Model-driven placement of compute tasks and data in a networked utility
An important problem in resource management for networked resource-sharing systems is the simultaneous allocation of multiple resources to an application. Self-optimizing systems must co-allocate resources in a way that reconciles competing demands and maximizes global system objectives under dynamic conditions. We propose a simple model-driven approach to estimate the performance of a candidate assignment of resources, and select the best candidate to meet local or global goals. In this work, we address the placement of batch compute tasks and data in a network of compute and storage sites. We use the model to select placements for a set of synthetic benchmarks and a functional MRI processing application. Our experiments show that the model predicts the performance of candidate assignments within 10% of the empirical values.
