Computer Science News
SIGCOMM Best Paper Award for Mark Handley
A paper by Prof. Mark Handley of UCL-CS and colleagues at University Politehnica of Bucharest, Romania and University of Cambridge, UK has won the best paper award at the prestigious SIGCOMM 2017 conference.
Their paper, "Re-Architecting Datacenter Networks and Stacks for Low Latency" describes a rather radical new approach to how to coordinate transfers of data among large populations of servers within a data center:
- use switches with tiny buffers;
- let senders blast fast from the start;
- under congestion, have switches trim packets to just headers, which indicate to the receiver which senders wish to send to it;
- and have receivers pull data from the exact senders who have data for them (now that they know who does).
This simple, elegant solution is very different from how data center networks work today:
- larger buffers that cause queues and thus delay;
- incasts of traffic from many senders cause severe congestion at switches;
- and long retransmission backoffs by senders, also leading to intolerably long delays.
It notably incorporates simple design changes to both switch hardware and the protocol stack software at end systems, and thus judiciously applies just the right architectural tweaks to reach a significant new point in the design space for data center congestion control.
The video below from the conference presentation illustrates the mechanism proposed.