Successfully reported this slideshow.
Your SlideShare is downloading. ×

Apache YuniKorn (incubating): A big milestone

Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Upcoming SlideShare
Fuel Plugins
Fuel Plugins
Loading in …3
×

Check these out next

1 of 17 Ad

More Related Content

Similar to Apache YuniKorn (incubating): A big milestone (20)

Recently uploaded (20)

Advertisement

Apache YuniKorn (incubating): A big milestone

  1. 1. Apache YuniKorn (incubating): A Big Milestone Thursday, 18 November 2021 Wilfred Spiegelenburg
  2. 2. © 2020 Cloudera, Inc. All rights reserved. 2 AGENDA Welcome Apache YuniKorn (incubating): A Big Milestone Gang Scheduling Kubernetes Support Performance Testing Future Development
  3. 3. © 2020 Cloudera, Inc. All rights reserved. 3 AGENDA Welcome Apache YuniKorn (incubating): A Big Milestone Gang Scheduling Kubernetes Support Performance Testing Future Development
  4. 4. © 2020 Cloudera, Inc. All rights reserved. 4 Apache YuniKorn (incubating): A Big Milestone In progress release: • Apache YuniKorn v0.12 • Renamed from v1.0.0 • Planned release date: December 2021 Next release: • Apache YuniKorn v1.0.0 • Planned release date: February 2022 • Kubernetes dependency and support
  5. 5. © 2020 Cloudera, Inc. All rights reserved. 5 What to expect in the next release ? Apache YuniKorn v0.12: • Kubernetes support added for: v1.20 and v1.21 • Node sorting improvement • Gang scheduling enhancements • Supportability improvements • Scheduler Interface update • Kubernetes dependency upgrade to v1.20 • Bug fixes
  6. 6. © 2020 Cloudera, Inc. All rights reserved. 6 AGENDA Welcome Apache YuniKorn (incubating): A Big Milestone Gang Scheduling Kubernetes Support Performance Testing Future Development
  7. 7. © 2020 Cloudera, Inc. All rights reserved. 7 Gang • Predefined set of resource requests • Defined as part of the application • Guaranteed resources: all or nothing Advantages of gang scheduling • Prevent starvation: FIFO ordering of applications • Predictable performance • Improved autoscaling FEATURE DEEP DIVE: Gang Scheduling Background
  8. 8. © 2020 Cloudera, Inc. All rights reserved. 8 FEATURE DEEP DIVE: Gang Scheduling Spark Example: no gang YuniKorn Spark Application Driver executor executor executor
  9. 9. © 2020 Cloudera, Inc. All rights reserved. 9 FEATURE DEEP DIVE: Gang Scheduling Spark Example: gang YuniKorn Spark Application Driver gang: 3 executors executor executor executor gang 1 gang 2 gang 3
  10. 10. © 2020 Cloudera, Inc. All rights reserved. 10 AGENDA Welcome Apache YuniKorn (incubating): A Big Milestone Gang Scheduling Kubernetes Support Performance Testing Future Development
  11. 11. © 2020 Cloudera, Inc. All rights reserved. 11 YuniKorn v0.12.0: • Support K8s 1.19, 1.20 & 1.21 • Dependency K8s 1.20 YuniKorn v1.0.0: • Support K8s . . . • Dependency K8s 1.22 – Deprecated APIs – Protobuf and gRPC changes • Scheduler Framework KUBERNETES SUPPORT Dependency vs Support
  12. 12. © 2020 Cloudera, Inc. All rights reserved. 12 AGENDA Welcome Apache YuniKorn (incubating): A Big Milestone Gang Scheduling Kubernetes Support Performance Testing Future Development
  13. 13. © 2020 Cloudera, Inc. All rights reserved. 13 PERFORMANCE TESTING YUNIKORN-842 Last published: • based on v0.8.0 (YUNIKORN-33) • Published result (website) • Run by: Alibaba New results: • Based on master (YUNIKORN-842) • Publishing in progress • Run by: National Taichung University of Education
  14. 14. © 2020 Cloudera, Inc. All rights reserved. 14 Scheduler Throughput • 20,000 / 50,000 pods • 2000 / 4000 nodes PERFORMANCE TESTING Preliminary results YuniKorn compared to Default • More than 3x faster • Predicate optimisation possible for even better performance
  15. 15. © 2020 Cloudera, Inc. All rights reserved. 15 AGENDA Welcome Apache YuniKorn (incubating): A Big Milestone Gang Scheduling Kubernetes Support Performance Testing Future Development
  16. 16. © 2020 Cloudera, Inc. All rights reserved. 16 New functionality and improvements: • Kubernetes dependency v1.22 – Indirect dependencies • Admission controller enhancements • Improved recovery stability • Scheduler framework (tech preview) FUTURE DEVELOPMENT Planned v1.0.0
  17. 17. Thank You

×