Applicable versions

Longhorn >= v1.1.2.

Symptoms

Not able to investigate the longhorn-manager performance bottlenecks from the external state of the longhorn processes.

Solution

To invesigate the longhorn-manager performance bottlenecks, the runtime CPU profiling data can be collected by pprof.

1. Forward the port 6060 from the longhorn-manager pod to local port 6060:

   kubectl port-forward ${pod-name} -n longhorn-system 6060:6060
   

2. Collect a 180-second CPU profile:

   wget -O profile.out "http://localhost:6060/debug/pprof/profile?seconds=180"
   

Related information

• Related Longhorn issue: https://github.com/longhorn/longhorn/issues/2715