How to diagnose nanosecond network latencies in rich end-host stacks
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
Tool called NSight reconstructs per-message lifetimes across kernel, NIC and user space by reconciling CPU and NIC traces into a single timeline with nanosecond precision. It uses low-overhead CPU hardware profiling so measurements don’t perturb microsecond paths, attributes tail spikes to concrete causes (head‑of‑line blocking, scheduling wakes, queueing) and pinpoints the offending functions and cores. Applied to memcached, the implementation cuts 99.9th‑percentile latency from 2.2 ms to 41 μs and guides iterative configuration fixes.
AI-written summary · about
chapters (8)
- 0:00Introduction to insight
- 1:25Challenges in latency
- 4:01Profiling methodology
- 7:24Message tracking techniques
- 10:41Analyzing latency causes
- 12:20Visualizing performance
- 14:05Case study results
- 16:19Conclusion and availability
related talks
22mLancet: A self-correcting Latency Measuring Tool
46mSelf-Programming Networks: Network Clock Synchronization, Event Playback and Real-time Action
17mUnderstanding and Profiling NVMe-over-TCP Using ntprof
26mSoNIC: Precise Realtime Software Access and Control of Wired Networks
01h 01mThe Performance Investigator's Field Guide
16mScalable Tail Latency Estimation for Data Center Networks
38mTechniques and Tools for a Coherent Discussion about Performance in Complex...
21mTime Synchronization using ML Techniques in Electronic Trading