End-to-end I/O Monitoring on a Leading Supercomputer
summary
Beacon implements end-to-end I/O tracing on the 40,960‑node Sunway TaihuLight, collecting per‑job traces from compute, forwarding, storage and metadata servers and aggregating ~100 GB raw/day with aggressive online+offline compression into about 10 TB over 18 months. It links jobs via scheduler and Lustre mappings, runs on Elasticsearch+Redis, detects modes (N-to-1, N-to-M), analyzes 116,000 jobs ≥32 nodes, exposes cache‑thrash and mapping over‑provisioning, drives fixes, and measures added cost at <1% overhead while operating at production scale on TaihuLight.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:24Supercomputer IO
- 1:57Backend IO
- 3:04Beacon II
- 4:25Cluster Architecture
- 6:05Beacon System
- 8:49IO Performance Analysis
- 11:14IO Modes
- 12:56IO Volume Comparison
- 16:22Hidden Problem
- 17:44Performance Overhead
- 18:56Conclusion
related talks
20mFull Lifecycle Data Analysis on a Large-scale and Leadership Supercomputer...
29mAutomatic, Application-Aware I/O Forwarding Resource Allocation
01h 00mSystems @Scale Summer Q&A Week 4
27mInsightful Traffic Monitoring: Harnessing Cilium for Comprehensive Network Observability
01h 06mBest practices when accessing Big Data or any other data!
29mFinding the Needle in a Haystack: Tenant-Level Network...
21mHigh Network Reliability & Availability in FE & BE for Scalable Training Solutions
41mScaling Observability: From Chaos to Clarity with Open Source Tools
