Using openshift v3.10.0+0c4577e-1 I'm seeing uneven loading across 4 VMs hosting all my pods which are endpoints to a single service route.

I've setup a pretty straight forward testing environment and I'm seeing some behavior that doesn't make any sense to me.

I've got a physical host setup running JMeter configured to make requests to a single service route IP backed by a collection of pods. The pods are hosting a very lightweight Wordpress site. Each pod is identical for the purposes of the test I'm doing.

The pods are split among 4 VMs. All 4 VMs are running on a single separate(from the JMeter host) physical host.

As I scaled out the number of VMs each additional VM continues to be less and less loaded. When I moved from 2 to 3 VMs (50% more VMs) I only saw a 33% improvement in the number of requests I could handle, moving from 3 to 4 VMs (33% more VMs) the improvement was around 18%. I'm not anticipating perfect scaling, but the graphs below don't make sense to me:

As you can see the first VM is using 100% of the available CPU (8 vCPUs), but each VM after uses less and less CPU. The 4th VM is down to ~75%. I'd expect the loading to be a lot closer.

The graphs show two tests, the first was with the roundrobin strategy, the second I switched to leastconn

Is this a problem with the load balancing strategies? Is there someway to better balance the requests across the VMs?