VMware vmxnet3 driver + multiple CPU sockets = poor performance

Performance testing of OpenEdge 11.5.1.044 running on a Red Hat 7.4 VM which sits on top of VMware ESXi 6.0 shows random networking pauses of between 2 and 30 seconds on the CONNECT and DISCONNECT statements.

Physical Server running VMware ESX 6 = HP DL560 with 40 CPUs and 512 GB of RAM.

VM configuration = Red Hat 7.4 with 32 CPU spread over 4 virtual sockets (initial test with 8 CPU on 1 virtual socket performed well) and 480 GB of RAM.

Each physical server is running only one Red Hat 7.4 VM.

All OpenEdge brokers are spawned and started so that the possibility that a new broker was starting up is removed as a possible cause of the delay.

Test code uses a simple repeat loop which does a connect to the database using -H with the IP address and -S with a port number followed by a disconnect.

The network delay disappears if strace is run on the VM which is hosting the OpenEdge database.
 

It was determined that there is a problem with the VMware vmxnet3 network driver when a VM is configured to use multiple cores spread across multiple virtual sockets.  Both the client and server side processes in OpenEdge were waiting for packets which had been sent but not received on the other end.

Upgrade to VMware ESXi 6.5 or later.