A novel Internet access service with online traffic engineering of elephant flows

Elephant flows, which are high-rate large-sized flows, can cause increased packet delays and losses in other flows. Large enterprises often upgrade their access links even before the load reaches a high level to avoid service degradations caused by elephant flows. This work proposes and demonstrates an alternative solution to access-link upgrades. The solution consists of deploying a high-speed Elephant Flow Traffic Engineering System (EFTES) that monitors access-link traffic, identifies elephant flows in real time, and instructs the router to add a firewall filter to isolate identified elephant flows to a separate queue. The EFTES algorithms were implemented in an innovative highspeed, multicore system called R-Scope, which leveraged features of high-performance network interface cards such as packet coalescence and kernel bypass. This EFTES implementation was evaluated using real 10-Gbps Center for Applied Internet Data Analysis (CAIDA) traffic traces. EFTES could handle a 1-min traffic trace that contained 1.5 M flows with 8 cores. Experiments were conducted on the NSF GENI and Chameleon testbeds at 1 and 10 Gbps, respectively, to demonstrate the value offered by EFTES to delay-sensitive flows in the presence of artificial elephant flows that were added to a background replay of the CAIDA packet traces.