Performance Interfaces for Network Functions
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
Proposes concrete “performance interfaces” for network functions and implements them with PIX, a tool that extracts programmatic latency expressions and free variables called performance critical variables from NF C code. Evaluated on 12 DPDK/ebpf NFs (including production NFs like CATRON), interfaces are 100–1000× simpler, predict latency with ~8–10% error across deployments (tails rise to ~15–25%), and extract in minutes (most <5 min, complex ones 20–32 min). Notes limits: single-threaded code, no queuing or interference modeling, relies on symbolic execution and an LLC-miss hardware model.
AI-written summary · about
chapters (15)
- 0:00Intro
- 0:24Semantic Interfaces
- 1:17An Ideal Interface
- 2:00Can there exist a performance interface?
- 2:12Performance Interfaces for NFS
- 2:59Outline
- 8:17What does a performance interface look like?
- 8:53Developer: Identify latency regressions Perf before Perfatter Performance
- 9:37Operator: Root-cause diagnosis
- 10:27Performance Interface extractor (PIX)
- 11:30PIX Overview
- 12:17PIX Back-end
- 14:04Evaluation
- 15:33Prediction accuracy across latency percentiles
- 16:15Time required to extract interfaces
related talks
26mPerformance Contracts for Software Network Functions
02h 53mPreview Sessions
13mLive in the Express Lane
25mDPDK performance optimization for edge computing scenario
31mA Practical, Research‑Backed Introduction To Cloud‑Native Kernel‑Bypass Networking
34mComprehensive Performance Benchmark on Various Well-Known CNI Plugins
15mFormal Methods for Network Performance Analysis
37mEfficient and Flexible Virtual Machine Networking Through eBPF