Bon, j'ai fait quelques tests supplémentaires quand même. Effectivement g77 et gfortran sont par défaut sans optimisation alors que ifort a -O2 par défaut.
J'ai donc testé plusieurs options pour chaque. Voici les meilleurs résultats pour chaque. ( avec -O3 c'est le mieux en fait).
measurement date = 08/12/11
log file = bench_g77.log
machine = 2* (Xeon5430 4*2.66 GHz), Linux 2.6.18-8.el5 x86_64 (g77 3.4.6)
compile with [ g77 -O3 ]
program weight ref.time time[sec] time/ref ref/time
( 1) matvec 8 296.00 21.44 0.0724 13.8040
( 2) matvecz 4 279.00 10.66 0.0382 26.1677
( 3) leqs4h 3 256.00 7.65 0.0299 33.4772
( 4) leqs1k 3 279.00 17.76 0.0636 15.7112
( 5) jacobi11 3 261.00 16.45 0.0630 15.8692
( 6) jacobi1h 3 124.00 10.32 0.0832 12.0132
( 7) runge 8 204.00 12.03 0.0590 16.9618
( 8) intgl4 4 300.00 19.43 0.0648 15.4377
( 9) mathfnc1 2 123.00 5.38 0.0438 22.8455
(10) mathfnc2 2 226.00 16.49 0.0730 13.7053
(11) intosc 2 236.00 5.76 0.0244 41.0007
(12) intpi3 2 122.00 5.32 0.0436 22.9410
(13) permute1 2 23.00 1.27 0.0550 18.1818
(14) permute2 2 54.00 1.28 0.0237 42.2535
(15) permute3 2 228.00 2.36 0.0104 96.4875
weighted arithmetic average of (ref/time) = 23.17
weighted geometrical average of (ref/time) = 20.01
1/weighted arithmetic average of (time/ref) = 18.32
fMark = 20.01
measurement date = 08/12/11
log file = bench_gfortran.log
machine = 2* (Xeon5430 4*2.66 GHz), Linux 2.6.18-8.el5 x86_64 (GNU Fortran 95 (GCC) 4.1.1 20070105 (Red Hat 4.1.1-52))
compile with [ gfortran -O3 ]
program weight ref.time time[sec] time/ref ref/time
( 1) matvec 8 296.00 16.81 0.0568 17.6117
( 2) matvecz 4 279.00 8.95 0.0321 31.1906
( 3) leqs4h 3 256.00 6.02 0.0235 42.5461
( 4) leqs1k 3 279.00 15.41 0.0552 18.1098
( 5) jacobi11 3 261.00 9.82 0.0376 26.5703
( 6) jacobi1h 3 124.00 4.06 0.0327 30.5569
( 7) runge 8 204.00 9.67 0.0474 21.1027
( 8) intgl4 4 300.00 29.93 0.0998 10.0227
( 9) mathfnc1 2 123.00 5.33 0.0433 23.0726
(10) mathfnc2 2 226.00 15.57 0.0689 14.5151
(11) intosc 2 236.00 4.86 0.0206 48.5797
(12) intpi3 2 122.00 4.87 0.0399 25.0565
(13) permute1 2 23.00 1.36 0.0592 16.8869
(14) permute2 2 54.00 1.37 0.0254 39.3873
(15) permute3 2 228.00 2.32 0.0102 98.1912
weighted arithmetic average of (ref/time) = 27.19
weighted geometrical average of (ref/time) = 23.75
1/weighted arithmetic average of (time/ref) = 21.34
fMark = 23.75
measurement date = 08/12/11
log file = bench_ifort.log
machine = 2* (Xeon5430 4*2.66 GHz), Linux 2.6.18-8.el5 x86_64 (ifort 10.1.008)
compile with [ ifort -O3 ]
program weight ref.time time[sec] time/ref ref/time
( 1) matvec 8 296.00 15.64 0.0528 18.9258
( 2) matvecz 4 279.00 8.75 0.0314 31.8894
( 3) leqs4h 3 256.00 5.71 0.0223 44.7944
( 4) leqs1k 3 279.00 14.63 0.0524 19.0743
( 5) jacobi11 3 261.00 8.89 0.0340 29.3720
( 6) jacobi1h 3 124.00 3.03 0.0245 40.8701
( 7) runge 8 204.00 8.53 0.0418 23.9128
( 8) intgl4 4 300.00 6.21 0.0207 48.2781
( 9) mathfnc1 2 123.00 1.53 0.0125 80.1825
(10) mathfnc2 2 226.00 3.77 0.0167 59.8834
(11) intosc 2 236.00 4.40 0.0187 53.5876
(12) intpi3 2 122.00 4.63 0.0379 26.3727
(13) permute1 2 23.00 0.92 0.0398 25.1092
(14) permute2 2 54.00 1.03 0.0191 52.3763
(15) permute3 2 228.00 2.36 0.0104 96.4467
weighted arithmetic average of (ref/time) = 37.07
weighted geometrical average of (ref/time) = 32.96
1/weighted arithmetic average of (time/ref) = 29.84
fMark = 32.96
On ne gagne pas grand chose entre -O2 et -O3 pour gfortran, g77 et ifort (0.2 sur le fMark).
On voit encore que ifort est plus rapide.
Kevin Kofler (./44) :
Quant aux paquetages non disponibles pour RHEL, va voir https://fedoraproject.org/wiki/EPEL et http://rpmfusion.org/ .
Merci! J'avais juste rpmforge jusqu'à présent, mais il y a peu de packages pour x86_64.el5
