PI is a bear, but as a "virtualization expert" which means I basically can use virtualization acronyms fluently without making too big a fool of myself, I can assure you that installing anything on bare metal will almost certainly slow it down in modern environments... with the exception of CML which is double virtualized (nested KVM in VMware).
If you inspect the virtual memory architecture of modern operating systems, virtualization is an amazing feature since the secondary lookup address table allows something like a monitored cache to occur.
If you want to make a real performance difference, I recommend using Cisco servers with Cisco 10Gbe NICs (think VIC 1225, VIC 1285 or at least P81E) to provide VMFEX to the PI virtual machines. This can actually give you a 15-30% boost in reality if you have a high bandwidth consumption. This is because it will remove the virtualization host from networking and free up massive CPU resources that are wasted by software based networking (hence the reason SDN VMware NSX style is so insanely expensive).
There are some other extreme methods you might look into... for example, consider using hardware accelerated Java. This requires special hardware and doesn't really provide too much benefit anymore now that dynamic recompilation and tracing JIT technology is mature.
I have found that there are some real benefits to replacing classic WLCs with 3850 switches at times, but I'd have to see your actual environment to see where your performance bottleneck is.
Can I recommend that you configure SNMP on Linux and use a monitoring system like Cacti to watch the actual PI server and monitor resource usage for a few days and identify what's actually causing the bottleneck?
