Hi amitlib and welcome to the community 
I'll directly side step question 1 but try to answer question 2 which might indirectly answer question 1 - confused. Excellent!
Because of the way vNUMA works when you are selecting 1 socket and 20 cores this is effectively disabling it plus it may be reducing CPU scheduling opportunities. Unless restricted by OS licensing, I would always start building the VM as wide and flat, i.e. use sockets before cores. In the case you have mentioned try building the VMs with 20 sockets and then retest. If you have licensing implication that prevents this number of sockets then build the VMs with 2 sockets and 10 cores each which will still honor the underlying NUMA configuration of the server assuming that all other things are built correctly, i.e. identical memory configuration across the memory banks for the pSockets.
I know this probably just for testing but also consider right sizing your VM's, i.e. do they really need 20 vCores each? Probably a little brutal but I start at one and then scale up - obviously you can make some educated guesses on what the minimum number should be, i.e. huge SQL server probably 1 is not a good starting point 
Check here for the performance implications of getting vNUMA wrong and bear in mind this is more from the guest than hypervisor but does illustrate a point: Does corespersocket Affect Performance? - VMware vSphere Blog
Hope this helps.