Re: Wrong compilation artifacts on ES40s?

From: Sanjeev B.S. <sanjeev_at_mbu.iisc.ernet.in>
Date: Mon 25 Jun 2001 00:17:48 +0400 (RET)

        I am trying to make ES40s work for amber6. I did benchmarking
        tests aganist a variety of platforms (linux, linux cluster, IBM
        SP3, etc.) and found that after running the benchmark test, the
        energy values were identical in all. Except on ES40s! I am
        attaching the MACHINE file and the mdout difference files. The
        first one in difference is the correct(?) one and the second
        one is from ES40s. (Incidentally even the number of
        processors change the output!)

There is nothing wrong with the results. At the end of the run:

< NSTEP = 100 TIME(PS) = 510.150 TEMP(K) = 296.88 PRESS = 0.00
< Etot = -57714.6839 EKtot = 13904.7963 EPtot = -71619.4802
< BOND = 513.3279 ANGLE = 1239.4034 DIHED = 975.5933
< 1-4 NB = 547.9299 1-4 EEL = 6666.7751 VDWAALS = 8145.3633
< EELEC = -89707.8731 EHBOND = 0.0000 CONSTRAINT = 0.0000
< Ewald error estimate: 0.8821E-04

---
>  NSTEP =   100  TIME(PS) =  510.150  TEMP(K) =   296.87  PRESS =      0.00
>  Etot   =  -57714.6108  EKtot   =   13904.7531  EPtot      =  -71619.3640
>  BOND   =     513.3283  ANGLE   =    1239.3997  DIHED      =     975.5875
>  1-4 NB =     547.9295  1-4 EEL =    6666.7675  VDWAALS    =    8145.4824
>  EELEC  =  -89707.8589  EHBOND  =       0.0000  CONSTRAINT =       0.0000
>  Ewald error estimate:   0.8949E-04
Trajectories diverge due to different floating point implementations,
and when you vary the number of processors, the order of evaluation
of global operations is indeterminate and this also has an effect. To 
really see if there is a problem, you would need to compare the
overall statistics for very long runs.
Bill Ross
------------------------------------------------------------------------------
Hello,
	I would only be glad if there is nothing worng with the results.
But somehow I am still suspicious because by default I guess Compaq gives
64bit architecture and it should be more accurate. Here it was only 100
steps. And for these 100 steps, IBM SP3 (2-16 processors), Octanes (SGI),
Alpha-Compaq work stations (stand alone), and Linux (1-12 processors) have
given same numbers upto the last digit. I agree that I may be wrong. I
shall do long time scale simulation soon and shall verify.
Thanking you very much,
-Sanjeev

Received on Sun Jun 24 2001 - 13:17:48 PDT