Vaidyanathan Srinivasan

Method to Fairly Distribute Power Saving Benefits in a Cloud among Various Customers

Cloud computing infrastructure offers new levels of flexibility and efficiency in a datacenter enabling more and more enterprises to adopt cloud and optimally utilize their IT infrastructure. Power management features in the platform and its exploitation through operating system policies enable significant cost savings by efficiently managing the power consumption. Most of the platform power saving features and policies are designed to exploit changes in server utilization. As the utilization decreases, more power saving features are turned-on, thereby reducing the servers power consumption. For a cloud service provider, cost savings due to power saving drives better ROI1. However for a cloud user who pays a fixed charge to the provider based on provisioned hours, there are no cost benefits due to lower power usage. In the absence of any benefit due to power savings getting passed to the end-user, there is less motivation for the cloud user to choose energy efficient policies, even for non-critical workloads. This paper describes a mechanism to pass on the benefits accrued due to power savings from cloud provider to the cloud consumer so as to encourage the consumer to choose power efficient policies and contribute to the overall reduction in power consumption and carbon footprint.

Power Optimization Techniques for DDR3 SDRAM

With memory contributing to a significant fraction of the overall power consumption, several power management techniques targeting the memory sub-system have been proposed by researchers. In this work, we propose two memory power optimization techniques. We first suggest dynamically varying the queue structure in a memory controller, and next propose an adaptive threshold technique for switching to low power SELF-REFRESH operating mode of the DDR3 SDRAM. With the proposed queue-resizing optimization, the power consumption reduces by up to 93%, while the adaptive threshold technique results in an additional 21% power savings, on an average, over a constant threshold implementation.

Power Supply Efficiency Aware Server Allocation in Data Centers

We address the issue of power efficiency at a relatively high level of abstraction -- tasks allocated on server systems. The domain is that of requests/tasks being forwarded to clusters of servers in data centers. We target the power delivery mechanism -- specifically, the power supply unit, and suggest techniques to distribute workload in a way that the total power is reduced by factoring this important parameter. We formulate the overall power minimization as an optimization problem that takes into account the variation of compute power and power supply efficiency with workload. Our proposed heuristic gives results that are close to that returned by standard NLP solvers, but takes a fraction of the execution time, making it a powerful alternative in a practical online workload distribution strategy.