Scaling Large Production Clusters with Partitioned Synchronization
summary
Clusters run 100,000 machines and process 3.1–4.4 billion tasks (3.3–4.4 million jobs) per month with 87% of tasks under 10s, so the scheduler is redesigned to avoid global sync bottlenecks. The team builds ParSync — a partitioned synchronization scheme that divides the master state into p ≥ n partitions, round‑robin syncing different partitions, and an adaptive quality‑first/latency‑first policy. Production deployment reaches about 40,000 tasks/sec on 100k machines, cuts conflicts, and preserves backward compatibility.
AI-written summary · about
chapters (15)
- 0:00Intro
- 0:15Scale of computer clusters
- 0:53Job/task statistics in Alibaba
- 1:27Single-master architecture
- 2:02Scheduling objectives
- 2:47Engineering objectives
- 4:14Shared-state architecture: benefits
- 5:02Shared-state architecture: limitations
- 6:03Conflict modeling
- 8:09Light weight simulation: results
- 9:29Observation
- 10:09ParSync
- 11:06Effects of synchronization strategies
- 12:00Adaptive scheduling strategy
- 13:35Conclusions
related talks
15mParaSync: Exploiting Fine-Grained Parallelism for Efficient File Synchronization
27mFrom Thundering Herd to Zero Outages: Building Reliable Inventory Sync
20mFine Grained Replicated State Machines for a Cluster Storage System
16mTowards Optimal Rack-scale μs-level CPU Scheduling through In-Network Workload...
43mScalability Intro – Shyam Jeedigunta, Google & Bob Wise, AWS (Any Skill Level)
22mAsynchronous Computing at Scale: Overview and Learning
30mDeep Dive: Scheduling SIG
01h 00mScaling Batch Applications