Automatic Workload Evaluation to Predict System Behavior (AWE)
Archana Sulochana Ganapathi, Armando Fox and David A. Patterson
RAD Lab Industrial Affiliates, UC Discovery and UC MICRO
The complexity of interactions in a datacenter environment coupled with the high churn rate of datacenter workload and software necessitates sophisticated models for predicting performance. There is a pressing need for tools to help operators predict scaling issues and system behavior to effectively make load balancing and resource scheduling decisions. We have come up with a methodology to address these issues.
We first identify a suitable set of workload features (e.g., URLs or database query operators) so that we can represent each trace element as a point in a high-dimensional feature space. We use kernel functions to define similarity between these points. Similarly, we collect performance metrics from a running system, and map these measurements into a high-dimensional "metric space." We then apply kernel canonical correlation analysis (KCCA)  to project both the workload and performance kernel matrices onto subspaces such that the two sets of projections are maximally correlated. The result of KCCA is a mapping from the transformed workload-feature space to the transformed metric space. We can leverage these mappings to predict performance of unseen workloads.
Once we derive useful models for predicting system behavior, we would like to use this information for conducting additional experiments without recreating the system hardware/software environment. For this purpose, we create a model-driven workload generator and complement it with an application simulator to reproduce system behavior. These tools will allow us to ask "what-if" questions for scaling and configuration management.
- F. R. Bach and M. I. Jordan, "Kernel Independent Component Analysis," International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2003.