IBM Israel Research Seminars

The performance of computer systems depends, among other things, on the system workload. Performance evaluations are therefore often done using logs of workloads on current productions systems, under the assumption that such real workloads are representative and reliable. Similarly, workload modeling is typically based on real workloads. However, real workloads may also contain anomalies that make them non-representative and unreliable.

A previously unrecognized type of anomaly is workload flurries: surges of activity with a repetitive nature, caused by a single user, that dominate the workload for a relatively short period. Under suitable conditions, such flurries can have a decisive effect on performance evaluation results. However, workloads containing such flurries are not representative of normal usage. Moreover, creating a statistical model based on such workloads or using workloads containing flurries directly is also not representative of flurries in general. This motivates the approach of identifying and removing the flurries, which will allow the evaluation of the workload under normal conditions. We demonstrate this for several evaluations of parallel systems, showing that the anomalies in the workload as embodied by flurries carry over to anomalies in the evaluation results, which disappear when the flurries are removed. Such an evaluation can then be augmented by a separate evaluation of the deviation caused by the flurry.