Fast and Efficient Container Startup at the Edge via Dependency Scheduling
summary
Container startup time is dominated by image pulls (average ~20s from a cloud repo), so the system treats image layers as dependency units and schedules pods to nodes that already hold the largest total size of matching layers. Implementation within Kubernetes adds a few hundred lines, a node-side layer tracker and scheduler logic, and is evaluated with 50k images profiled plus synthetic requests and simulations, showing 1.4–2.3x startup speedup and reduced disk and CPU overhead; the change is merged into mainline Kubernetes.
AI-written summary · about
chapters (6)
related talks
38mHow Can Pod Start-up Be Accelerated on Nodes in Large Clusters?
35mFrom Minutes to Seconds: How we reduced ML models cold start time by 90% on Kubernetes
30mWhen Things Get Big
30mSpeeding Up Analysis Pipelines with Remote Container Images
34mWhere's My Pod, Already? A Tale of Cold-Start Struggles on Kubernetes
30mMaximizing Launch Reliability: Controlled Lift-off for Reliable Application Startup on…
34mPanel: Improving and Managing Kubernetes at Scale
08m⚡ Fast Restarts, Not Just Fast Starts: Accelerating Pod Recovery