Get off my GPU! Better Sharing of AI Accelerators in OpenStack with Blazar Reservations
summary
Blazar, the OpenStack reservation service, is presented as a way to reduce GPU contention in private clouds by allowing users to reserve AI/ML-optimised resources in advance. StackHPC’s work with the upstream community adds support for reserving GPU-accelerated flavors and related performance features (for example CPU pinning) to improve access to and utilization of accelerators in OpenStack.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 8:44What is Blazar?
- 14:09Compute reservations
- 22:25Better GPU access
- 30:59Cloud credit management
- 34:09Next steps and integrations
- 36:43Q&A
related talks
20mThe Coral Reef Cloud: Autoscaling, Reservation, and Preemption
32mSmart GPU Management: Dynamic Pooling, Sharing, and Scheduling for AI Workloads in Kubernetes
16mBOSH & GPU - a Match Made in Heaven
43mSupporting NASA Science with OpenStack
38mStop Allocating GPUs, Start Delivering Intelligence: An Enterprise...
31mAI, CERN, and the Quest for GPU Custody: How CERN Leverages DRA for Efficient GPU Sharing
36mOpenPOWER and GPU Porting Platform for FOSS Using OpenStack at OSUOSL
04mA Cloud Native Stack From Bare Metal To Tokens for Large-Scale AI Inference