Integrated Networking and Scheduling for Centralized Transaction Servers

Network-attached storage devices provide ways to build large scalable file systems that distribute data and load across multiple network nodes. Even though data is distributed, most file systems require a shared server to manage centralized data for clients. This paper investigates ways to use a scheduler to support latency guarantees for requests to a centralized data server. We develop a simple model for the server, and use a simulator to test scheduling strategies. We find that to guarantee average latency we may use measured latency as criteria for admission control. This method has the additional advantage of automatically adapting queue lengths to load. Using measured latency as an input for scheduling priority also provides an efficient scheduling policy.