Harmony: A Congestion-free Datacenter Architecture
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
Claims a datacenter architecture that guarantees bounded queueing at each switch so messages once sent observe a worst‑case delay and no congestion‑related drops. Implements a distributed RSVP‑style admission control that decomposes physical links into K virtual links and a Δ admission delay, with theory and simulations showing small K keeps queuing ≲10 packets, sustains ~45% load in a representative point, matches FastPass-like delay bounds for loads <0.9, and costs only a tiny additive latency that shrinks with message size.
AI-written summary · about
chapters (5)
- 0:00The challenge of networks
- 2:43Inefficiency of current tools
- 8:39Introducing Harmony
- 13:50Protocol design and results
- 18:36Benefits and future outlook
related talks
16mManaging Congestion Control Heterogeneity on the Internet with Approximate Performance...
18mCCC: Re-architecting Delay-based Congestion Control in Datacenter Networks
12mTwenty Years After: Hierarchical Core-Stateless Fair Queueing
25mApproximating Fair Queueing on Reconfigurable Switches
17mBackdraft: a Lossless Virtual Switch that Prevents the Slow Receiver Problem
22mProgrammable Calendar Queues for High-speed Packet Scheduling
16mAchieving Wire-Latency Storage Systems by Exploiting Hardware ACKs
14mRDC: Energy-Efficient Data Center Network Congestion Relief with Topological