About

MicroFuge is a distributed middleware that provides performance isolation for cloud storage system by building a lightweight and accurate performance model of the underlying storage system using measured data, mapping performance isolation requirements into latency-based deadlines, and then adding a deadline-conscious, performance model-driven caching and scheduling layer to the cloud storage stack. Compared to using Memecached, our prototype implementation can reduce the deadline-missed rate of a modified YCSB workload by more than 40% in an EC2 deployment.

System Overview

MicroFuge provides performance isolation through two middleware components. Deadline Cache (DLC) Builds a performance model Uses multiple LRU queues for deadline-aware evictions Deadline Schedule (DLS) Performs intelligent replica selection. Implements feedback-driven deadline-aware scheduling. Optionally performs admission control.

DLC offers adaptive deadline-aware caching Multiple LRU queues enable DLC to perform deadline-aware evictions. Uses Each eviction victim is selected by computing the Modified Recency Value DLC uses an adaptive policy that considers both the client request rate for each deadline range and the underlying system’s performance to update the adaptive divisors.

With MicroFuge’s distributed design, each DLS is responsible for scheduling client access to a subset of distributed data servers. Each scheduler performs three tasks to provide performance isolation in the scheduling layer. DLS will select the replica which is most likely to meet request’s deadline. In order to make the selection, DLS relies on the latency modeling component which uses previous request latencies to predict incoming request’s latency There are cases where server load just exceeds its capacity. DLS additionally provides an optional admission control mechanism which performs early rejection of requests which are likely to miss their deadlines.