SCS Undergraduate Thesis Topics
|Ram Raghunathan||Mor Harchol-Balter||The Design and Implementation of a Power-Aware Load Balancer|
Energy costs for data centers are doubling every five years and have already crossed $19 billion. However, much of this power is wasted as servers are mostly idle. Idle servers can also consume as much as 60% of peak power consumption. We introduce a power management algorithm called AutoScale which reduces power consumption by over 30% while delivering response times that are only slightly longer, and still meet service level agreements. AutoScale works by dynamically provisioning data center capacity as needed. AutoScale is load-oblivious and can also be deployed as a distributed application. It is also computationally cheap. We evaluate AutoScale in a testbed structured as a multi-tier data center. A new benchmark is developed to test jobs which work by performing key-value requests upon a data storage. Emphasis is given on testing by implementation rather than by simulation and on comparison against existing power management techniques.