Improving Distributed Filesystem Performance by Combining Replica and Network Path Selection

Abstract

Distributed filesystems are often the primary bandwidth consumers of large-scale datacenter networks. Unsurprisingly, the datacenter network is often the performance bottleneck for distributed filesystems. Yet even with this close relationship, current distributed filesystems and networks are designed independently and communicate over narrow interfaces that expose only their basic functionalities. Even network-aware distributed filesystems only make use of rudimentary network information, and are not reciprocally involved in making network decisions that affect filesystem performance.

In this thesis, we introduce Mayflower, a new distributed filesystem co-designed with the control plane of its underlying datacenter network. This design approach enables Mayflower to combine both filesystem and network information to make replica selection and dynamic flow scheduling decisions. By having more information and controlling both the filesystem and the network, Mayflower can perform optimizations that are unavailable to conventional distributed filesystems and network control planes. Our evaluation results using a real implementation show that Mayflower reduces average read completion time by more than 60% compared to HDFS with ECMP.