Started in 2016, Power Efficiency and Sustainability is a collaboration between Azure, CO+I, and AHSI to harvest stranded datacenter resources via smart performance-aware power capping and oversubscription. Datacenters often leave resources stranded or underutilized (e.g., space, cooling, networking), because of a perceived lack of available power. To meet increasing capacity demand, more datacenters have to be built. However, allocated power is also underutilized in existing datacenters and power draw spikes rarely, opening the opportunity to improve efficiency by increasing the server density. Our current research efforts focus on three areas:
- Build capping systems that maximize oversubscription within performance envelopes
- Leverage reserve power and mitigate emergencies in the face of failures
- Reduce the granularity of capping and oversubscription to virtual machines, instead of servers
- Leverage ML techniques to prevent performance impact to important workloads
As of June 2023, we have deployed our power capping system to millions of servers across all Microsoft datacenters, and freed up hundreds of MWs of harvested power. Our capping system also allowed Bing and Bing Ads to safely enable turbo boost in their fleet, resulting in ~20% performance improvements. We have also reduced the amount of reserved power in Microsoft’s most modern datacenters to zero, freeing up hundreds of MWs more of power for allocation to customers.
人员
Ricardo Bianchini
Technical Fellow, Corporate Vice President, Microsoft Azure
Alok Kumbhare
Principal Research SDE
Pulkit Misra
Principal Research SDE
Mariana Rodrigues
Senior Research Software Engineer