Causal Structure Learning for Dependency Analysis of Performance Data

This paper proposes causal models for representing relational information about the variables of a performance analysis. Causal structure learning algorithms are used for constructing such models from experimental data. However, the existing algorithms had to be extended to capture the complexity of performance models, since they contain a mixture of continuous and discrete variables, nonlinear relations and deterministic relations. To handle the first two cases, we use a general conditional independence test based on the mutual information between probabilistic variables, where the underlying probability distribution of the experimental data is estimated by Gaussian kernel density estimation. Deterministic relations between variables make that these variables contain the same information about other related variables. We use a the complexity of the relations as a criterion to decide upon which of the equivalent relations is the direct relation. Experiments with the aztec benchmark application for solving linear equations in parallel gave accurate models, providing insight in how each variable affects the overall performance.