David Klee discusses an important hypervisor-level metric:
VMware’s CPU Co-Stop metric shows you the amount of time that a parallelized request spends trying to line up the vCPU schedulers for the simultaneous execution of a task on multiple vCPUs. It’s measured in milliseconds spent in the queue per vCPU per polling interval. Higher is bad. Very bad. The operating system is constantly reviewing the running processes, and checking their runtime states. It can detect that a CPU isn’t keeping up with the others, and might actually flag a CPU is actually BAD if it can’t keep up and the difference is too great.
This is extremely useful information for DBAs in virtualized environments. My crude and overly simplistic answer is, don’t over-book vCPUs on hosts running important VMs like your SQL Server instances.