Server Partitioning Without VMs - for Flexibility and Performance

Learn about advanced server partitioning techniques that achieve flexible CPU and memory allocation without the overhead of virtual machines in this 38-minute conference talk from the Linux Foundation. Discover how to efficiently utilize servers with tens or hundreds of CPUs by partitioning them to expose only specific CPU fractions to container sets, thereby improving performance, hardware utilization, and mitigating noisy neighbor problems. Explore dynamic CPU set management including containers that can grow and shrink based on demand, grouping strategies using container names, labels, QoS classes, or namespaces, and pre-allocation of isolated CPUs for latency-critical workloads. Master techniques for allowing containers to burst beyond their partitions when free CPUs are available while maintaining comprehensive observability of cluster partitions, including detailed views of exact CPUs and containers within each partition. Examine practical implementations using NRI plugins and the balloons policy that overcome Kubernetes CPU manager limitations without VM overhead, with applications extending to large VM environments as well.