>>>One day I thought outside normal frames, created for each VM a special directory
>>outside all VHD on the host machine, cut the RAM of each VM by more than half and
>>started & accessed all tables on the host machine. As no LAN wires stemmed data flow,
>>the host machine now had excess memory to buffer data access for all 3 VM and
>>this used the physical RAM much better than "fixed" RAM buffering the "local" vhd data.
>>
>>>Benefit of buffering on host gave a stronger boost than piping the data through
>>TCP/IP stack without physical cables added as overhead....
>>
>>But, IIRC, you had to employ the TCP/IP to access host machine from inside the VM, cableless and simulated probably, but still used. The overhead you avoided seems to be the emulated disk access inside the VM and its memory management.
>
>Yupp. But the speed benefit resulted IMO from option to pool disk caching on host for all VMs.

By adding network shares to your host you effectively added a RAMdisk file server to the hypervisor's vswitch/virtual network. Network access to that from your VMs can be fast if the VLAN is private and doesn't interact with any physical network hardware.

In this sort of situation it can be even more important to remember relatively simple things such as enabling disk write caching. Yes, you do need some protection against power failures (laptop's battery, or a UPS for a desktop machine) but gains can be considerable, especially for anyone still using spinning rust.