NetHint: White-Box Networking for Multi-Tenant Data Centers
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
NetHint exposes three lightweight hints — a hierarchical virtual topology, per-link residual bandwidth and number of competing flows, and a 100 ms update period via two-level allgather — so tenants compute topology- and congestion-aware transfer schedules without per-tenant probing. Evaluation uses a testbed emulating 240 racks (10 ms collection latency) and simulations with 3× oversubscription, showing improved completion times: 2.7× for allreduce, 1.5× for broadcast, and 1.2× for MapReduce shuffle versus the black-box baseline; code is open source.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:37Data-Intensive Applications Are Movin The Cloud
- 1:01Today's Cloud Offers a "Black-Box" Ab
- 2:17Data-Intensive Applications Can Adapt
- 3:08Data-Intensive Applications Have Incer Adapt Traffic
- 4:48Mismatch!
- 6:05Strawman White-Box Solution
- 7:24NetHint Overview
- 9:49What is in the Hint? - Cont'd
- 11:33Timely NetHint with Low Cost
- 12:31Overhead of NetHint's Monitoring Pla
- 13:04Adapting Transfer Schedules with Net
- 13:35Other Questions to Answer
- 14:13Evaluation
- 14:48NetHint on Testbed
- 16:04Summary
related talks
16mSöze: One Network Telemetry Is All You Need for Per-flow Weighted Bandwidth Allocation...
29mFinding the Needle in a Haystack: Tenant-Level Network...
18mMeasuring Congestion in High Performance Datacenter Interconnects
20mNetKernel: Making Network Stack Part of the Virtualized Infrastructure
30mThe Gashlycrumb Tinies of AI Networking You Must Know (or Languish!)
16mCredence: Augmenting Datacenter Switch Buffer Sharing with ML Predictions
25mVirtualizing Traffic Shapers for Practical Resource Allocation
13mContracting Wide-area Network Topologies to Solve Flow Problems Quickly