Again Percentiles! My anomaly detection tool (SonR) now has option to use percentiles to calculate UCL and LCL (Control Limits).
Let's go a listen about percentiles at the following CMG international conference session:
PERF
Whether externally mandated or internally tracked, the enterprise relies on governance of application service response time objectives. In many cases, achieving service requirements in terms of the average response time may not deliver an experience that delights the consumer. The consumer may request a deeper level of governance. Service providers want to achieve the promised objectives and, on the other hand, avoid over-provisioning. This paper explores rules of thumb that can be applied to estimate 90th or 95th percentiles for service response times, based on the measured or predicted mean. The risk assessment behind these recommendations is described in the paper. Various types of networks were modeled and analyzed. Even though classical queueing models rely on strict assumptions (and rarely met in the real world), it was found that the classical M/M/1 model provided a useful upper bound. Another function was evaluated for tighter accuracy.
Presenter bio: Dr. Salsburg is an independent consultant. Previously, Dr. Salsburg was a Distinguished Engineer and Chief Architect for Unisys Technology Products. He was founder and president of Performance & Modeling, Inc. Dr. Salsburg has been awarded three international patents in the area of infrastructure performance modeling algorithms and software. In addition, he has published over 70 papers and has lectured world-wide on the topics of Real-Time Infrastructure, Cloud Computing and Infrastructure Optimization. In 2010, the Computer Measurement Group awarded Dr. Salsburg the A. A. Michelson Award.
Presenter bio: Co-founder and Chief Scientist, BGS Systems, 1975 - 1998