Sprint ability: How well does your software exploit bursts in processing capacity?

Abstract

Computer systems constrain their processing rates to stay within power, cost and answer quality budgets. Sprinting mechanisms increase processing rates by exceeding budgets for short bursts before reverting back to safe processing rates. Sprinting mechanisms can speed up query processing and reduce queuing delay, but it is challenging to set policies on when and how hard to sprint. This paper discusses sprint ability, i.e., performance under a set sprinting policy divided by the best performance achieved by any competing policy. System managers can use sprint ability to diagnose slowdown caused by poor sprinting policies. As sprinting mechanisms proliferate, system managers will need tools to measure and manage sprint ability, creating new research problems. For example, new techniques to model response time under various sprinting mechanisms and policies will be needed. Approaches to adapt sprinting policies as workloads and underlying systems change will also be needed. For this paper, we provided a first step toward these problems by building a simulator that models response time for Internet services, as an efficient mean to explore the large parameter space of sprinting policies. We set up our simulator to capture key features of sprinting policies, i.e., sprinting frequency and magnitude, studied in recent papers: ApproxHadoop and Adrenaline. In our tests, the policies proposed in those papers achieved only 71% and 75% sprint ability. Further, the difference between best and worst policies varied across sprinting mechanisms. These results confirm the need for research on techniques to efficiently manage sprinting mechanisms.