Abstracts Category : Other

Centralized performance control for datacenternetworks

by Jonathan Ph Perry

An ideal datacenter network should allow operatorsto specify policy for resource allocation between users orapplications, while providing several properties, including lowmedian and tail latency, high utilization (throughput), andcongestion (loss) avoidance. Current datacenter networks inheritthe principles that went into the design of the Internet, wherepacket transmission and path selection decisions are distributedamong the endpoints and routers, which impede obtaining the desiredproperties. Instead, we propose that a centralized controllershould tightly regulate senders' use of the network according tooperator policy, and evaluate two architectures: Fastpass andFlowtune. In Fastpass, the controller decides when each packetshould be transmitted and what path it should follow. Fastpassincorporates two fast algorithms: the first determines the time atwhich each packet should be transmitted, while the seconddetermines the path to use for that packet. We deployed andevaluated Fastpass in a portion of Facebook's datacenter network.Our results show that Fastpass achieves high throughput comparableto current networks at a 240 x reduction is queue lengths, achievesmuch fairer and consistent flow throughputs than the baseline TCP,scales to schedule 2.21 Terabits/s of traffic in software on eightcores, and achieves a 2.5 x reduction in the number of TCPretransmissions in a latency-sensitive service at Facebook. InFlowtune, congestion control decisions are made at the granularityof a flowlet, not a packet, so allocations change only whenflowlets arrive or leave. The centralized allocator receivesflowlet start and end notifications from endpoints, and computesoptimal rates using a new, fast method for network utilitymaximization. A normalization algorithm ensures allocations do notexceed link capacities. Flowtune updates rate allocations for 4600servers in 31 ps regardless of link capacities. Experiments showthat Flowtune outperforms DCTCP, pFabric, sfqCoDel, and XCP on tailpacket delays in various settings, and converges to optimal rateswithin a few packets rather than over several RTTs. EC2 benchmarksshow a fairer rate allocation than Linux's Cubic.Advisors/Committee Members: Hari Balakrishnan, Devavrat Shah (advisor).