[Pw_forum] I/O performance on BG/P systems

[Gabriele Sclauzero]

Dear all,

     I've been running some large systems on BG/P and I've met 
surprisingly large timings for disk writing operations (wavefunctions).
As an example, this is part of the timing for a system with 408 atoms, 
904 KS states, and FFT grid: (220,220,512)

      General routines
      calbec       :   1165.61s CPU   1165.61s WALL (     337 calls)
      fft          :     74.20s CPU     74.20s WALL (     636 calls)
      ffts         :      1.39s CPU      1.39s WALL (      78 calls)
      fftw         :   2042.87s CPU   2042.87s WALL (  206112 calls)
      interpolate  :      5.75s CPU      5.75s WALL (      78 calls)
      davcio       :   1083.30s CPU   1083.30s WALL (      38 calls)

The davcio cputime looks pretty high to me, compared to what I've seen 
in other computing clusters. I know that the BGP architecture is pretty 
different than what I've been using so far, but the system 
administrators claim that the IBM GPFS file system should be very 
efficient and fast.
Before moving some complaints to the sysadmins, I would like to have 
some feedback from those of you who have used this kind of architecture.
Are you aware of similar kind of I/O issues? Is there any performance 
tuning that can be done on the software/compiler side?



Regards,

Gabriele

-- 

Gabriele Sclauzero, EPFL SB ITP CSEA
PH H2 462, Station 3, CH-1015 Lausanne