On Thu, Mar 07, 2002 at 09:45:38PM -0800, Kyndig Renshai wrote: > > I'm using oscar 1.2.1. Linux 7.1. on 3 266mhz and a 166mhz machine. I have tried >2 of the 3 solutions to get all the machines working as a cluster. (In this attempt I >removed the smp kernel rpm and lam rpm and replaced them with i386's. So far while >the 3 266mhz will function correctly, however, they still will not run the tests >(last part of the oscar suit -located in OSCAR_tests). > The 166 hd machine dies at the end of the installation - where the image is >transferred (says something about an illegal instruction after downloading the entire >image) (the partition table is screwed somehow. - I tried reinstalling linux on it - >and it mentioned that as a problem).
The illegal instruction was based on the fact that there is architecture skew accross versions of glibc. In the current version of OSCAR we attempted to make things as easy as possible by autodetecting the architecture on the head node, building the image for that arch, and hiding from the users the select box for that in the image building panel (it is there if you run tksis manually). I think the first 2 things are valid, but we are definitely going to unhide the select box for 1.3, as far more users than we expected are trying to run Pentium 1 class machines with a Pentium 2 head node. I think the group also didn't really think there were many Pentium boxes still out there being used in this environment (the chip did stop production 5 years ago if I remember correctly), which is just ignorance on our part. I don't think that any of the core members have Pentium 1 machines to test on though, so support in that area is only going to be as good as the feedback we get during beta release cycles. > Ok - a possible feature for the oscar team is creating images for pre-686 machines >(which might also include a probe for the types of hd present on those machines >before creating the image. Maybe an ambitious feature might be dual boot images?) >One other suggestion is - the need for a diagram that shows how the components are >interacting. An architecture doc if you don't mind and a small user manual since this >is a system that you've put together. Pictures are worth a billion. We detect the type of hd (scsi or ide) on the head node, and set up the image for that type by default (though it is changeable by the user). Beyond that things should just work. If they don't, I suggest registering a bug against SystemImager (http://sf.net/projects/systemimager) with your hardware and what errors you are getting. > To date I still get the feeling that oscar is intended towards a homogeneous >collection of machines. Of course the team argues against this. A reminder though - >Beowulf clusters were aimed at making use of machines that were tossed aside for >upgrades. I don't see any reason why commodity machines cannot include low end (once >before high end products) as clusters. It is less so now than in the past, but testing does tend to happen in a more homogeneous environment. I just built a cluster from 3 PIII SMP and one much slower PIII UP box in our lab yesterday. The Slower PIII had a different scsi controler than the 3 fast machines, and I think it had different ethernet as well. It all worked, which is something that would not have been true for OSCAR 1.1. > I am currently trying to put together a grid as a class project - and creating a >cluster as a resource was just one aspect of that project. I've created a simple >beowulf cluster the semester before with all the machines - so I had thought that the >automated version would have been a quick way do this over again. However I'm not so >sure. Everything went well I will try solution 3 - which is fixing the partition >table first - then creating an image appropriate for the hd. Hopefully this will >work. Ok time for spring break. Please continue to provide feedback. Hopefully we can help get you up and running. -Sean > Ren > "Brian W. Barrett" <[EMAIL PROTECTED]> wrote: On Thu, 7 Mar 2002, Kyndig Renshai >wrote: > > > 1. Is there an .i386.rpm version for the > > lam-6.5.5-usysv.1oscar.i686.rpm sitting around in your > > cvs? Can someone send me a copy. Is there any harm > > of using the lam-6.5.6-sysv.1.i386.rpm (hq of lam) > > instead? > > There is only one real difference - the LAM/MPI rpm from the LAM/MPI web > page will install into /usr, while the RPM with OSCAR will install in > /opt/lam-6.5.5. > > > 2. I notice this persistent error during installation > > (etc/profile.d/mpi-01lam.sh - file/dir cannot be > > found). I do not know if this could explain why tests > > (OSCAR_tests/test_cluster) would not run. Can someone > > say why - and where I can find this script? Or can > > someone send me a copy. > > Which version of OSCAR are you using? This is a bug in some of the LAM > RPM code. > > Brian > > -- > Brian Barrett > Graduate Student, Open Systems Lab, Indiana University > http://www.osl.iu.edu/~brbarret/ > > > > --------------------------------- > Do You Yahoo!? > Try FREE Yahoo! Mail - the world's greatest free email! -- _______________________________________________________________________ Sean Dague [EMAIL PROTECTED] http://dague.net/sean _______________________________________________________________________ Attachment Converted: "d:\mail\attach\Re [Oscar-users] Creating clus"
