Previous Thread
Next Thread
Print Thread
c34b1 ifort problem
#17207 02/26/08 05:14 AM
Joined: Nov 2007
Posts: 7
M
maluyao Offline OP
Forum Member
OP Offline
Forum Member
M
Joined: Nov 2007
Posts: 7
I'm using Charmm c34b1 on CentOS Linux 5.1 X86_64. My intel ifort version is 10.1 20080112.

Charmm was compiled by ifort with mpich-127/mpich2/openmpi/lammpi

I also compile a charmm with ifort and without any mpi.

Modify c34b1/test/cbenchtest/mbcodyn.inp

from

dyna cpt start nstep 1000 timestep 0.001 nbscale @s imscale @t -

to

dyna cpt start nstep 10000000 timestep 0.001 nbscale @s imscale @t -

With mpirun, all Charmmm program parallel on 8core within a node.

All of these program stop on step 2425

charmm output
Code:
 
SELECTED IMAGES ATOMS BEING CENTERED ABOUT 0.000000 0.000000 0.000000
UPDECI: Nonbond update at step 2400
UPDECI: Image update at step 2425

SELECTED IMAGES ATOMS BEING CENTERED ABOUT 0.000000 0.000000 0.000000
UPDECI: Nonbond update at step 2425
TOTAL ENERGY CHANGE EXCEEDED
20. KCAL AND 10% OF THE TOTAL KINETIC ENERGY IN THE LAST STEP
PREVIOUS E = -0.3384E+05 CURRENT E = 0.2290E+07 KINETIC = 0.3909E+07


Writing RESTART FILE with previous and current coordinates,
which may be read by: READ COOR DYNR .... (see io.doc).
NOTE!! THIS FILE C A N N O T BE USED TO RESTART A RUN!!!



***** LEVEL -2 WARNING FROM <DYNAMC> *****
***** ENERGY CHANGE TOLERANCE EXCEEDED
******************************************
BOMLEV ( -3) IS NOT REACHED. WRNLEV IS 5


Parallel load balance (sec.):
Node Eext Eint Wait Comm List Integ Total
0 149.9 30.3 1.1 13.3 37.4 4.5 236.5
1 148.9 32.2 0.2 14.1 36.9 4.4 236.7
2 147.6 33.5 0.2 14.1 36.9 4.4 236.6
3 147.9 33.0 0.3 13.5 37.1 4.4 236.2
4 148.5 31.6 1.2 13.3 37.6 4.4 236.6
5 149.6 30.9 0.8 14.1 36.8 4.4 236.6
6 150.4 30.1 0.8 14.1 36.9 4.4 236.7
7 150.8 30.0 0.5 13.5 37.1 4.4 236.2
PARALLEL> Average timing for all nodes:
8 149.2 31.5 0.6 13.7 37.1 4.4 236.5
VCLOSE: Closing unit 1 with status "KEEP"

$$$$$$ New timer profile Local node$$$$$

First List 0.45 Other: 0.00
Comm coords 5.31 Other: 0.00
dynamc 0.15 Other: 0.00
Dynamics total 10.08 Other: 4.62
Image Update 3.86 Other: 0.00
List time 37.79 Other: 33.94
Direct Ewald time 100.21 Other: 0.00
Fill charge grid 14.66 Other: 0.00
Scalar sum 5.40 Other: 0.00
Grad sum 10.47 Other: 0.00
FFTcomm 10.48 Other: 0.00
FFT 16.59 Other: 6.11
Recip Ewald time 47.85 Other: 0.72
Ewald time 148.08 Other: 0.02
Nonbond force 149.87 Other: 1.79
Bond energy 1.98 Other: 0.00
Angle energy 1.38 Other: 0.00
Dihedral energy 1.30 Other: 0.00
Restraints energy 0.01 Other: 0.00
INTRNL energy 24.52 Other: 19.85
Comm force 8.37 Other: 0.00
Energy time 190.01 Other: 7.25
Total time 238.89 Other: 0.56

$$$$$$ Average profile $$$$$

First List 0.45 Other: 0.00
Comm coords 5.51 Other: 0.00
dynamc 0.14 Other: 0.00
Dynamics total 10.26 Other: 4.81
Image Update 3.86 Other: 0.00
List time 37.31 Other: 33.46
Direct Ewald time 99.09 Other: 0.00
Fill charge grid 14.02 Other: 0.00
Scalar sum 5.48 Other: 0.00
Grad sum 9.50 Other: 0.00
FFTcomm 11.13 Other: 0.00
FFT 17.15 Other: 6.68
Recip Ewald time 48.36 Other: 1.23
Ewald time 148.06 Other: 0.00
Nonbond force 149.19 Other: 1.11
Bond energy 1.86 Other: 0.00
Angle energy 1.19 Other: 0.00
Dihedral energy 1.26 Other: 0.00
Restraints energy 0.01 Other: 0.00
INTRNL energy 24.46 Other: 19.79
Comm force 8.40 Other: 0.00
Energy time 190.30 Other: 7.54
Total time 238.89 Other: 0.56

NORMAL TERMINATION BY END OF FILE
MAXIMUM STACK SPACE USED IS 772278
STACK CURRENTLY IN USE IS 0
MOST SEVERE WARNING WAS AT LEVEL -2
HEAP PRINTOUT- HEAP SIZE 10240000
SPACE CURRENTLY IN USE IS 0
MAXIMUM SPACE USED IS 3483268
FREE LIST
PRINHP> ADDRESS: 1 LENGTH: 10240000 NEXT: 0

$$$$$ JOB ACCOUNTING INFORMATION $$$$$
ELAPSED TIME: 3.98 MINUTES
CPU TIME: 0.02 SECONDS


Re: c34b1 ifort problem
maluyao #17208 02/26/08 11:10 AM
Joined: Feb 2004
Posts: 147
Forum Member
Offline
Forum Member
Joined: Feb 2004
Posts: 147
I have also observed differences in behaviour when using CHARMM executables generated using the Intel compilers for x86_64 for several recent stable CHARMM versions, when compared with executables produced with gfortran and the PathScale compilers for x86_64. The Intel ones either produce significantly different numerical results or throw errors which seem to me related to numerical differences (like energy tolerance exceeded) for some test inputs like those provided with the CHARMM distribution (f.e. both the older benchmark with classical dynamics and the PME one); in the same conditions the executables produced with gfortran and PathScale compilers give very similar or identical numerical results.

I can't provide too much data at the moment, but I wanted to say that the original poster is not alone in getting strange results when the Intel compilers are used.

Re: c34b1 ifort problem
bogdan #17209 02/26/08 04:49 PM
Joined: Jun 2007
Posts: 32
M
mks Offline
Forum Member
Offline
Forum Member
M
Joined: Jun 2007
Posts: 32
Which version of ifort compiler you are using?

Re: c34b1 ifort problem
mks #17210 02/27/08 07:04 AM
Joined: Nov 2007
Posts: 7
M
maluyao Offline OP
Forum Member
OP Offline
Forum Member
M
Joined: Nov 2007
Posts: 7
My intel ifort version is 10.1 20080112.

Re: c34b1 ifort problem
maluyao #17211 02/28/08 07:41 PM
Joined: Jun 2007
Posts: 32
M
mks Offline
Forum Member
Offline
Forum Member
M
Joined: Jun 2007
Posts: 32
Thanks Maluyao,

I can see in your first post that you have used version 10.1. I have also observed some strange behavior of charmm after compiling with ifort version 10.1. Since Bogdan in his post has not mentioned any thing thing about version of compiler, I wanted to know his version of compiler, and did he observed odd behavior with only version 10.1 or with older version also(9.1 etc).

Last edited by mks; 02/28/08 10:31 PM.
Re: c34b1 ifort problem
mks #17212 02/29/08 04:11 PM
Joined: Feb 2004
Posts: 147
Forum Member
Offline
Forum Member
Joined: Feb 2004
Posts: 147
Sorry for the incomplete information...

I have certainly seen this behaviour with version 10.1.008 on x86_64.

However, I have seen a similar behaviour (it was ~2 years ago and I don't remember all the details to say whether it's the same or only similar) with an older release, 8.something also on x86_64; the same version on i686 worked fine.

Re: c34b1 ifort problem
bogdan #17213 06/26/08 07:54 PM
Joined: Jun 2008
Posts: 4
D
Forum Member
Offline
Forum Member
D
Joined: Jun 2008
Posts: 4
Just for some more information on this problem, I see this behavior when compiling with ifort 10.1.015 on X86_64.
I have tried compiling with different flags. Of the following 5 versions of charmm run on the same script, the first 3 run to completion but the last 2 seem to die:

TOTAL ENERGY CHANGE EXCEEDED
20. KCAL AND 10% OF THE TOTAL KINETIC ENERGY IN THE LAST STEP
PREVIOUS E = -0.7934E+05 CURRENT E = -0.7583E+05 KINETIC = 0.3467E+05

***** LEVEL -2 WARNING FROM <DYNAMC> *****
***** ENERGY CHANGE TOLERANCE EXCEEDED
******************************************
BOMLEV ( 0) IS REACHED - TERMINATING. WRNLEV IS -5


charmm_1: -132 -w95 -cm -align all -i8 -mcmodel=medium -shared-intel -auto-scalar -xT -axT -ip -O3

charmm_2: -132 -w95 -cm -align all -i8 -mcmodel=medium -shared-intel -auto-scalar -xT -axT -ip -mp1 -O3

charmm_3: -132 -w95 -cm -align all -i8 -mcmodel=medium -shared-intel -auto-scalar -c -O2

charmm_4: mpif90 -132 -w95 -cm -align all -i8 -mcmodel=medium -shared-intel -auto-scalar -xT -axT -mp1 -O3 *CRASH*

charmm_5: mpif90 -132 -w95 -cm -align all -i8 -mcmodel=medium -shared-intel -auto-scalar -fp-model strict -O3 *CRASH*


The main difference between the sets is the -ip flag which "enables additional interprocedural optimizations for single file compilation."

Not sure exactly what is happening...

Re: c34b1 ifort problem
davecap #17214 09/10/08 03:00 PM
Joined: Oct 2007
Posts: 9
N
Forum Member
Offline
Forum Member
N
Joined: Oct 2007
Posts: 9
Quote:



charmm_1: -132 -w95 -cm -align all -i8 -mcmodel=medium -shared-intel -auto-scalar -xT -axT -ip -O3

charmm_2: -132 -w95 -cm -align all -i8 -mcmodel=medium -shared-intel -auto-scalar -xT -axT -ip -mp1 -O3

charmm_3: -132 -w95 -cm -align all -i8 -mcmodel=medium -shared-intel -auto-scalar -c -O2

charmm_4: mpif90 -132 -w95 -cm -align all -i8 -mcmodel=medium -shared-intel -auto-scalar -xT -axT -mp1 -O3 *CRASH*

charmm_5: mpif90 -132 -w95 -cm -align all -i8 -mcmodel=medium -shared-intel -auto-scalar -fp-model strict -O3 *CRASH*


The main difference between the sets is the -ip flag which "enables additional interprocedural optimizations for single file compilation."

Not sure exactly what is happening...




Actually are the last two mpi and the first 3 serial?
Could it be mpi introducing problems or am I misreading something?


Moderated by  BRBrooks, bucknerj, lennart, rmv 

Link Copied to Clipboard
Powered by UBB.threads™ PHP Forum Software 7.7.4
(Release build 20200307)
Responsive Width:

PHP: 5.6.33-0+deb8u1 Page Time: 0.009s Queries: 30 (0.004s) Memory: 0.9594 MB (Peak: 1.0744 MB) Data Comp: Off Server Time: 2020-10-01 17:31:36 UTC
Valid HTML 5 and Valid CSS