Non-intrusive Performance Management for Computer Services

摘要

Networked computer services are increasingly hosted on shared consolidated physical resources (servers, storage, network) in data centers. Thus, some form of resource control is required to ensure contractual performance targets for service customers under dynamic workload and system conditions. This paper proposes a solution for resource control that maximizes the yield of the performance contracts given the available physical resources, while it does not require any modifications to the clients' and the computing services' software or hardware. Our approach achieves this by manipulating the flow of requests into the service by using one or more proxies between the clients and the service. This paper evaluates Proteus, a prototype implementation of the proposed approach, on two different services: a 3-tier e-commerce system and a networked file service. We show that existing proxies for the two respective protocols (HTTP and NFS RPC) can easily be modified to use Proteus to schedule their requests. Once the modified proxies have been deployed, our approach is transparent to clients and services. Moreover, we show that, in contrast to prior art, our solution (1) is stable when workloads and systems change, (2) automatically tunes itself to different services, (3) can enforce flexible quality of service specifications, and (4) correctly detects and reacts to contention of internal service resources.