Performance of a system is a function of the system properties, and the workload seen by the system. One of the best ways to improve performance in systems is to tune or design the system based on the input workload. Localities in workloads greatly dictate the benefits one can extract from various cache hierarchies of the system stack. However, existing synthetic workload generators fail to reproduce traces that are a good representative of the original application, in terms of temporal and spatial localities. Additionally, existing workload generators are not flexible, and cannot handle cases that mimic changes in application behavior. Hence a probabilistic workload generator framework that produces synthetic trace with similar characteristics and locality as the original application, and has the support to accept or tune various workload parameter values to mimic existing or predicted workloads is presented. Apart from that, this workload generator has integration with a replay engine to issue trace IOs to a real system, or a storage simulator. Microsoft Research Traces were used for validating the tool, and the results show with up to 90% confidence that the ordering of synthetic trace is similar to the real trace. This tool can be used to study workloads in various environments like VM, cloud, database etc., and perform system optimizations or load studies.
University of Minnesota M.S. thesis. June 2014. Major: Electrical Engineering. Advisor: David J. Lilja. 1 computer file (PDF); vi, 108 pages, appendix p. 99-108.
Intelligent Block Level I/O workload characterization for a temporal and spatial locality aware workload generator.
Retrieved from the University of Minnesota Digital Conservancy,
Content distributed via the University of Minnesota's Digital Conservancy may be subject to additional license and use restrictions applied by the depositor.