Data Center Workload Consolidation based on Truncated Singular Value Decomposition of Workload Profiles

Kurzfassung/Abstract

In today’s data centers, typically thousands of enterprise applications with varying workload behaviors are hosted. As energy usage is one of the key cost drivers in data centers, workload consolidation is increasingly used to host multiple applications on a single server, sharing and multiplexing a server’s capacity over time. To minimize the number of required, energyconsuming servers, IT managers need to decide which applications should be combined on which server. For that purpose, typically application workload levels are predicted for a planning period such as a month in a defined granularity (e.g., over 5-minute intervals). Then integer programs are used to minimize the amount of required servers, while for each interval constraints ensure that the aggregated workloads of applications assigned to a server must not exceed a server’s capacity. As such problems are NP-hard and computationally intractable for data centers with hundreds of servers and fine-grained workload data, approximations are applied to find at least a good solution, often abandoning the chance to find the optimum. In this paper we propose a novel approach based on applying Singular Value Decomposition to the workload data to reduce the dimensionality of the problem by capturing workload features in order to make the problem computationally tractable. We interpret the coordinates of the time-series projections along the first right singular vectors as indicators for workload levels and complementarities and propose a model to solve the consolidation problem with these few indicators only. We evaluate the model using industry data.