An edge-queued datagram service for all datacenter traffic
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
EQDS claims to move queuing to senders and enforce delivery via receiver-driven grants, implemented with either a packet trimming backend (Intel Tofino, Nvidia, Broadcom) or an RTS fallback. Measured results: cuts TCP flow completion time ~2×, boosts NVMeOF throughput ~30%, single x86 core DPDK implementation saturates 100 Gb/s with 3 KB packets, kernel path ~50 Gb/s, Amazon test on 1,000 VMs shows baseline 9 ms incast buffering versus ~20 µs median added latency with EQDS, and 12-packet per-port buffers suffice with trimming.
AI-written summary · about
chapters (15)
- 0:00Intro
- 2:30Datacenter networking toda
- 3:51EQDS: an edge-queued datagram
- 4:56Edge Queued Datagram Service
- 6:03EQDS concept: reordering & retransmissions
- 7:15EQDS concept: queuing disciplines
- 7:58EQDS building blocks
- 8:17NDP backend: small buffers + packet to
- 10:01RTS backend (1RMA) - no switch supp
- 10:34EQDS host API and tunnel protocol
- 12:06Deploying EQDS
- 12:44Disaggregated storage: reducing flow ca
- 13:22Incast: reducing latency at scale
- 14:09Sharing in clouds What happens if different VMs use conflictin
- 14:36Edge-Queued Datagram Service (EQDS) move buffering out of the network, into the en
related talks
25mEyeQ: Practical Network Performance Isolation at the Edge
33mBringing packet queueing to XDP
17mWhite-Boxing RDMA with Packet-Granular Software Control
34mCoupling Eventdev Usage with Traffic Metering & Policing (QoS)
22mLIVE Q&A #3
46mNetwork Traffic Quality of Service (QoS) Classes for Containers & Pods Via EBPF/XDP
25mKeynote: Network Transformation: Are We Ready for 5G and Edge Clouds?
38mTelecom Packet Processing and Correlation Engine Using DPDK