Portland State University. Department of Computer Science
Date of Award
Doctor of Philosophy (Ph.D.) in Computer Science
1 online resource (xii, 151 pages)
Computer networks -- Energy conservation, Computer network architectures, Computer networks -- Workload
Present-day datacenter networks (DCNs) are designed to achieve full bisection bandwidth in order to provide high network throughput and server agility. However, the average utilization of typical DCN infrastructure is below 10% for significant time intervals. As a result, energy is wasted during these periods. In this thesis we analyze traffic behavior of datacenter networks using traces as well as simulated models. Based on the insight developed, we present techniques to reduce energy waste by making energy use scale linearly with load. The solutions developed are analyzed via simulations, formal analysis, and prototyping. The impact of our work is significant because the energy savings we obtain for networking infrastructure of DCNs are near optimal.
A key finding of our traffic analysis is that network switch ports within the DCN are grossly under-utilized. Therefore, the first solution we study is to modify the routing within the network to force most traffic to the smallest of switches. This increases the hop count for the traffic but enables the powering off of many switch ports. The exact extent of energy savings is derived and validated using simulations. An alternative strategy we explore in this context is to replace about half the switches with fewer switches that have higher port density. This has the effect of enabling even greater traffic consolidation, thus enabling even more ports to sleep. Finally, we explore a third approach in which we begin with end-to-end traffic models and incrementally build a DCN topology that is optimized for that model. In other words, the network topology is optimized for the potential use of the datacenter. This approach makes sense because, as other researchers have observed, the traffic in a datacenter is heavily dependent on the primary use of the datacenter.
A second line of research we undertake is to merge traffic in the analog domain prior to feeding it to switches. This is accomplished by use of a passive device we call a merge network. Using a merge network enables us to attain linear scaling of energy use with load regardless of datacenter traffic models. The challenge in using such a device is that layer 2 and layer 3 protocols require a one-to-one mapping of hardware addresses to IP (Internet Protocol) addresses. We overcome this problem by building a software shim layer that hides the fact that traffic is being merged. In order to validate the idea of a merge network, we build a simple mere network for gigabit optical interfaces and demonstrate correct operation at line speeds of layer 2 and layer 3 protocols. We also conducted measurements to study how traffic gets mixed in the merge network prior to being fed to the switch. We also show that the merge network uses only a fraction of a watt of power, which makes this a very attractive solution for energy efficiency.
In this research we have developed solutions that enable linear scaling of energy with load in datacenter networks. The different techniques developed have been analyzed via modeling and simulations as well as prototyping. We believe that these solutions can be easily incorporated into future DCNs with little effort.
Yi, Qing, "Power-Aware Datacenter Networking and Optimization" (2017). Dissertations and Theses. Paper 3474.