AMBER Archive (2003)

Subject: Re: AMBER: tru64 alpha

From: Robert Duke (rduke_at_email.unc.edu)
Date: Wed Oct 15 2003 - 22:41:57 CDT


Mu -
I guess I missed the fact that your first email actually had a reference to a pmemd allocation routine in it. SO, it is possible that Bill Ross' comments about ulimit are appropriate, but I don't know what the default for that would be (type ulimit and see what it says; if it is not huge or unlimited, read the man pages and bump up allowed memory). I have not tried using prun at PSC, but instead use qsub (parallel queues submission). Lemieux is an ES45 alphaserver; I don't know what an SC45 is, but I don't have a world of experience on the alphaservers (wish I had one, believe me). With alphaservers that are essentially smp's with large shared memories, there can be problems with running out of memory if you run a really large problem on a large number of processors. Thus, I know I have run a 91K atom problem on 24 processors on salk at psc, but I think I failed to get it to run on 32 processors, and I think it has something to do with the shared memory (I perused man pages and never actually found an option for specifying memory usage in the executable). I don't know if this had anything to do with other things running on the smp. I would try running on 8 processors or so first, and then scaling up, just in case it is something like this (I have seen this on the Cray T3E at PSC and on salk, but nowhere else). In general, you should hit this problem later with pmemd than you do with sander or sander7, since pmemd actually uses only a fraction of the memory used by the sanders (through pairlist compression, among other things). Let me know if it works or fails again.
Regards - Bob
  ----- Original Message -----
  From: Mu Yuguang (Dr)
  To: amber_at_scripps.edu
  Sent: Wednesday, October 15, 2003 10:59 PM
  Subject: RE: AMBER: tru64 alpha

  Dear Robert,

  I try quadrics, because the switch -lelan also used here for other parallel jobs.

  The compilation looks fine.

  When I use prun to submit the job, the errors appeared as I showed in the last mail.

  The machine here is called AlphaServer SC45 having 42*4 processors.

  Maybe the problem is from the prun. What kind of tools you have used fro submitting jobs there ?

  Rgds

  yuguang

  -----Original Message-----
  From: Robert Duke [mailto:rduke_at_email.unc.edu]
  Sent: Thursday, October 16, 2003 10:51 AM
  To: amber_at_scripps.edu
  Subject: Re: AMBER: tru64 alpha

  Mu -

  All the Machine files that contain "axp" somewhere in their name are for the Tru64 OS running on Compaq or HP alphaservers. The generic mpi one is src.pmemd/Machines/Machine.axp_f90_mpi. I have used the Machine.axp_f90_mpi.ev7 file for the salk servers at PSC, and the first file (Machine.axp_f90_mpi) is just a generalization of this one that you can't use on the jonas front-end at PSC because it has a different chip architecture (ie., you can use the generic one if your target machine and build machine have the same chip). There are also axp files with a "quadrics" extension, which you can use if you have a quadrics mpi interconnect between the nodes, as is the case for lemieux.psc.edu, the "terascale computer" at PSC. You should not have any problems with dynamic memory allocation in pmemd. Sorry for the late reply; the UNC email server was dead for 2 days, which kind of backed things up. If you have problems, let me know more about your machine config.

  Regards - Bob Duke

    ----- Original Message -----

    From: Mu Yuguang (Dr)

    To: amber_at_scripps.edu

    Sent: Tuesday, October 14, 2003 9:19 PM

    Subject: AMBER: tru64 alpha

    Dear All,

    Do you have experiences in install PMEMD in tru64 alpha machine ?

    Thanks in advance

    yuguang

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu