Re: Computer cluster
I have a lot of old HW complete with CPUs and MOBOs lying around, so got the same idea in my head about a year ago. Did my research and discovered the following:
Originally Posted by Tristan_Williams
1. Though perhaps beyond reach for a newbie, it really isn't that hard for a seasoned Linux veteran to build a beowulf cluster.
2. The problem isn't the physical assembly, the connection topology or gathering together the apps; it's what to do with the cluster once it is built.
3. My initial rationale was to run folding@home and contribute to a good cause. I soon discovered that it is more CPU efficient to run eight separate instances of F@H than one instance in a cluster of eight CPUs.
4. Same is true for virtually all standard apps: they are not designed to run efficiently in a cluster. The only apps that do are exotic and highly specific-use apps like the imaging software in CGI render farms or nuclear explosion modeling on supercomputers.
5. Unless you are prepared to write similar code, your cluster will be more an intellectual curiosity than a real-world tool.
6. Clusters are very finnicky. As only one example, they don't mix and match well. Unless your nodes are absolutely identical, the entire cluster will run at the speed of its weakest node. Or the choke point may be the network switch. Or whichever is the weakest link in your chain.
Here's the small article that started me down the clustering road. It has links to other relevant sites. But peek under the initial surface layers, and you will start discovering the limitations of clusters. I'm still somewhat curious about the concept and may even try it out one day seeing as how I've already got most of the pieces ― but all of that additional inquiry sure took most of the wind out of my sail.
Last edited by DuckHook; June 11th, 2014 at 02:44 AM.
Newb: How far must I jump to clear the ledge halfway down?
Guru: It's bad to jump off cliffs. Let's look at better options.
Newb: Stop harping about "best practices" and just tell me.