servers

The scarcity of publicly available storage workload traces of production servers impairs characterization, modeling research, and development efforts across the storage industry. Twelve sets of storage traces from a diverse set of Microsoft Corporation production servers were captured using ETW (event tracing for windows) instrumentation. Windows server 2008 dramatically increases the breadth and depth of ETW instrumentation, and new trace capture and visualization tools are available in the Windows Performance Tools kit. Additional analytical tools were developed to analyze and visualize traces captured from Exchange, software build and release, Live Maps, MSN storage, security authentication, and display advertisement platform servers. This paper contains a first set of characterizations for these traces, including simple block-level statistics, multi-parameter distributions, rankings of file access frequencies, and more complex analyses such as temporal and spatial self-similarity measurements. Trace data visualizations enable the examination of workload parameters, subcomponents, phases, and deviations from predicted behavior.