[MITgcm-support] Something very slow has happened to the pressuresolver

[Dimitris Menemenlis]

Christopher, I am no expert on solve_for_pressure but first thing I would look at is the number of cg2d_iters that it does.  Set ALLOW_DEBUG and debuglevel=1

-----Original Message-----

From:  "Christopher L. Wolfe" <clwolfe at ucsd.edu>
Subj:  [MITgcm-support] Something very slow has happened to the pressuresolver
Date:  Wed Feb 27, 2008 2:04 am
Size:  2K
To:  mitgcm-support at mitgcm.org


Hello modelers,

I ran MITgcm on a Cray X1E (Phoenix at ORNL) about a year ago with good results. I recently got more time on this machine and downloaded the latest MITgcm checkpoint (59n), but was dismayed to find that this checkpoint is A LOT (like 5x) slower than the one I was originally using (58q_post). I ran freshly compiled versions of checkpoint_58q_post and checkpoint_59n with identical model configurations, compiler options, and CPP options (to the extent possible) and discovered that the pressure solver in 59n takes ten times (!) more CPU time on the X1E than the pressure solver from 58q_post. No other routine seems to haveacquiredmore than about 10% greater overhead. Interestingly, I have not noticed significant increases in the pressure solver overhead on IBM sp4 machines.


Before I start a flurry of diffs, does anyone have any idea what has has been done to the pressure solver over the past year that would have had such a significant impact on its performance on the Cray?

Thanks in advance,
Christopher

-----------------------------------------------------------