The combination of virtual machines and multicore CPUs creates a whole new world in which the number of CPUs available can be set by the software. If there are, say, four cores, and each can run, for example, up to eight virtual machines, a single (desktop) CPU can be configured as a 32-node multicomputer if need be, but it can also have fewer CPUs, depending on the software. Never before has it been possible for an application designer to first choose how many CPUs he wants and then write the software accordingly. This is clearly a new phase in computing.

Moreover, virtual machines can share memory. A typical example where this is useful is a single server hosting multiple instances of the same operating systems. All that has to be done is map physical pages into the address spaces of multiple virtual machines. Memory sharing is already available in deduplication solutions. Deduplication does exactly what you think it does: avoids storing the same data twice. It is a fairly common technique in storage systems, but it is now appearing in virtualization as well. In Disco, it was known as transparent page sharing (which requires modification to the guest), while VMware calls it content-based page sharing (which does not require any modification). In general, the technique revolves around scanning the memory of each of the virtual machines on a host and hashing the memory pages. Should some pages produce an identical hash, the system has to first check to see if they really are the same, and if so, deduplicate them, creating one page with the actual content and two references to that page. Since the hypervisor controls the nested (or shadow) page tables, this mapping is straightforward. Of course, when either of the guests modifies a shared page, the change should not be visible in the other virtual machine(s). The trick is to use copy on write so the modified page will be private to the writer.

If virtual machines can share memory, a single computer becomes a virtual multiprocessor. Since all the cores in a multicore chip share the same RAM, a single quad-core chip could easily be configured as a 32-node multiprocessor or a 32-node multicomputer, as needed.

The combination of multicore, virtual machines, hypervisor, and microkernels is going to radically change the way people think about computer systems. Current software cannot deal with the idea of the programmer determining how many CPUs are needed, whether they should be a multicomputer or a multiprocessor, and how minimal kernels of one kind or another fit into the picture. Future software will have to deal with these issues. If you are a computer science or engineering student or professional, you could be the one to sort out all this stuff. Go for it!