Closed-loop Network Performance Monitoring and Diagnosis with SpiderMon
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
SpiderMon detects performance problems by adding an accumulated queuing delay header on data‑plane packets and triggering when that sum exceeds a threshold, then issues spider packets that pull recent per‑switch telemetry from circular buffers. Switches maintain lightweight causality state with a timeout Bloom filter and a purple traffic meter to trace backward paths and heavy senders. Collected epochs are aligned by sequence numbers, a wait‑for graph identifies flows with high blocking degree, and evaluation on Tofino and BMv2 shows high precision/recall with far less telemetry and switch memory overhead.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 0:35Observations
- 1:01Root causes
- 2:01Existing solutions
- 3:51SpiderMon
- 4:22Alwayson Monitor
- 4:56Diagnosis
- 5:32Trigger
- 7:49Timeout Filter
- 8:33Timeout Filter Example
- 9:33Purple Traffic Meter
- 9:46Telemetry Data
- 10:16Aligning Data
- 11:09Wait For Graph
- 11:55Calculate Degree
- 13:05Evaluation
- 14:35Complexity
- 15:15Summary
related talks
21mDistributed Network Monitoring and Debugging with SwitchPointer
20mNetwork Monitor: A Tale of ACKnowledging an Observability Gap
16mMicroMon: A Monitoring Framework for Tackling Distributed Heterogeneity
18mMeasuring Congestion in High Performance Datacenter Interconnects
46mLessons learned from monitoring a large production network
46mSelf-Programming Networks: Network Clock Synchronization, Event Playback and Real-time Action
12mDebugging Transient Faults in Data Centers using Synchronized Network-wide Packet
25mStroboscope: Declarative Network Monitoring on a Budget