Re-architecting datacenter networks and stacks for low latency and high performance
Authors: Mark Handley (University College London), Costin Raiciu, Alexandru Agache, and Andrei Voinescu (University Politehnica of Bucharest), and Andrew Moore, Gianni Antichi, and Marcin Wójcik (University of Cambridge)
This paper, acknowledging SSICLOPS, was submitted and accepted in Sigcomm 2017, LA, USA and it has got the best paper award. Congratulations to our colleagues!
Modern datacenter networks provide very high capacity via redundant Clos topologies and low switch latency, but trans- port protocols rarely manage to deliver performance match- ing the underlying hardware. We present NDP, a novel data- center transport architecture that achieves both near-optimal completion times for short transfers and high throughput in a wide range of scenarios including incast. NDP builds upon Cut Payload (CP) which cuts packet payloads when switches overflow, but remedies CP’s shortcomings, and implements a novel high performance multipath-aware transport protocol. Headers of packets whose payload was cut due to congestion give the receiver a complete view of instantaneous demand from all senders. NDP is primarily a receiver-driven trans- port protocol, as the receiver is the only entity that can ac- curately manage this demand and prioritize between traffic from different senders during incast scenarios.
We implemented NDP in Linux end systems using DPDK, in a software switch and in hardware switches based on the NetFPGA-SUME platform. We evaluate NDP’s performance both in our implementation and in large-scale simulations. NDP achieves slightly lower short-flow completion times than DCTCP running over lossless Ethernet using PFC, while having better ability to prioritize traffic from stragglers. At the same time, for large transfers in heavily loaded Clos topologies, it can achieve more than 95% of the available network capacity due to its excellent multipath capability, beating DCTCP by approximately 40%.