Maximum Response Time Indicator: Detecting Chaotic Behavior in Performance Data

Many recent studies of computing systems have revealed the presence of chaotic, power-tailed (PT), heavy-tailed behavior. For instance, PT properties have been found present in UNIX CPU process lifetimes, disk le sizes, and network traÆc. The signi cance of this is that in many cases PT phenomena can cause serious performance degradations, especially its impact upon the maximum response time. This paper presents heuristics that aid in determining the presence of PT phenomena. We also discuss a function called the Maximum Response Time Indicator which estimates the probability the response time of a given workload, transaction, device, etc., of interest will exceed some prede ned value. This function can be particularly useful for capacity planning purposes if highly-variable or PT behaviors have been observed.