Causal Models for Parallel Performance Analysis

This paper proposes causal models to enhance the performance analysis of parallel processing. Causal models explicitly denote the relations among the variables involved. This makes it possible to automate the modeling task as well as to present the user a clear and understandable performance analysis. It is a flexible approach, since new environment variables can easily be integrated and performance can be estimated from incomplete knowledge. Since independency among variables is the key information, it can help the construction of a performance model that separates application and system dependency.