Ergodic Continuous Hidden Markov Models for workload characterization

In this paper we present a novel approach for accurate characterization of the execution workload run by a computer. Usually, workload characterization is performed by measuring the type and amount of resources requested during a program execution (for instance the usage of CPU, I/O, network, etc.). The sequence of measures is then treated as a stochastic process and analyzed with statistical techniques. The novelty of our approach is that we instead use directly the sequence of memory references generated during the execution of a program. The sequences of memory references are treated as sequences of floating point numbers, and analyzed with signal processing techniques. In the feature extraction phase we use spectral analysis while in the pattern matching phase we use ergodic continuous hidden Markov models (ECHMMs). The ECHMM models estimated in an initial training phase can be used both for online workload classification of a running process and for synthetic traces generation. Several processes of the same workload are necessary to obtain an HMM model of the workload. The proposed algorithms is evaluated via trace driven simulations using the SPEC 2000 workloads. We show that ECHMMs describe address memory sequences; average classification accuracy is about 76% with eight different workloads.