Hmm. I had high hopes on the coalescence being the issue. Did you do another vsish after the config settings changes, to make sure that the problematic host wasn't showing the coalescence settings anymore?
I'm losing the bubble, but are you using the vmxnet3 driver in your VMs? If so, one thing to try would be switching to the E1000 vNIC on the VMs on the problematic host. While vmxnet3 has a lot of advantages, it IS still less tested than the native E1000 drivers that come with the guest OSes (e.g., we just stumbled across a quite interesting defect in vmxnet3 concerning > 32 multicasts concurrently).
If E1000 "fixes" it, then it becomes a question of whether you want to stick with the E1000, or do more work to nail down exactly what feature/capability is problematic in the vmxnet3 driver, so you can try to configure it differently or turn it off. I don't know how much the advantages of vmxnet3 (lowered CPU, performance increase) matter in your case.
Also to consider - are you at the most recent vSphere version that you can be? the NIC drivers used at the host level get regularly updated, and there might be something at that level.