Hi I was just really into the fact the you can build a cluster using your old computers, and I was wondering...say if you were to perform a task on the host computer, does it mean that the task is autmatically split into little bits and sent to other computers to process the information and sends it back to the host?? or do you need very specific programms to do that?? what what kind of programms are they apart from science and research programs??
Oh one more thing can you create a network of both PPC computers and x86 computers together to make a cluster??
Thanks

Most multithreaded applications can be split across multiple machines in a cluster. A unithreaded application generally cannot be.

Just to add on to this: it depends onthe application. Some multithreaded applications can be made to run on more than one machine easily. But if the application relies on shared memory or another form of intra-machine IPC, porting it becomes a lot more difficult. Most parallel programs are written that way explicitly using MPI or PVM to save the programmer from having to implement low level communication routines. You can also use OpenMosix which simply migrates processes but does not provide for explicit parallel communication so much.

You can network together x86s and PPCs in the same cluster, but given that the same code won't run on both systems (unless you use a scripting language, Java, or some other CPU neutral sort of code, which is a bad idea if you really want performance) there's no much of a point. Plus cluster nodes generally are meant to be as close in speed as possible to make load-balancing easy.

You *CAN* cluster x86 and ppc using some proprietary software, but not in the traditional manner. You can perform more of a distributed computing system (like distributed.net) does. This is, however, taking a look at the problem from a whole different perspective.

Called a Beowulf, its a neat idea, though with standard x86 computers, especially older ones, you aren't really going to see an appreciable gain in preformance, even multithreaded ones. The main problem is this, you need to have computers that are almost if not totally identical, reason being that if you have computers that aren't then the processing load bearing is unbalanced. You'll have one or a couple of the faster computers doing most of the work, while the slower ones just sit on the network and process little bits. Thing about that is that they contribute to lag in the system because of the addressing needed to have everycomputer know that they are there.

Another big thing to consider is the interlink. I'm assuming that you will be using Cat5 cable with a 100 switch, or rack switch depending on your fundage. The problem with this solution over proprietary ones like Inifiniband or some of the stuff that Cray has to offer, is the massive lag that you are going to see. Usually, they use a standard IP addressing on a special network with the nodes, all that passing of data and ipaddressing and caching of ipaddresses, combined with the physical limitations of the cable can lead to a really laggy multicpu setup. You'll have a "supercomputer" technically, but chances are that you'll have a better experience with a single modern cpu. Lag and latency on that are limited to your bus and local bottlenecks, not network bottlenecks, which can really add up.


Check out the Stone Soup Computer, just google that. And the Big Mac at Virginia Tech. An x86 and a PPC beowulf, respectively.

You *CAN* make use of slower computers, provided that your computational units are proportionately larger. And you can load-balance across a heterogenous network, it just takes a bit of extra work.