Making Sense of Multi-Threaded Application Performance at Scale with NonSequitur

on YouTube ↗19m

" summary

NonSequitur compresses multi‑threaded traces with a domain‑specific merge of identical call stacks, leaves long events uncompressed so latency spikes remain visible, and enforces per‑interval size limits so millions of events reduce to a few hundred. Evaluation uses real traces from RocksDB and WiredTiger, a 42‑participant study that reports participants scoring up to 11× higher and spending about 3× less time on some tasks. Compression is lossy, yields non‑linear x‑axes and thread views that require hovering to align corresponding intervals.

AI-written summary · about

" chapters (10)

1. 0:00Performance bugs in multi-threaded apps
2. 1:25Performance comprehension challenges
3. 2:27Limitations of flame graphs and Chrome Trace
4. 3:52NonSequitur’s compression-based approach
5. 5:40Call stack compression trade-offs
6. 7:27Visualization with long event thresholds
7. 9:36NonSequitur vs alternative tools
8. 12:46Thread activity comprehension study
9. 13:51Findings and tool capabilities
10. 14:13Q&A

" related talks

• 01h 02mWriting high performance code in .NETBart De Smet · NDC Sydney 2017
• 32mSoftware Performance: A Shape Not a NumberKay Ousterhout · Strange Loop 2018
• 42mWhen the OS gets in the wayMark Price · Strange Loop 2015
• 48mFantastic Performance and Where to Find ItRichard Warburton · JAX London 2019
• 15mMimesys: Generating Realistic Executable Testing Environments from Resource Usage TracesAditya Akella, Joydeep Biswas, Daehyeok Kim, Donghyun Kim, Zichao Hu · 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26)
• 24mDIY::Thread.profile -- Light-Weight Profiling in Pure RubyJason R. Clark · MountainWest RubyConf 2013
• 49mResponding in a Timely MannerMartin Thompson · GOTO Copenhagen 2014
• 27mNeed for SPEED: Site Performance Efficiency, Evaluation and DecisionKingsum Chow, Zhihao Chang · SREcon21