2009/9/4 ron minnich <[email protected]>: > On Fri, Sep 4, 2009 at 6:40 PM, Sebastian Lara<[email protected]> wrote: > >> Just because we start using ROCKS Clusters. Are there some cluster >> distributions that can run without local disk on nodes? > > Yes. I strongly suggest you take a look at this: http://onesis.org/ > > No local disks required. How big is your cluster? If less than 128 > nodes, just run NFS root with one NFS root server. onesis is > incredibly clever, in that you can easily configure it to put each > nodes /tmp, /var, and so on in local ramfs or on a local disk. > > We use it to run a 4400 node (not a typo) system at sandia: it scales. > We have many different types of installations, and it runs well one > even very small systems, like my Geode clusters. > > And, it's very network oriented, but allows you to have data on local > disks. I'm really sold on it. One of our newer interns, Chris Kinney, > who is also on this list, can tell you more. He set up an 80-node > cluster, with no previous experience, in an afternoon. > > I think if you went with onesis you could avoid having to refresh your > bios. Your life would be easier. > > For compute node clusters, in fact, the best thing you can do is yank > the disks and throw them away -- unless you need them for local data > storage. They tend to cause trouble. I have not built a disk-based > cluster in 10 years, and I've built clusters that range in size from 4 > nodes to 2048 nodes. Local disks are just trouble. > > While I respect the work the Rocks guys have done, I think onesis is a > good way to go. So does Sun: they use onesis for their commercial > cluster offerings. > > Thanks
This sounds really good. I will definitely try this. Thanks. -- Sebastián Lara Menares Ingeniería Civil Electrónica Universidad de Concepción -- coreboot mailing list: [email protected] http://www.coreboot.org/mailman/listinfo/coreboot
