[weblacky]

I don't know terribly much after this when it comes to putting together the cluster systems. I think what I'm saying is mostly correct, but I've not had first-hand experience, it's all from reading.

Would you need to understand is the server you got is meant to be part of a cluster. Yes it is a complete computer that can boot the operating system to some form of terminal. As some people have mentioned there are compute nodes and some computer notes may actually contain graphics cards. You don't technically need graphics cards, you can use a terminal to do a command line only style interaction.

But basically you're supposed to have a special NumaLink router to attach these heavy cables too them in a star pattern and then designated a startup sequence to get them all online. They each boot in an order and in the end they look like a single system with the ability to perform process migration across nodes. You could easily find yourself with a system that has 30 or 60 CPU's and 10 gigs of RAM or more.

One thing that makes SGI clusters very very different from Beowolf clusters or other such things is that SGI was one of the few companies that actually creates a single system image system. I don't know if they did this before or after they acquired Cray supercomputer. But it was for the same purpose, that's my understanding.

So when you put an SGI cluster together you don't have 30 systems with two processors, you have a single system with 60 processors. It's the same with the memory although it obviously distinguishes between local memory and remote memory. I'm unsure if the linking technology is a variant of Infiniband or not, it seems to work very similarly. Entire processes can be migrated from node to node, I don't know exactly how threading spreads out but I'm assuming it's something similar. However this does mean that the system is not robust enough to withstand a loss of a node during operation.

Unlike modern cloud computing which is a bunch of conventional hardware with enormous amount of redundancy everywhere and an API layer that allows for sudden disappearances of nodes and redundant data being loaded and continuing, SGI systems work more on restores and checkpoints with ultra-reliable hardware. So they're very inflexible to a sudden machine going down. But from my understanding I don't believe it crashes the entire cluster but I do believe it would affect all processes on that cluster node and errors out the applications involved with that failed node...

You could of course still use technologies like MPI if that's how you wish to write it, but with a single system image a normal threaded application or a multi process application should scale linearly across the system due to the single system image design. This is not easy to achieve and you won't find many other platforms that can do this. But what it allowed was an ease of programming instead of relying on message passing or shared memory within infiniband or some sort of networking technology you were abstracted away from that as if you were actually working with a giant SMP architecture.

So that's one of the good things, unfortunately I don't know much about the MIPS processors to know how to get them to perform more or stream anything efficiently that would make them a good candidate to actually run as a cluster. Within the past two years several nearly complete clusters have come up both on online auction sites and on this site with almost 0 takers. That's not surprising because they take an enormous amount of power and what kind of performance are you getting compared to today's standard servers? Putting together two or three of these notes is a great little project. But don't go nuts thinking that if you look 30+ of them together that suddenly you're going to rival what you could get for today's graphics cards and multi CPU x86 style servers.

One thing you might do really depends on if you're involved in the software of things you could definitely build a three or four node cluster that's racked with a large amount of memory and the fastest processors you could find as a way of being a compilation point for someone in the community to very quickly run nightly builds or something for the new freeware software packages being developed by some of the members. I'm not one of those members but I'm simply stating that.

Outside of compiler access unless you're going to utilize any of the old tools for some special kind of video work or something you may just find at the end of this you hit a wall of what to really do with it. There is a glut of these cluster systems because once they didn't have a chance of outperforming a high-end desktop station for a fraction of the price outside of the novelty of the operating system that really wasn't much about them people were interested in. The rackmount systems aren't particularly visually appealing, most of them did have graphical options and so they could be used like a workstation. They are the only ones that can be linked together in more than just a pair. There's a series of rackmount Tezro stations that are allowed to be put into a pair with a NumaLink cable directly between them. None of the workstations, desktops, can be NumaLinked!!!!

But as I said you could get a pair of rackmount Tezros and link them in a special pairing only without a router. But I've never seen it done recently. Rackmount Tezro are actually quite rare. Some of these cluster nodes can be modded, I don't have any information on that end for you.

What I'd suggest is if you're interested in learning and potentially providing compilation space if you get this one working and feel good about all of it look into getting maybe 2 to 3 additional notes and calling it a day. You should also get your hands on an L2 controller, if you can. Look on the various online auction sites normally people post them for a couple hundred dollars but sometimes someone post for literally $75, that's how I got the one I have. So I don't use mine very much but as a troubleshooting tool because they also link to desktop versions of the same architecture, fuel and Tezro... But they have limited usefulness on those platforms.

The L2 runs an embedded Linux system that basically helps give you a terminal to each of your systems, interfaces with the L1 for the different cluster nodes, as well as provide power up sequence and error storage. I think you can boot a cluster system without one but I think they're supposed to be quite handy so consider trying to get your hand on one after you've gotten your first node to work.