Linux 64bits vs 32bits?
- Torben
- Topic Author
- Offline
- Premium Member
-
Less
More
- Posts: 91
- Thank you received: 1
12 years 11 months ago - 12 years 11 months ago #6318
by Torben
Linux 64bits vs 32bits? was created by Torben
Hello,
I am considering upgrading to a x86_64 system in order to run a contemporary Salome-MECA. Is there any performance related change to expect due to the 64bits system compared to 32bits?
Thanks for your input.
Edit:
After googling a bit more I found this benchmark for floating points operations:
illustration taken from here: www.phoronix.com
So it seems to work better with _64bits all other equal. However If You have experience to share, please do.
I am considering upgrading to a x86_64 system in order to run a contemporary Salome-MECA. Is there any performance related change to expect due to the 64bits system compared to 32bits?
Thanks for your input.
Edit:
After googling a bit more I found this benchmark for floating points operations:
illustration taken from here: www.phoronix.com
So it seems to work better with _64bits all other equal. However If You have experience to share, please do.
Last edit: 12 years 11 months ago by Torben.
- Torben
- Topic Author
- Offline
- Premium Member
-
Less
More
- Posts: 91
- Thank you received: 1
12 years 10 months ago - 12 years 10 months ago #6337
by Torben
Replied by Torben on topic Re: Linux 64bits vs 32bits?
For your information I just did a testcase. Exact same hardware (dual boot, actually) and 3GB RAM. Clearly this is not enough to include the benefits of the higher address space. My x64 test system is booted on a 16GB USB stick so to filter out the inferior disc performance the job was executed on the hard disk in both cases.
Size of the job can be described by:
Run 1: Linux Mint 9, x86_32, SALOME-MECA-2010.2-LGPL-i386, Timings:
Run 2: Linux Mint 9, x86_64, SALOME-MECA-2010.2-LGPL-x86_64, Timings:
So clearly there is a performance improvement by using 64bits systems and versions of Salome-MECA. Even for a "small" job not even close to the 4GB address space limitation a factor of improvement of 1,28 was found.
There are oher challenges of x86_64 systems, but if Aster performance matters it seems worth considering. What puzzles me a bit though is that if I test the total execution tlme from hitting the "RUN" button in Salome-MECA until the run status changes away from RUNNING in the object browser, the performance difference is much less obvious as the total elapsed time measured with a stop watch is around 85 seconds in either case.
Size of the job can be described by:
--- NOMBRE TOTAL DE NOEUDS : 60095 DONT : 35768
NOEUDS "LAGRANGE"
--- NOMBRE TOTAL D'EQUATIONS : 157676
--- TAILLE DU PROFIL MORSE DE LA TRIANGULAIRE SUPERIEURE (FORMAT SCR): 4570293
--- DONC LA TAILLE DE LA MATRICE EST:
--- EN SYMETRIQUE NNZ= 4570293
--- EN NON SYMETRIQUE NNZ= 8982910
Run 1: Linux Mint 9, x86_32, SALOME-MECA-2010.2-LGPL-i386, Timings:
********************************************************************************
* COMMAND : USER : SYSTEM : USER+SYS : ELAPSED *
********************************************************************************
* init (jdc) : 2.40 : 0.23 : 2.63 : 3.85 *
* . compile : 0.00 : 0.00 : 0.00 : 0.00 *
* . exec_compile : 0.22 : 0.01 : 0.23 : 0.23 *
* . report : 0.02 : 0.00 : 0.02 : 0.01 *
* . build : 0.00 : 0.00 : 0.00 : 0.00 *
* DEBUT : 0.03 : 0.05 : 0.08 : 0.09 *
* LIRE_MAILLAGE : 0.15 : 0.00 : 0.15 : 0.35 *
* CREA_MAILLAGE : 0.48 : 0.00 : 0.48 : 0.52 *
* AFFE_MODELE : 0.12 : 0.00 : 0.12 : 0.36 *
* DEFI_MATERIAU : 0.01 : 0.00 : 0.01 : 0.00 *
* AFFE_MATERIAU : 0.01 : 0.00 : 0.01 : 0.01 *
* AFFE_CARA_ELEM : 0.14 : 0.00 : 0.14 : 0.17 *
* AFFE_CHAR_MECA : 2.57 : 0.01 : 2.58 : 2.70 *
* MECA_STATIQUE : 29.38 : 3.21 : 32.59 : 40.25 *
* CALC_ELEM : 5.33 : 0.02 : 5.35 : 5.43 *
* CALC_NO : 2.91 : 0.06 : 2.97 : 2.94 *
* IMPR_RESU : 1.29 : 0.09 : 1.38 : 1.45 *
* FIN : 0.07 : 0.18 : 0.25 : 1.41 *
* . part Superviseur : 2.44 : 0.28 : 2.72 : 4.12 *
* . part Fortran : 42.45 : 3.57 : 46.02 : 55.57 *
********************************************************************************
* TOTAL_JOB : 44.89 : 3.85 : 48.74 : 59.69 *
********************************************************************************
Run 2: Linux Mint 9, x86_64, SALOME-MECA-2010.2-LGPL-x86_64, Timings:
********************************************************************************
* COMMAND : USER : SYSTEM : USER+SYS : ELAPSED *
********************************************************************************
* init (jdc) : 2.26 : 0.32 : 2.58 : 2.59 *
* . compile : 0.00 : 0.00 : 0.00 : 0.00 *
* . exec_compile : 0.19 : 0.04 : 0.23 : 0.23 *
* . report : 0.02 : 0.00 : 0.02 : 0.02 *
* . build : 0.00 : 0.00 : 0.00 : 0.00 *
* DEBUT : 0.03 : 0.07 : 0.10 : 0.10 *
* LIRE_MAILLAGE : 0.10 : 0.00 : 0.10 : 0.11 *
* CREA_MAILLAGE : 0.42 : 0.00 : 0.42 : 0.42 *
* AFFE_MODELE : 0.05 : 0.00 : 0.05 : 0.03 *
* DEFI_MATERIAU : 0.00 : 0.00 : 0.00 : 0.00 *
* AFFE_MATERIAU : 0.03 : 0.00 : 0.03 : 0.02 *
* AFFE_CARA_ELEM : 0.25 : 0.00 : 0.25 : 0.16 *
* AFFE_CHAR_MECA : 2.12 : 0.01 : 2.13 : 2.14 *
* MECA_STATIQUE : 22.54 : 1.70 : 24.24 : 23.24 *
* CALC_ELEM : 3.59 : 0.00 : 3.59 : 3.59 *
* CALC_NO : 2.20 : 0.01 : 2.21 : 2.21 *
* IMPR_RESU : 1.90 : 0.03 : 1.93 : 1.93 *
* FIN : 0.12 : 0.34 : 0.46 : 0.47 *
* . part Superviseur : 2.31 : 0.39 : 2.70 : 2.71 *
* . part Fortran : 33.31 : 2.09 : 35.40 : 34.31 *
********************************************************************************
* TOTAL_JOB : 35.62 : 2.48 : 38.10 : 37.02 *
********************************************************************************
So clearly there is a performance improvement by using 64bits systems and versions of Salome-MECA. Even for a "small" job not even close to the 4GB address space limitation a factor of improvement of 1,28 was found.
There are oher challenges of x86_64 systems, but if Aster performance matters it seems worth considering. What puzzles me a bit though is that if I test the total execution tlme from hitting the "RUN" button in Salome-MECA until the run status changes away from RUNNING in the object browser, the performance difference is much less obvious as the total elapsed time measured with a stop watch is around 85 seconds in either case.
Last edit: 12 years 10 months ago by Torben.
Moderators: catux
Time to create page: 0.123 seconds