Stuffing 10K Pods on a Commodity Server - And Starting Them in Milliseconds via Kubernetes
summary
Claim: commodity servers host thousands of pods and start them in tens of milliseconds. Implementation uses Unikraft unikernels, a new VM runtime runu, and a Crossplane provider plus a CLI called craft cloud. Measured cold boot times are ~16–30 ms, snapshot wakeups 40–60 ms for heavier apps, and a demo runs ~1,075 nginx unikernel instances on one server scaled to zero (idle instances consume no memory), live demo works.
AI-written summary · about
chapters (10)
- 0:00Introduction and motivation
- 1:28Problems with cloud scaling
- 3:56Cold starts and limitations
- 5:00Overhead in cloud stacks
- 6:42Unikernels as solution
- 7:24What is a unikernel
- 8:06Demo and deployment
- 9:48Scaling to zero demo
- 10:10Running 10K pods
- 11:13Q&A session
related talks
15mLeveraging Unikernels and Kubernetes to (Transparently)...
25mDeploying Unikernels in Production with Kubernetes
39mServerless Platform for Large Scale Mini-Apps: From Knative to Production
34mWhere's My Pod, Already? A Tale of Cold-Start Struggles on Kubernetes
30mWhen Things Get Big
39mDelivering Serverless Experience on Kubernetes: Beyond Web Applications
30mMaximizing Launch Reliability: Controlled Lift-off for Reliable Application Startup on…
25mOne Size Fits None: Lightweight Workload-Specialized Confidential Containers