Kubernetes Batch Processing at Scale - A Scheduling Perspective
Talkat Kubernetes Batch + HPC Day Europe 2023 · on Tuesday, 2023-04-18· at 15:20· Hall 7, Room A | Ground Floor | Europe Complex·
summary
ByteDance builds a custom Kubernetes scheduler that replaces the default to run hundreds of millions of batch pods daily and a largest cluster with >20,000 nodes. The system uses a dispatcher/scheduler/binder trio for distributed, optimistic-concurrency scheduling, adds gang scheduling and domain resource fairness (DRF), uses node partitioning and result caching to hit ~5000 pods/sec, raises cluster utilization to ~60%, and colocates batch and microservices successfully.
AI-written summary · about
chapters (14)
- 0:00Start
- 2:30Ecosystem
- 4:02Large Scale
- 4:50Evolution
- 7:40Scalability
- 10:20Architecture
- 11:16Scheduling framework
- 12:53Concurrent scheduling
- 13:38Result caching
- 14:24Gang Scheduling
- 15:46Microtopology scheduling
- 16:46Benchmarks
- 17:42Future work
- 18:33Open Source
related talks
25mScale your Batch / Big Data / AI Workloads Beyond the Kubernetes Scheduler
31mHow We Scale up to 2k Nodes for Batch Jobs Using Cluster Autoscaler
29mScale Your Batch / Big Data / AI Workloads Beyond the K8s...
30mDeep Dive: Scheduling SIG
41mHow Kubernetes and Spring Batch Process Over 12,000 Hours of Video per Month
33mBatch Systems in Production with Kueue: Multi-Tenancy and Fungibility
29mData Processing Efficiency: Optimizing Batch Workloads on Kubernetes With Custom Schedulers
23mPractice of Fine-grained Cgroups Resources Scheduling in Kubernetes