Near-Optimal Datacenter Transport Protocol

Project goals:

  • Achieves high throughput (or network utilization).
  • Maintains low latency for short flows.

Generally speaking, current solutions fall into two catergories:

  • Rate control protocols (e.g. TCP, DCTCP, DCQCN, HPCC)
  • Admission control protocols (e.g. pHost, NDP, Homa)

However, none of them achieve the goal mentioned above.

Problems

1. Rate Control: high latency for short flows

  • Requires congestion signals to tune the flow rate, which are not timely: it usually takes Multiple RTTs to converge to “right” rates impacting short flows during the convergence period.
  • Becomes even less effective, as bandwidth continuely increases (we are at 100Gbps!): short flows transmission has already been done before sources get the signals from the network.

2. Admission Control: Near-Zero Utilization

Existed designs (e.g. pHost, NDP, Homa) desire to be “coordination-free” : end hosts make admission control decisions without coordinating with other end hosts.

  • Theoretically, due to no coorination, the worst case network utilization is , where n is the number of end hosts in the network; as n increases, the utilization drops to near zero.
  • We also show this phenomenon using unmodified simulators from these designs.

Papers

  • Under submission.

Download

(To be updated..)

Members