Detecting Ephemeral Optical Events with OpTel
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
A centralized vendor-agnostic controller standardizes optical device models, deploys push-based telemetry agents, and uses a cloud-backed collector cluster to stream metrics at one-second granularity. OpTel runs in Tencent’s backbone, detects about 2× more optical events (15% of events last <10s), and classifies ephemeral vs persistent faults; SNMP-style polling spikes device CPU to ~96% at 0.1s while OpTel stays ~19→25%, and troubleshooting that previously took 5–10 minutes now completes in seconds.
AI-written summary · about
chapters (11)
- 0:00Start
- 1:38Optical Failures
- 2:28Talk outline
- 2:42Traditional Telemetry Pipeline
- 6:06Existing Telemetry System
- 6:42OpTel high-level design
- 7:21Design and implementation
- 12:29Data collection overheads
- 13:19Detecting optical events
- 14:16Troubleshooting optical events
- 14:42Concluding summary
related talks
16mOpenOptics: Enabling Open Research and Implementation of Optical Data Center Networks
18mOpenTelemetry and Observability: What, Why, and Why Now?
20mObservability and Open Telemetry
30mTelemetry: Industry Status, Challenges, and OSS Opportunities
19mOpenTelemetry's Path to Observability's Future
37mObservability For You And Me With Opentelemetry
34mPanel: OpenTelemetry: Bridging Platform and Enablement
24mOTel-y Oops: Learning From Our Observability Blunders