GreenMap: MapReduce with Ultra High Efficiency Power Delivery

Energy consumption has become a significant fraction of the total cost of ownership of data centers. While much work has focused on improving power efficiency per unit of computation, little attention has been paid to power delivery, which currently wastes 10-20% of total energy consumption even before any computation takes place. A new power delivery architecture using series- stacked servers has recently been proposed in the power community. However, the reduction in power loss depends on the difference in power consumption of the series-stacked servers: The more balanced the computation loads, the more reduction in power conversion loss.

In this preliminary work, we implemented GreenMap, a modified MapReduce framework that assigns tasks in synchronization, and computed the conversion loss based on the measured current profile. At all loads, GreenMap achieves 81x-138x reduction in power conversion loss from the commercial-grade high voltage converter used by data centers, which is equivalent to 15% reduction in total energy consumption. The average response time of GreenMap suffers no degradation when load reaches 0.6 and above, but at loads below 0.6, the response time suffers a 26-42% increase due to task synchronization. For the low-load region, we describe the use of GreenMap with dynamic scaling to achieve a favorable tradeoff between response time and power efficiency.