Kubernetes Performance Tuning

Cloud Orchestration - Kubernetes Application Notes 23.7.0

This section provides a configuration example for Kubernetes performance tuning for SR-IOV.

The machine in this example includes the following CPUs:

Copy
Copied!
            

numactl --hardware available: 2 nodes (0-1) node 0 cpus: 0 1 2 3 4 5 12 13 14 15 16 17 node 0 size: 31990 MB node 0 free: 25314 MB node 1 cpus: 6 7 8 9 10 11 18 19 20 21 22 23 node 1 size: 32237 MB node 1 free: 27135 MB node distances: node 0 1  0: 10 21  1: 21 10

3 CPUs are reserved on NUMA, and 1 for the system and for Kubernetes. Edit /var/lib/kubelet/config.yaml:

Copy
Copied!
            

cpuManagerPolicy: static reservedSystemCPUs: 6-8 topologyManagerPolicy: single-numa-node

Using isolcpus kernel boot command-line isolates the CPUs from the kernel scheduler. This will ensure that a user-space process will not be scheduled by the kernel. In this example, CPUs 0-5 and 9-23 are isolated. isolcpus=0,1,2,3,4,5,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24.

Warning

In case of changes in the reservedSystemCPUs or in the cpuManagerPolicy the /var/lib/kubelet/cpu_manager_state should be deleted, and the kubelet should be restarted.

© Copyright 2023, NVIDIA. Last updated on Oct 15, 2023.