Routing Mesh / VIP slow compared to DNSRR

Not sure where to go on this issue I encountered… hoping someone can help.

Running latest Docker EE on Windows Server 2019 as a single node swarm. If I deploy a service to the default ingress network with default options aside from publishing a port, the service works as expected.

If I deploy the same service but via dnsrr (not using the default ingress network) and attach an nginx container to act as a reverse stream proxy, the service works as expected.

The problem is performance. When deployed to the default ingress network and publishing a port, the operation I’m testing takes approximately 2 seconds. When deployed to the non-default network with nginx container attached to proxy, the operation takes 1 second.

Timing of the operation within the bounds of the container it is routed to. I’m not sure if ingress or egress (or both) is where the delay happens.

Any tips to debug this? Has anyone else experienced a similar performance issue with the default endpoint mode? Ideally, I’d like to just use the default ingress network with vip endpoint-mode and publish a port as this is a simpler configuration but double the response time is too much of a cost.

If you’re using DNSRR are you checking that your replicas are evenly hit? I am thinking that the calling service may be using only one of the replicas because it got only one DNS.