Using cause-effect analysis to understand the performance of distributed programs

Abstract Understanding the performance of distributed programs can be very difficult, since a program’s performance depends on characteristics of the application, the underlying hard- ware, the software environment, and interactions among all three. In this paper we present cause-effect analysis (CEA), a general approach to understanding distributed program performance that facilitates performance analysis, tuning, and prediction. Using detailed program traces gathered at execution time as input, CEA automatically generates ex- planations for important performance phenomena, identify- ing code segments that are responsible for the occurrence of the phenomena. We illustrate our approach by describing CEA techniques for three classes of overheads in distributed programs: con- tention, synchronization, and communication. Using the ex- planations produced by CEA, we are able to understand and minimize common performance problems in real appli- cations including load imbalance, false sharing, and resource contention.