Get More Computing Power by Helping the OS Scheduler
Talkat Kubernetes Batch + HPC Day Europe 2022 · on Tuesday, 2022-05-17· at 14:55· Room 4F | Event Center·
summary
Adds CRI Resource Manager (CRI‑RM) to nodes to take over CPU and memory pinning via the CRI host manager, replacing direct kubelet→container runtime calls and raising sustained core turbo by packing threads into fewer cores and aligning pods on the same NUMA/L3 domains. Provides zero-configuration CPU and memory pinning and extensible policies (examples: port boost and balloon pools), implements cache, memory-bandwidth and block‑IO controls, and ships patches into CRI/containerd. Case studies report major improvements in database and AI training workloads without modifying Kubernetes or apps.
AI-written summary · about
chapters (5)
related talks
41mAn Essential Relationship between Real-time and Resource Partitioning
13mHarvesting Spare CPU Resources in Container Systems
22mRestartable Sequences: Scaling Per-Core Shared Memory Use in Containers
44mRestartable Sequences: Scheduler-Aware Scaling of Memory Use on Many-Core Systems
45mDynamic vCPU priority boosting in KVM for latency sensitive workloads
43mExpanding the Memory Limits with CXL Software Stack
01h 36mLinux Memory Management and Containers
37mCRI-O: Under the Hood