Hi Junichi, On Monday 21 May 2001 17:15, Junichi Uekawa wrote: > Do you mean, some architecture where a standard system-wide configuration > file exists so that mpich and dsh can rip off the list of machines from one > location? > > That would be good, and Debianlike.
Yeah, definitely :) Not just mpich and dsh, but any package that is a beowuf program could use that. I have in mind a nice scripting environment not unlike the menu system. Athough a beowulf admin could certainly do all that by himself, some convenience wouldn't hurt. We can also build cool administration utils on top of that. So, how would it look like? We should have an easy way of defining beowulf configurations. Some standard location for beowulf configuration would be the first step. + /etc/beowulf |___ clusters -- names of clusters |___ cluster1 |___ cluster2 Specifying the content of those configuration files would be an important task :) Then, some utility like `update-beowulf' could run package-specific filters. We can also provide system tools in beowulf packages with wrappers that automatically update their nodes list, etc. Inserting some fault tolerance there wouldn't be difficult. We could also determine standard system services such as logging, monitoring and testing for the beowulf clusters. My idea is simply to store configuration in the beowulf "interface" node. Next, we could define what a beowulf configuration is. What configuration and programs beowulf "interface" and "compute" nodes will have. Those configurations would be automated through some packages that oversee tweaking of various parameters. Now that fai is finally released we could also try to hook fai into all that. Of course, debian-installer development is going on and we should see how scriptable that is. We could certainly use a client/server version of debian-installer to do "magic-floppy" or netboot installs. Installation is a very important part of any beowulf setup. So are upgrades, configuration changes etc. I remember how many times nodes have been re-installed or scripts written just to change a small thing at our cluster. Surely these are off the top of my head. :) Kind of a wishlist as a suffering beowulf admin, and would be fun to implement. Looking forward to your suggestions. Thanks,
