Network Based Virtual Memory

Consider an enterprise level computer network, with dozens, hundreds, or perhaps even thousands of computers networked together. Each of these computers has a given amount of RAM, and when that memory is full it turns to the rather slow virtual memory system to fill the demand, causing performance to drop quite rapidly. But with Ethernet and fibre channel speeds already in the 10gbps range, and poised to go even higher, memory swapping need no longer be limited by the relatively pokey hard drive interface.

In its simplest form, this idea could be implemented as a server with a large amount of RAM, and a fast connection to all of the computers on the network. Individual computers could request swap space on the server, and while it wouldn't be quite as fast as native RAM, it should be a substantial improvement over caching to a hard drive.

But for every computer requiring more RAM than it has available, there are likely several others that aren't making use of the memory they have. The server could then itself swap out the contents of its own memory to other computers on the network. In essence, the entire pool of RAM installed in all of the computers on the network becomes available to use as swap space.

Of course, networks are inherently unreliable, so it would require some clever data management to keep the engineering workstation designing the critical part under a tight deadline from crashing when the secretary reboots the damn office PC because Word keeps printing garbage. To that end, individual computers would also be responsible for caching memory locally, in case the pages they want no longer exist on the network for some reason when they go back to retrieve them. Although this might seem to negate the performance benefits of this system, it wouldn't really be an issue as long as memory is constantly being cached in the background while other applications are running.

However, this doesn't help much when more active memory is requested than is currently available on the local computer, either because it's all in use or because the system hasn't had a chance to cache the unused pages yet. In that case, the computer could make a special request to the swap server, essentially saying "I need memory RIGHT NOW. Here's a page of memory, but just hold on to it for a bit because I'll want it back very soon." With such a request, the server won't farm out the page to any other computers on the network, but will retain it instead in its own memory. In this way, the computer doesn't have to bother with caching it locally, because it can expect with a high degree of certainty that the page will still be available when needed. However, it's expected that the computer will ask for the page back as soon as possible, and so there's an upper limit on how much memory the server will allow to be stored this way.

Since RAM tends to be the limiting factor in performance—far more these days than processor speed, hard drive space, or any other specification—having such a system in place could extend the useful life of practically every computer on the network. The money saved by not upgrading hardware as often could then be put into upgrading the network, which would in turn improve the performance of all of the individual computers.