SPARK: Integrated Resource Allocation in Virtualization-Enabled SAN Data Centers

In this paper, we present a novel framework called SPARK (Stable-Proposals-And-Resource-Knapsacks) for addressing the combined placement of application data and CPU in virtualized SAN data centers. Unlike previous approaches, which address only CPU placement or only storage placement, SPARK enables both in an integratedmanner taking into account the differing degrees of proximity and affinity among the different node pairs in the data center.

SPARK is based on two well-studied problems – Stable Marriage and Knapsacks – and is simple, fast, intuitive, iterative, and versatile. It yields good quality placements – in our synthetic experiments, it is consistently within 4% of the optimal values computed using Linear Programming based methods for a wide range of workloads and experiments. At the same time, it is one to two orders of magnitude faster than LP and scales to much larger problem sizes that arise in practice. In comparison with other natural candidate algorithms, SPARK outperforms them by at least 30-40%, demonstrating faster speed and improved versatility at the same time.

The fast running time of SPARK makes it highly suitable for dealing with dynamic data center complexities: workload surges, growth, node failures and downtimes that need quick response. Its built-in versatility positions it well to accommodate policies and constraints that may arise in practical deployments.