PARALLEL DATA LAB 

PDL Abstract

SOFTScale: Stealing Opportunistically For Transient Scaling

Carnegie Mellon University School of Computer Science Technical Report CMU-CS-12-111R, August 2012.

Anshul Gandhi, Timothy Zhu, Mor Harchol-Balter, Michael Kozuchy

School of Computer Science
Carnegie Mellon University
Pittsburgh, PA 15213

http://www.pdl.cmu.edu/

Dynamic capacity provisioning is a well studied approach to handling gradual changes in data center load. However, abrupt spikes in load are still problematic in that the work in the system rises very quickly during the setup time needed to turn on additional capacity. Performance can be severely affected even if it takes only 5 seconds to bring additional capacity online. In this paper, we propose SOFTScale, an approach to handling load spikes in multi-tier data centers without having to over-provision resources. SOFTScale works by opportunistically stealing resources from other tiers to alleviate the bottleneck tier, even when the tiers are carefully provisioned at capacity. SOFTScale is especially useful during the transient overload periods when additional capacity is being brought online.

Via implementation on a 28-server multi-tier testbed, we investigate a range of possible load spikes, including an artificial doubling or tripling of load, as well as large spikes in real traces. We find that SOFTScale can meet our stringent 95th percentile response time Service Level Agreement goal of 500ms without using any additional resources even under some extreme load spikes that would normally cause the system (without SOFTScale) to exhibit response times as high as 96 seconds.

FULL PAPER: pdf