Distributed Traffic Flow Consolidation for Power Efficiency of Large-Scale Data Center Network

Abstract

Power optimization for data center networks (DCNs) has recently received increasing research attention, since a DCN can account for 10 to 20 percent of the total power consumption of a data center. An effective power-saving approach for DCNs is traffic consolidation, which consolidates traffic flows onto a small set of links and switches such that unused network devices can be shut down dynamically for power savings. While this approach has shown great promise, existing solutions are mostly centralized and do not scale well for large-scale DCNs. In this article, we propose DISCO, a <b><u>DIS</u></b> tributed traffic flow <b><u>CO</u></b> nsolidation framework, with correlation analysis and delay constraints, for large-scale data center network. DISCO features two distributed traffic consolidation algorithms that provide different trade-offs (as desired by different DCN architectures) between scalability, power savings, and network performance. First, a flow-based algorithm is proposed to conduct consolidation for each flow individually, with greatly improved scalability. Second, an even more scalable switch-based algorithm is proposed to consolidate flows on each individual switch in a distributed fashion. We evaluate the DISCO algorithms both on a hardware testbed and in large-scale simulations with real DCN traces. The results show that, compared with state-of-the-art centralized solutions, DISCO can achieve nearly the same power savings while decomposing the global problem into sub-problems that are three orders of magnitude smaller. As a result, DISCO can run <inline-formula><tex-math notation="LaTeX">$\mathbf {10^4}$</tex-math></inline-formula> to <inline-formula><tex-math notation="LaTeX">$\mathbf {10^6}$</tex-math></inline-formula> times faster for a DCN at the scale of 10K servers. The convergence of DISCO has also been proven theoretically and examined experimentally.