Explore a conference talk from NSDI '22 that introduces the concept of performance interfaces for network functions. Delve into how these interfaces can succinctly describe a system's performance, similar to how semantic interfaces describe functionality. Learn about PIX, a tool that automatically extracts performance interfaces from network function implementations. Discover how developers and operators can leverage these interfaces to identify performance regressions, diagnose and fix performance bugs, and assess the latency impact of NIC offloads. Gain insights into the evaluation of PIX on 12 network functions, including production-level examples, and understand how the resulting performance interfaces offer accuracy while being significantly simpler than the original code.
Overview
Syllabus
Intro
Semantic Interfaces
An Ideal Interface
Can there exist a performance interface?
Performance Interfaces for NFS
Outline
What does a performance interface look like?
Developer: Identify latency regressions Perf before Perfatter Performance
Operator: Root-cause diagnosis
Performance Interface extractor (PIX)
PIX Overview
PIX Back-end
Evaluation
Prediction accuracy across latency percentiles
Time required to extract interfaces
Taught by
USENIX
Related Courses
-
Closed-Loop Network Performance Monitoring and Diagnosis with SpiderMon
-
How to Diagnose Nanosecond Network Latencies in Rich End-Host Stacks
-
RingLeader - Efficiently Offloading Intra-Server Orchestration to NICs
-
Formal Methods for Network Performance Analysis
-
Automated Verification of Network Function Binaries
-
Collie - Finding Performance Anomalies in RDMA Subsystems
