Gromacs slow.

botheredbothered Manchester UK
edited August 2004 in Folding@Home
Why are gromacs so slow? I mean it's just maths. What do they do differantly that takes three times as long to complete?

Comments

  • shwaipshwaip bluffin' with my muffin Icrontian
    edited August 2004
    They are not necessarily doing the same kind of math. What can you do faster:

    3x4
    or
    3.1124 x 4.214

    Also, they may even be doing more math in a given frame.
  • edited August 2004
    It all depends on the individual work unit you are working on too. There are some 21 and 24 point Gromacs out there and there also are some 230 pointers too. Before Stanford reworked the points on Tinker work, the Gromacs work was the best units to get for points production but when Stanford switched benchmark machines to the P4 2.8 from the Celeron 500, the Tinker points worth went up a whole bunch. That is due to the P4's weak native FPU unit and Tinker not using any SSE or SSE2 optimizations, which is weaker than even the old Celeron's or P3's FPU unit. Since the Athlon and Duron processors have a much superior native FPU unit they are much more efficient at non-optimized Tinker wu's, which makes them produce a higher points/hour on Tinker wus.
  • botheredbothered Manchester UK
    edited August 2004
    It's like 230+ tinker, 24 hours. 230+ gromac, 2 1\2 days.
    Oh well, it's the work that's important. Is there any way to make sure eyes only gets all gromacs?
  • edited August 2004
    There's no real way to specify only Gromacs or only Tinker work, unfortunately, and I wish there was so I could keep the Tinkers off my P4 rigs.

    One thing that comes to mind on your time disparity bothered, is that it almost looks like the folding client isn't using SSE on your machine with Gromacs work. What client are you running (graphical or console) and are you using the -forceasm or -forcesse switch on the shortcut you are starting the client with?
  • botheredbothered Manchester UK
    edited August 2004
    graphical v4. I think I set -forceasm when I set it up but it was ages ago and I'm not 100% sure. I don't 'start' the client, it runs whenever the PC's on.

    edit,
    just checked, -advmethods -forceasm are what's set.
  • EyesOnlyEyesOnly Sweden New
    edited August 2004
    bothered wrote:
    Is there any way to make sure eyes only gets all gromacs?

    Hey i heard that. :mean::rant: :bawling:

    muddocktor wrote:
    There's no real way to specify only Gromacs or only Tinker work, unfortunately, and I wish there was so I could keep the Tinkers off my P4 rigs.

    One thing that comes to mind on your time disparity bothered, is that it almost looks like the folding client isn't using SSE on your machine with Gromacs work. What client are you running (graphical or console) and are you using the -forceasm or -forcesse switch on the shortcut you are starting the client with?

    I too wish there was some way so i could get only tinkers. :rarr: I'm running the console installed by firedaemon with no switches. Someone said a while back that they weren't needed so i didn't add any. Considering that this is the second post i read today where bothered makes up for his name :) and wish that i slow down indicates that i'm doing something right.

    Hopfully i wont see any more of those 20 minutes frame times. :eek: I hate them. :rant: :banghead: :mean: :bawling:
  • mmonninmmonnin Centreville, VA
    edited August 2004
    Gromacs do like 3x more work, as in data calculations, than tinkers do. Thus the time difference. Also one thing, it depends on the CPU. If its a P4 then it should be about the same times. An athlon, how ever will kick ass with tinkers.
  • witenoizwitenoiz 19,356 miles East of Kansas City, MO Member
    edited August 2004
    See this link from Missileman http://www.short-media.com/forum/showthread.php?t=17891 If I have a unit that shuts down improperly, I alway restart it and then shut it down with Ctrl + C and then start it again. I keep this very simple - always on the desktop - never use flags - always move low pointers to the slowest box. I have had some success with this arrangement. :fold: Jack
Sign In or Register to comment.