Backpressure Flow Control
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
Claims a practical per‑hop flow control design that runs in hardware, implemented in P4 and implemented on Tofino2 and evaluated in packet‑level ns3 simulation. The switch tracks only active flows, assigns empty physical queues first to avoid collisions, propagates queue assignment in packet headers, and pauses upstream queues until all offending packets drain. Results show near‑ideal throughput and tail latency, with 2.3–60× lower tail latency for short flows and 1.6–5× better average completion time for long flows versus deployed end‑to‑end schemes.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:17Lots of Existing Congestion Control Protocols. Congestion control goals
- 1:52E2E Feedback Loops are too slow for • High feedback delay: network round-trip-time (E2E RTT) Acting on stale information can hurt performance
- 2:48Link Speed | Uncontrolled Traffic • No feedback in the first RTT
- 6:27Challenges in per-hop per-flow Flow Co
- 7:34Backpressure Flow Control (BFC)
- 8:24Only Track Active Flows • Active flow: flow with packets queued at the switch
- 10:07Dynamic Queue Assignment
- 11:57Communicate State across Switches
- 14:49Evaluation: Simulation (Incast)
- 16:02Thank You • Per-hop per-flow flow control is great
related talks
17mPyrrha: Congestion-Root-Based Flow Control to Eliminate Head-of-Line Blocking in...
31mFlow-based Packet Processing Framework on DPDK and VPP
19mNet-P4ct: Enhanced WAN Bandwidth Fair Sharing Using P4 Programmable Switches
16mManaging Congestion Control Heterogeneity on the Internet with Approximate Performance...
16mFLB: Fine-grained Load Balancing for Lossless Datacenter Networks
25mApproximating Fair Queueing on Reconfigurable Switches
16mFlow Scheduling with Imprecise Knowledge
18mRevisiting Congestion Control for Lossless Ethernet
