Network-Based Data Processing Architecture for Reliable and High-Performance Distributed Storage System

摘要

In the era of post peta scale computing, high-performance and reliable storage systems have become much more important. Close cooperation between network and storage is an emerging issue. This paper proposes a network-based data processing architecture to build reliable and high-performance distributed storage system using future programmable network devices. Distributed storage systems use replication or erasure coding for ensuring reliability. However, they require additional data transfer and computing resources. Satisfying both reliability and performance is an important issue for storage systems. Recent studies related to Software Defined Networking (SDN) imply that programmable network switch will become more functional. Currently, SDN intends to provide a flexible routing mechanism. Network switches are starting to have intelligent mechanisms and are expected to have a capability for data processing. In our proposed architecture, storage controller functionality is offloaded to a programmable network switch to eliminate additional data transfer. We conducted experiments to show an advantage of the proposed network-based data processing mechanisms for erasure coding and show an optimized design for distributed storage systems. With the proposed method, the performance gain of a reliable data storage system is 44% compared with a client compute case.