CSR: Small: A Server-Centric Approach to Data Center Networks

University Of California-Los Angeles, Los Angeles CA

Investigators

Abstract

Data centers have become the main networking infrastructure to support online services, ranging from end-user applications (e.g., Web search and IM) to distributed system operations (e.g., GFS and MapReduce). In this project, we explore a different design paradigm to data center networks (DCNs). We take a new server-centric solution approach, in contrast to the switch-centered paradigm. We push intelligence into servers, which generally have more open and standardized hardware architecture and software platforms. Multiple layers of low-cost, commodity off-the-shelf, mini-switches are used to connect servers. We leverage the large number of servers in a data center. Even when each server just adds one more link to the rest of servers, we obtain many links in the DCN system given the large server population. These added links, without using high-end switches but using only mini-switches, provide the foundation for enhancing scalability, inter-server capacity, and fault tolerance. Taking the approach, we can design DCNs with high scalability, enhanced fault-tolerance, and high inter-server capacity, and better support data center applications that exhibit various traffic patterns of one-to-one, one-to-several, one-to-all, and all-to-all communications. Our design also opens more space for stimulating innovations in the DCN software through the more open programming platform on servers, while abandoning expensive, proprietary core switches. The results in scalable and modular data centers and associated management tools are expected to show that, the added complexity into servers will incur only minor systems overhead but produce large performance gains.