Virtualizing Traffic Shapers for Practical Resource Allocation
summary
Claim: a small set of hardware shapers can emulate many more by periodically mapping many classes onto few shapers using TCAM mapping, adaptive partitioning algorithms (dynamic programming and a greedy approximation), and per-class demand counting. Prototype runs on an Arista 7048 prototype with Python/EOS scripts and microbenchmarks show adaptation costs of a few milliseconds and feasible 50 ms intervals; simulations replaying Bing top-of-rack traces report about 3% error with eight WFQs, and substantial error reduction versus static or random grouping.
AI-written summary · about
chapters (5)
- 0:00Start
- 4:26What Are Shapers
- 6:01High Level Problem
- 11:16Weighted Fader Allocation
- 13:28Evaluation Results
related talks
17mPrecise Data Center Traffic Engineering with Constrained Hardware Resources
16mTowards Optimal Rack-scale μs-level CPU Scheduling through In-Network Workload...
16mSöze: One Network Telemetry Is All You Need for Per-flow Weighted Bandwidth Allocation...
16mManaging Congestion Control Heterogeneity on the Internet with Approximate Performance...
25mApproximating Fair Queueing on Reconfigurable Switches
17mNetHint: White-Box Networking for Multi-Tenant Data Centers
17mTowards provably performant congestion control
22mAdapting TCP for Reconfigurable Datacenter Networks