Bazaar: Enabling Predictable Performance in Datacenters

Abstract

The disconnect between the resource-level interface exposed by today’s cloud providers and tenant goals hurts both entities. Tenants are encumbered by having to translate their performance and cost goals into the corresponding resource requirements while providers suffer revenue loss due to uninformed resource selection by tenants. We explore this disconnect in the context of data analytics applications and present Bazaar, a system that enables predictable performance for such applications in multi-resource, multi-tenant datacenters. Bazaar allows tenants to express high-level goals and predicts the resources needed to achieve them. Since multiple resource combinations may achieve the same goal, Bazaar chooses the combination most suitable for the provider. Using large-scale simulations and deployment on a small testbed, we demonstrate that Bazaar enables a symbiotic tenant-provider relationship. Tenants achieve their performance goals. At the same time, holistic resource selection benefits providers in the form of increased goodput.