From: Jason S. <js...@li...> - 2005-11-22 23:40:27
|
On Tue, Nov 22, 2005 at 01:57:50PM -0800, Craig Hughes wrote: > On Nov 22, 2005, at 1:26 PM, Adam Ernst wrote: > > >For those of us coming in late on this discussion... > > > >What's this project's purpose? How big will these "gumstix > >clusters" be, and what will people do with them that's not doable > >with regular computer clusters? > > Well, I think my main idea initially was to provide some publically > accessible compile farms for building stuff to run on gumstix. That > is, allow native compilation at reasonable speeds (using distcc/ > ccache) for stuff which can't be cross-compiled in the buildroot due > to "issues" with the upstream package. For example stuff like the > perl build really wants to run natively on the machine it's being > compiled for. You can kind of get miniperl to compile with a cross- > compiler, but it's not easy, and miniperl ain't perl. I gave up > entirely on trying to get PHP to cross compile, though I compiled it > natively on a gumstix using an NFS mounted root_fs, and that worked > nicely (but took a long time). I vote for iSCSI for the OS storage, and either iSCSI or AFS for the build storage. I can set up a common AFS root that we can all share for builds. Of course, that assumes that the other sites can talk to Cal quickly enough for that to be practical; otherwise, we should probably use a source control system hosted by sourceforge or another highly available provider logically close to all three sites. We're also going to need an admin mailing list. Any suggestions on who it should host it? > But that's not by any means the only thing one could do with a > gumstix cluster (gumwad?). http://en.wikipedia.org/wiki/Gumwad > The intention is basically to provide the hardware out there to help > people potentially come up with great ideas. I think the main > advantage to a gumstix cluster vs any other kind of cluster, if > there is one, is the power consumption. a 6-node gumstix+netstix > cluster will have about 2.4GHz of (integer) processing power, but > only draw about 12 watts of power. By contrast, your typical > single-CPU desktop will probably be drawing somewhere in the 100-200 > watt range. 6 gumstix+netstix also fits in about 1/50th of the > space of a 1U rackmount enclosure, if you arrange them nicely. I > think if one thinks about power supply issues carefully, you could > probably fit 300 or 400 gumstix into a 1U enclosure, for about > 120-160GHz of CPU, though to do that you'd probably want to not use > the etherstix, but do something a bit smarter by interconnecting the > busses of the PXAs more directly. That SMC ethernet controller chip > sucks a ton of power, and seems to turn most of what it draws into > heat... We're next door to the mechanical engineering building, and I hear they build weird enclosures like that if you ask them nice. If you're serious about doing this, we can build a custom multiport switching power supply around one of the analog IC vendors' chips for each (say) 50 gumstix. There's also the small matter of getting several hundred gumstix for a gumwad, but that's a small implementation detail. > As for what I'll be sending to each of these 3 sites, I haven't > completely made up my mind yet (gotta keep some surprises back), but > it'll probably be something like: > > 6 connex-400xm > 6 netMMC > 1 connex-200 > 1 netDUO > cables, power bricks, etc as makes sense > > The idea would be that the connex200+netDUO could be used as an > access control system to the cluster. The site will need to provide > a >=7 port hub or switch as a backbone, but I suspect each of the 3 > sites probably has some spare network ports lying around which could > be used. You have no idea. May I also suggest some serial hardware for implementing stomith and differentiating between network and node failures? Oh, and several hundred people walk by our lab each day. If you provide me with a nice-looking mount for the cluster, I can mount the cluster in the secure display case outside our lab. -- - Jason Best of all is never to have been born. Second best is to die soon. |