AMBER Archive (1999)

Subject: Problem with bench test on PIII450

From: Claudio Chuaqui (chuaqui_at_rowland.org)
Date: Fri Aug 13 1999 - 12:33:54 CDT

Hello,

I have been porting AMBER4.1 to a dual PentiumIII 450. With some effort
I have manged to get things working well in serial and parallel (by
replacing SGI_MP directives with OpenMP directives supported by
the pgf77/pgcc compiler), except for one glitch which I can't seem to
overcome. I have been using the "bench" test suite provided with
the distribution to test things out. Perhaps someone can point
me in the right direction...

System: dual PIII450 running RedHat 6.0 (2.2.5-15smp kernel)

Both in serial and parallel, under every possible compiler/optimization
directive combination, using pgf77, g77 and even fort77,
I obtain the following (identical) differences at the first time step
for the dnawat/Ewald test case.:

b1_md4.out.dif:

188,189c188,189
< NSTEP = 1 TIME(PS) = 0.101 TEMP(K) = 249.75 PRESS =
-2047.64
< Etot = -24172.9092 EKtot = 3874.6436 EPtot =
-28047.5528 

---
>  NSTEP =     1  TIME(PS) =    0.101  TEMP(K) =   249.75  PRESS =
-2047.40
>  Etot   =  -24172.8891  EKtot   =    3874.6428  EPtot      =
-28047.5319

329,334c329,334
<  NSTEP =   100  TIME(PS) =    0.200  TEMP(K) =     5.65  PRESS =
630.69
<  Etot   =      22.8409  EKtot   =      87.5834  EPtot      =
108.6791
<  BOND   =       7.5226  ANGLE   =      18.9307  DIHED      =
7.8277
<  1-4 NB =       9.3561  1-4 EEL =      29.6332  VDWAALS    =
209.7362
<  EELEC  =     320.7651  EHBOND  =       2.0258  CONSTRAINT =
0.0000
<  EKCMT  =      54.8849  VIRIAL  =    1149.1345  VOLUME     =
381.2115
---
>  NSTEP =   100  TIME(PS) =    0.200  TEMP(K) =     5.65  PRESS =
630.66
>  Etot   =      22.8213  EKtot   =      87.5914  EPtot      =
108.6732
>  BOND   =       7.5223  ANGLE   =      18.9312  DIHED      =
7.8272
>  1-4 NB =       9.3558  1-4 EEL =      29.6337  VDWAALS    =
209.7410
>  EELEC  =     320.7652  EHBOND  =       2.0259  CONSTRAINT =
0.0000
>  EKCMT  =      54.8862  VIRIAL  =    1149.0809  VOLUME     =
381.2251
336c336
<  Ewald error estimate:   0.2978E-04
---
>  Ewald error estimate:   0.7049E-04

Given that there are differences at the first time step, I am assuming
they
are not attributable to machine precision, but rather, that
there exists a more serious underlying problem. I assume that the
problem lies with the use of the Ewald  routine for this test. I have
made bugfixes 32, 78, and 88, to no avail. The same code/test case
compiled and run our SGIs shows no such differences.

Has anyone observed similar discrepancies for Pentium based machines?

Thanks very much for your help,

Claudio Chuaqui
Rowland Institute for Science
Cambridge, MA
chuaqui_at_rowland.org