Power-Optimized Processor Slowdown Control.

For user-serving computational workloads in data centers (like ML/AI inference tasks with e-commerce applications), managing both the latency experienced by the user and the energy consumption of such workloads is essential. The slowdown quality-of-service metric is increasingly used in real-world systems, as it captures the perceived user delay proportional to the expected job service time. In this work, we introduce a framework to control job slowdown in a data center with multi-class workloads, by adjusting the processor rate. Increasing the processor rate decreases slowdown at the cost of additional power; the goal is to select the processor rate to optimize this tradeoff. We present optimal rate control characterizations for both 1) when the sizes of all jobs in the queue are observed and 2) when only the head-of-line job size and the total queue length are observed. We compare the controls analytically and numerically, and demonstrate how this framework can be employed to achieve desired slowdown targets with minimal power.