Per-Application Power Delivery

Datacenter servers are often under-provisioned for peak power consumption due to the substantial cost of providing power. When there is insufficient power for the workload, servers can lower voltage and frequency levels to reduce consumption, but at the cost of performance. Current processors provide power limiting mechanisms, but they generally apply uniformly to all CPUs on a chip. For servers running heterogeneous jobs, though, it is necessary to differentiate the power provided to different jobs. This prevents interference when a job may be throttled by another job hitting a power limit. While some recent CPUs support per-CPU power management, there are no clear policies on how to distribute power between applications. Current hardware power limiters, such as Intel's RAPL throttle the fastest core first, which harms high-priority applications. In this work, we propose and evaluate priority-based and share-based policies to deliver differential power to applications executing on a single socket in a server. For share-based policies, we design and evaluate policies using shares of power, shares of frequency, and shares of performance. These variations have different hardware and software requirements, and different results. Our results show that power shares have the worst performance isolation, and that frequency shares are both simpler and generally perform better than performance shares.