Full Lifecycle Data Analysis on a Large-scale and Leadership Supercomputer...
summary
Six years of 40 TB over six years from Sunway TaihuLight (41,508 nodes) and bacon monitoring (deployed April 2017) analyze 6 million jobs totaling 4 billion core-hours to surface concrete problems: system anomalies and OST “fslow” cause multi-hour outages and up to 2x slowdowns, metadata (getattr ≈9%) drives hotspots while some jobs touch up to 26,000 files, and average job wait time peaks near 4,000s with 3% waiting as long as 80,000s. A Dynamic priority scheduling prototype promotes long-waiting jobs and demonstrably prevents starvation while preserving utilization.
AI-written summary · about
chapters (6)
- 0:00Six years of supercomputer data
- 3:56Bacon IO monitoring
- 7:15System anomalies and utilization
- 11:36Metadata performance and access
- 16:35Job scheduling delays
- 19:42Key takeaways
related talks
25mEnd-to-end I/O Monitoring on a Leading Supercomputer
01h 06mBest practices when accessing Big Data or any other data!
31mFrom OOMs to 2x Throughput: Scaling Multimodal Data Curation on Ray
26mUncovering Access, Reuse, and Sharing Characteristics of I/O-Intensive Files
39mMaturing Your Data Architecture in a Week: How Bluesky Survived
11mProcessing TBs of Astronomy Data at Scale with the AC3 Horizon Research Project
36mRapid Processing of NASA Satellite Data Stored with Ceph
42mMulti-GPU Accelerated Processing of Time-Series Data of Huge Academic