Computational sprinting is a class of mechanisms that boost performance but dissipate additional power. We describe a sprinting architecture in which many, independent chip multiprocessors share a power supply and sprints are constrained by the chips' thermal limits and the rack's power limits. Moreover, we present the computational sprinting game, a multi-agent perspective on managing sprints. Strategic agents decide whether to sprint based on application phases and system conditions. The game produces an equilibrium that improves task throughput for data analytics workloads by 4–6x over prior greedy heuristics and performs within 90% of an upper bound on throughput from a globally optimized policy.
Modern datacenters oversubscribe their power supplies to enhance performance and efficiency. A conservative datacenter that deploys servers according to their expected power draw will under-utilize provisioned power, operate power supplies at sub-optimal loads, and forgo opportunities for higher performance. In contrast, efficient datacenters deploy more servers than it can power fully and rely on varying computational load across servers to modulate demand for power.4 Such a strategy requires responsive mechanisms for delivering power to the computation that needs it most.
Computational sprinting is a class of mechanisms that supply additional power for short durations to enhance performance. In chip multiprocessors, for example, sprints activate additional cores and boost their voltage and frequency. Although originally proposed for mobile systems,13, 14 sprinting has found numerous applications in datacenter systems. It can accelerate computation for complex tasks or accommodate transient activity spikes.16, 21
No entries found