Power Provisioning for a Warehouse-sized Computer

Power Provisioning for a Warehouse-sized Computer

What is the Problem?

Power provisioning for datacenters is challenging because there's a significant gap between maximum power draw ratings and actual maximum power consumption, leading to underutilization of power infrastructure. Building datacenter power capacity is expensive in nature, so this gap represents a significant opportunity for cost savings by ammortizing the cost of building out infrastructure.

Summary

The authors present a 6-month study of power usage patterns across large-scale workloads to evaluate opportunities for maximizing power capacity utilization. They find a pattern of significant underutilization, which could be used to deploy additional servers within the same power budget.

Key Insights

• The diversity of workloads across a datacenter, as well as the variability between peak and average power consumption, can be safely exploited to increase server density by better matching power provisioning to actual demand
• Power capping provides a safety mechanism enabling more aggressive deployment with minimal risk

Notable Design Strengths

• Evaluated effectiveness for both well-tuned applications and more realistic workloads
• Large-scale real-world measurements across three major workloads (Websearch, Webmail, MapReduce)

Limitations/Weaknesses

• Wasn't very prescriptive about how to design more cost-effective systems within the context of the findings, just said to reduce idle power
• Cooling infrastructure and other relevant costs are factored out of the analysis

Summary of Key Results

• CPU dynamic voltage/frequency scaling might yield moderate energy savings (up to 23%)
• They were able to host between 7% and 16% more computing equipment for individual (well-tuned) applications, and as much as 39% in a real datacenter running a mix of applications

Open Questions

• Does the flat-tax assumption hold up across all datasets?
• Is there a general framework system designers can use to decrease their idle power consumption to near-zero?