Sky Computing
summary
Sky Computing proposes an intercloud broker that maps application requirements to services across clouds; prototypes Sky Pilot (ML job broker using Ray) and Sky Plane (data transfer overlay) already run. Results: a three-stage ML pipeline runs 47% faster and 60% cheaper across clouds, a large open LLM is trained on spot instances for $300, and spot-only training cuts cost ~70% with little time loss. Sky Plane outperforms AWS DataSync; multiple smaller clouds (Lambda Labs, IBM, Oracle, Samsung SDS) adopt the system.
AI-written summary · about
chapters (17)
- 0:00Introduction to Sky
- 1:28What is Sky Computing
- 3:36Motivations for Sky
- 4:57Challenges of Intercloud
- 8:33The Sky Approach
- 11:15Machine Learning Example
- 14:23Sky Broker Architecture
- 17:07Why Sky Will Happen
- 24:54Sky Development Philosophy
- 28:51Impact of Sky Computing
- 33:27Sky for Single Clouds
- 35:27SkyPilot Prototype
- 39:28Spot Instance Optimization
- 41:48High-Level Service Example
- 45:26Sky Ecosystem Growth
- 46:09SkyPlane Data Transfer
- 49:55Conclusion and Summary
related talks
16mSky Computing: Charting a New Course in AI Cloud Infrastructure
52mJoint Keynote Address
17mSkyPilot: An Intercloud Broker for Sky Computing
16mSkyplane: Optimizing Transfer Cost and Throughput Using Cloud-Aware Overlays
11mSkyPilot: Simple Orchestration of Workloads
13mSpace Clouds
11mUnlocking the Power of Machine Learning for Cloud Workload Opt.
16mThe Supercloud