5:04 p.m.
Thanks to all of you who have run the matrix_perf tests. This will
be helpful in setting gretl's (internal, default) parameters for
using the system BLAS versus OpenMP (where available), versus our
own single-threaded matrix multiplication code.
Let me address some of the little "issues" that have come up, then
I'll offer a brief discussion of the results people have posted.
Issues
* Artur getting "error in function printvec, line 1": I think your
gretl sources must be not quite up to date. I'm not seeing this
problem.
* "nan" getting printed in the Gflops column in some cases. This is
due to division by zero, really meaning that the timer offered by
the OS is not very good. In the latest matrix_perf update (0.4) I've
converted such results to "inf", which is not true (!) but at least
should get the rank-ordering right.
* Some uncertainty over gretl's identification of the BLAS variant
against which gretl is linked. As I mentioned, gretl is relying on
the "ldd" program for this info.
I thought that in Artur's case gretl was misindentifying the BLAS as
the Netlib "reference" version when it must be some optimized
variant; that was because it was showing as 5-7 times as fast as
"vanilla" on some of the experiments. But Artur tells me he's using
Debian/Ubuntu libblas version 1.2.20110419-5, and that's Netlib OK.
I now think that the Debian guys must be building libblas with very
aggressive optimization at the compiler level, probably inducing
vectorization for amd64. So it seems our detection was right, but my
expectation that Netlib BLAS couldn't be so fast relative to vanilla
was wrong.
Then there was Hélio seeing on stderr, "detect blas: confused, found
too many libs!". This was because we have, in the ldd output:
libblas.so.3 => /lib64/libblas.so.3
libf77blas.so.3 => /usr/lib64/atlas/libf77blas.so.3
libcblas.so.3 => /usr/lib64/atlas/libcblas.so.3
The first of these lines suggested plain Netlib to gretl, while the
others suggested the BLAS was ATLAS. I'm not sure about this, but I
suspect it really is ATLAS, with ATLAS BLAS having been placed at
/lib64/libblas.so.3 via the "alternatives" mechanism that Sven
mentioned. It would be good if that could be confirmed; then we ccan
adjust our detection code accordingly.
Discussion
I'll just mention some points that seem of interest to me.
* If Hélio's BLAS is really ATLAS (on Fedora 19), it's surprisingly
unimpressive, coming a poor second to gretl's openmp code (or even
third). Maybe it's single-threaded? Or maybe it's not really ATLAS
after all?
* Henrique's Mac Mini (plus follow-up on the MacBook Pro): Apple's
VecLib seems to do quite nicely!
* Ignacio's Dell Xeon: Netlib does quite well in some cases but is
clearly single-threaded and is beaten out by OpenMP on larger
problems.
For comparison, I've posted results for my two systems (a desktop
named "myrtle" and laptop named "waverley") at
http://ricardo.ecn.wfu.edu/~cottrell/tmp/myrtle-mperf.txt
http://ricardo.ecn.wfu.edu/~cottrell/tmp/waverley-mperf.txt
These show openblas operating at up to 50 times the speed of
"vanilla" and up to 10 times the speed of gretl's openmp -- which
explains why I bothered building openblas for the 64-bit Windows
gretl packge. I would recommend installing openblas on Linux,
though you'd want to make sure that if you do so, it's a build that
uses OpenMP rather than "raw" pthreads, otherwise it will not play
nicely with gretl's internal uses of OpenMP.
Allin
7:54 p.m.
* If Hélio's BLAS is really ATLAS (on Fedora 19), it's
surprisingly
unimpressive, coming a poor second to gretl's openmp code (or even third).
Maybe it's single-threaded? Or maybe it's not really ATLAS after all?
No, it was not the ATLAS lib.
See if you can use the reports I send.
8:19 a.m.
On Wed, 18 Jun 2014, Hélio Guilherme wrote:
> * If Hélio's BLAS is really ATLAS (on Fedora 19), it's
surprisingly
> unimpressive, coming a poor second to gretl's openmp code (or even third).
> Maybe it's single-threaded? Or maybe it's not really ATLAS after all?
>
> No, it was not the ATLAS lib.
See if you can use the reports I send.
Thanks, Hélio. So it turns out gretl was right to find your original BLAS
set-up confusing!
Allin
8:39 a.m.
Hi,
these are my results. Machine: Intel(R) Core(TM) i7-4500U CPU @ 1.80GHz,
16GB RAM, Debian Testing/Unstable/Experimental (mixed).
Gretl configured: by "'--enable-build-doc' '--enable-openmp'
'--with-odbc''" and compiled by GCC 4.9.0.
All the Best,
Marcin
######################################################
dgemm experiment 1, variant 1, speed in Gflops
m n k vanilla openmp netlib
128 128 128 1,2844 8,4280 2,3635
128 128 256 2,3648 9,8147 8,7647
128 128 512 2,5742 10,958 8,6539
128 128 1024 2,7452 4,5530 9,1735
128 128 2048 2,7503 5,5682 9,2473
result: netlib dominates for mnk >= 16777216
openmp dominates for mnk < 16777216
dgemm experiment 1, variant 2, speed in Gflops
m n k vanilla openmp netlib
128 128 128 2,4507 3,6337 7,8852
256 256 128 2,7755 8,6371 8,1046
512 512 128 2,7741 9,9523 8,3603
1024 1024 128 2,8880 9,6166 8,5778
2048 2048 128 2,8775 8,2817 8,7284
result: netlib dominates for mnk >= 536870912
dgemm experiment 1, variant 3, speed in Gflops
m n k vanilla openmp netlib
128 128 128 2,5831 8,4664 7,7957
256 256 256 2,9340 11,584 8,6843
512 512 512 3,0497 11,710 9,7398
1024 1024 1024 2,3554 7,4442 10,085
2048 2048 2048 2,4262 5,1813 10,295
result: netlib dominates for mnk >= 1073741824
openmp dominates for mnk < 1073741824
dgemm experiment 2, variant 1, speed in Gflops
m n k vanilla openmp netlib
8 8 8 0,77332 0,21404 0,61904
16 8 8 0,86759 0,37646 0,98366
32 8 8 1,1219 0,30031 1,3463
64 8 8 1,3700 0,82423 1,6513
128 8 8 1,4883 1,2634 1,7845
256 8 8 1,5561 1,5149 1,9103
512 8 8 1,6070 1,6414 1,6082
1024 8 8 1,6016 1,8070 1,8167
2048 8 8 1,5968 1,7187 1,8652
4096 8 8 1,5807 1,4528 1,7123
result: netlib dominates for mnk >= 65536
dgemm experiment 2, variant 2, speed in Gflops
m n k vanilla openmp netlib
10 2 1000 2,0464 0,68509 2,6785
20 2 1000 2,3376 2,4716 3,0151
40 2 1000 2,0023 3,2774 3,0313
80 2 1000 2,4535 4,5667 2,8541
160 2 1000 2,7358 4,4775 2,8616
320 2 1000 2,8717 6,4803 2,8914
640 2 1000 2,4402 5,3052 1,8185
1280 2 1000 2,3685 4,9653 1,9498
2560 2 1000 2,3710 4,0467 1,6928
5120 2 1000 2,3778 4,3298 1,6417
result: openmp dominates for mnk >= 80000
netlib dominates for mnk < 80000
dgemm experiment 2, variant 3, speed in Gflops
m n k vanilla openmp netlib
10 10 1000 2,1420 2,7267 3,1789
20 10 1000 2,4351 5,0807 3,8299
40 10 1000 2,1367 7,2032 4,0107
80 10 1000 2,5392 9,3613 5,6899
160 10 1000 2,8350 10,593 5,6848
320 10 1000 2,8115 10,835 6,0015
result: openmp dominates for mnk >= 200000
netlib dominates for mnk < 200000
Operating system: Linux (64-bit)
BLAS library: Netlib
Number of processors: 4
OpenMP enabled: yes
Performance summary:
vanilla -
dominates outright in 0 out of 6 tests
openmp -
dominates outright in 0 out of 6 tests
dominates in 2 test(s) for mnk >= (80000, 200000)
dominates in 2 test(s) for mnk < (16777216, 1073741824)
netlib -
dominates outright in 0 out of 6 tests
dominates in 4 test(s) for mnk >= (16777216, 536870912, 1073741824,
65536)
dominates in 2 test(s) for mnk < (80000, 200000)
######################################################
--
Marcin Błażejowski
GG: 203127
9:11 a.m.
Am 17.06.2014 23:04, schrieb Allin Cottrell:
Thanks to all of you who have run the matrix_perf tests. This will
be
helpful in setting gretl's (internal, default) parameters for using the
system BLAS versus OpenMP (where available), versus our own
single-threaded matrix multiplication code.
Let me address some of the little "issues" that have come up, then I'll
offer a brief discussion of the results people have posted.
Issues
* Artur getting "error in function printvec, line 1": I think your gretl
sources must be not quite up to date. I'm not seeing this problem.
Indeed, re-compiling gretl helped to solve the problem.
Artur
12:12 p.m.
On Tue, 17 Jun 2014, Allin Cottrell wrote:
Thanks to all of you who have run the matrix_perf tests. This will be
helpful
in setting gretl's (internal, default) parameters for using the system BLAS
versus OpenMP (where available), versus our own single-threaded matrix
multiplication code.
Sorry I'm late. A few more interesting results here: two machines with
same operating system (64-bit debian). One is a low-end dual core, the
other is a modern machine with avx and two physical processors, each with
8 hyperthreaded cores.
The results follow, but I believe the moral of the story (also, having
seen the results others posted earlier) is quite evident: the "right"
software setup depends heavily on what your hardware/software combination
is.
Machine #1:
? matrix_perf(1234)
dgemm experiment 1, variant 1, speed in Gflops
m n k vanilla openmp netlib
128 128 128 1.1414 3.8052 4.4472
128 128 256 1.6901 3.9216 9.6005
128 128 512 1.7049 4.0992 12.939
128 128 1024 1.7066 4.0703 14.629
128 128 2048 1.6609 3.0855 14.598
result: netlib dominates
dgemm experiment 1, variant 2, speed in Gflops
m n k vanilla openmp netlib
128 128 128 1.6039 2.8291 12.559
256 256 128 1.6551 3.7319 13.157
512 512 128 1.5689 3.1081 12.157
1024 1024 128 1.7065 3.2241 13.810
2048 2048 128 1.4343 3.1975 12.901
result: netlib dominates
dgemm experiment 1, variant 3, speed in Gflops
m n k vanilla openmp netlib
128 128 128 1.6506 3.6948 9.4904
256 256 256 1.6570 3.6680 12.913
512 512 512 1.4917 3.4395 16.028
1024 1024 1024 0.70373 1.4884 18.831
2048 2048 2048 0.78776 1.5937 17.032
result: netlib dominates
dgemm experiment 2, variant 1, speed in Gflops
m n k vanilla openmp netlib
8 8 8 0.46703 0.37920 0.32284
16 8 8 0.63081 0.60029 0.54729
32 8 8 0.73601 0.89405 0.92477
64 8 8 0.90807 1.1178 1.3149
128 8 8 0.99528 1.4653 1.6911
256 8 8 1.0726 1.5460 1.5603
512 8 8 1.0186 1.8071 2.1102
1024 8 8 1.1191 1.6092 2.1806
2048 8 8 1.0810 1.6515 2.2472
4096 8 8 1.1118 1.6283 2.2618
result: netlib dominates for mnk >= 2048
vanilla dominates for mnk < 2048
dgemm experiment 2, variant 2, speed in Gflops
m n k vanilla openmp netlib
10 2 1000 1.4020 1.0541 2.9305
20 2 1000 1.2824 1.3566 3.2125
40 2 1000 1.4981 1.9113 2.3364
80 2 1000 1.6409 3.3097 3.0373
160 2 1000 1.6510 3.4486 2.8229
320 2 1000 1.2217 2.5190 2.2135
640 2 1000 0.80308 1.6429 1.9382
1280 2 1000 0.78715 1.5664 1.8491
2560 2 1000 0.71995 1.5641 1.8383
5120 2 1000 0.80175 1.5103 1.7516
result: netlib dominates for mnk >= 1280000
dgemm experiment 2, variant 3, speed in Gflops
m n k vanilla openmp netlib
10 10 1000 1.3822 2.9129 7.9036
20 10 1000 1.2999 3.3758 9.7947
40 10 1000 1.5159 3.2116 10.159
80 10 1000 1.6433 3.8072 11.100
160 10 1000 1.4395 4.1740 11.270
320 10 1000 1.1835 2.6255 9.5855
result: netlib dominates
Operating system: Linux (64-bit)
BLAS library: Netlib
Number of processors: 2
OpenMP enabled: yes
Performance summary:
vanilla -
dominates outright in 0 out of 6 tests
dominates in 1 test(s) for mnk < 2048
openmp -
dominates outright in 0 out of 6 tests
netlib -
dominates outright in 4 out of 6 tests
dominates in 2 test(s) for mnk >= (2048, 1280000)
Machine #2:
? matrix_perf(1234)
dgemm experiment 1, variant 1, speed in Gflops
m n k vanilla openmp netlib
128 128 128 0.90944 1.6489 3.6727
128 128 256 1.0361 12.066 4.0342
128 128 512 1.0363 13.360 3.2194
128 128 1024 2.1998 14.647 4.3699
128 128 2048 2.2040 15.157 4.4580
result: openmp dominates for mnk >= 4194304
netlib dominates for mnk < 4194304
dgemm experiment 1, variant 2, speed in Gflops
m n k vanilla openmp netlib
128 128 128 0.99176 7.2687 3.3558
256 256 128 1.0369 10.092 4.5736
512 512 128 1.0251 10.393 5.5707
1024 1024 128 1.0528 12.044 5.8882
2048 2048 128 1.5099 12.170 5.7919
result: openmp dominates
dgemm experiment 1, variant 3, speed in Gflops
m n k vanilla openmp netlib
128 128 128 1.0068 8.0279 3.1235
256 256 256 1.0678 12.979 5.2844
512 512 512 1.2698 14.865 6.8351
1024 1024 1024 2.3056 15.893 7.5533
2048 2048 2048 1.8804 19.930 10.979
result: openmp dominates
dgemm experiment 2, variant 1, speed in Gflops
m n k vanilla openmp netlib
8 8 8 0.31635 0.060822 0.40654
16 8 8 0.37316 0.13059 0.64134
32 8 8 0.48452 0.25061 0.89951
64 8 8 1.0088 0.35585 1.1348
128 8 8 1.2879 0.56582 1.3032
256 8 8 1.3405 0.83658 1.3995
512 8 8 1.3665 1.0493 0.61674
1024 8 8 1.3656 1.2329 0.87132
2048 8 8 1.3579 1.3256 1.0189
4096 8 8 1.3372 1.3590 0.59412
result: openmp dominates for mnk >= 262144
dgemm experiment 2, variant 2, speed in Gflops
m n k vanilla openmp netlib
10 2 1000 0.93204 0.27760 0.39253
20 2 1000 1.8316 0.51884 0.71238
40 2 1000 1.6916 0.60586 0.85907
80 2 1000 1.9516 1.0208 1.2188
160 2 1000 2.1520 1.1731 1.4970
320 2 1000 2.2201 1.4638 1.6964
640 2 1000 2.3148 1.8720 1.9165
1280 2 1000 2.2985 1.9854 2.3177
2560 2 1000 2.1321 2.4638 2.2796
5120 2 1000 1.8681 1.6833 1.5077
result: vanilla dominates for mnk >= 10240000
dgemm experiment 2, variant 3, speed in Gflops
m n k vanilla openmp netlib
10 10 1000 0.75487 0.64920 2.2521
20 10 1000 1.4513 0.99118 3.2809
40 10 1000 1.7493 1.5317 4.3326
80 10 1000 2.0249 2.1957 4.8481
160 10 1000 2.1919 3.3981 5.6694
320 10 1000 2.2979 5.4260 6.2356
result: netlib dominates
Operating system: Linux (64-bit)
BLAS library: Netlib
Number of processors: 32
OpenMP enabled: yes
Performance summary:
vanilla -
dominates outright in 0 out of 6 tests
dominates in 1 test(s) for mnk >= 10240000
openmp -
dominates outright in 2 out of 6 tests
dominates in 2 test(s) for mnk >= (4194304, 262144)
netlib -
dominates outright in 1 out of 6 tests
dominates in 1 test(s) for mnk < 4194304
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
2:10 p.m.
I'm also late to this party. Here are my results from 64-bit Windows 7:
? matrix_perf(0)
dgemm experiment 1, variant 1, speed in Gflops
m n k vanilla openmp openblas
128 128 128 0.51479 2.4324 0.49444
128 128 256 1.3617 2.8694 7.5620
128 128 512 1.3178 3.2178 7.3135
128 128 1024 1.4547 3.2327 6.8889
128 128 2048 1.3938 3.2093 7.5701
result: openblas dominates for mnk >= 4194304
openmp dominates for mnk < 4194304
dgemm experiment 1, variant 2, speed in Gflops
m n k vanilla openmp openblas
128 128 128 1.1285 2.8591 6.7398
256 256 128 1.3835 3.0657 4.7816
512 512 128 1.2706 3.1040 6.3758
1024 1024 128 1.3621 3.1572 6.9379
2048 2048 128 1.3089 3.0210 6.5357
result: openblas dominates
dgemm experiment 1, variant 3, speed in Gflops
m n k vanilla openmp openblas
128 128 128 1.3427 2.2489 3.7473
256 256 256 1.3720 3.2505 7.2717
512 512 512 1.4024 3.2591 7.0756
1024 1024 1024 1.1824 2.9743 8.2696
2048 2048 2048 1.1475 2.5411 8.1775
result: openblas dominates
dgemm experiment 2, variant 1, speed in Gflops
m n k vanilla openmp openblas
8 8 8 0.51573 0.037176 0.57536
16 8 8 0.63935 0.073940 0.86400
32 8 8 0.72757 0.14150 1.3334
64 8 8 0.82663 0.26622 1.5487
128 8 8 0.78510 0.44572 1.9989
256 8 8 0.90749 0.68418 1.2459
512 8 8 0.96497 0.88337 1.8486
1024 8 8 0.87171 1.1631 1.7002
2048 8 8 0.95462 1.2275 2.0776
4096 8 8 0.78187 1.4584 1.9049
result: openblas dominates
dgemm experiment 2, variant 2, speed in Gflops
m n k vanilla openmp openblas
10 2 1000 1.2314 0.62643 2.7906
20 2 1000 1.2397 0.97835 3.0542
40 2 1000 1.2850 1.5531 3.4026
80 2 1000 1.3963 2.2204 3.3032
160 2 1000 1.3556 2.3921 2.9239
320 2 1000 1.3506 2.1077 2.8484
640 2 1000 1.1736 1.9558 2.0293
1280 2 1000 1.0521 1.8458 1.9304
2560 2 1000 1.0719 1.7515 1.7272
5120 2 1000 1.0384 1.9335 1.8989
result: openmp dominates for mnk >= 5120000
openblas dominates for mnk < 5120000
dgemm experiment 2, variant 3, speed in Gflops
m n k vanilla openmp openblas
10 10 1000 1.0817 1.9209 4.1470
20 10 1000 1.0978 2.4580 4.2459
40 10 1000 1.2954 2.1036 6.0306
80 10 1000 1.3542 1.9040 6.2067
160 10 1000 1.4430 2.0096 5.9852
320 10 1000 1.4470 2.0040 5.4259
result: openblas dominates
Operating system: Windows (64-bit)
BLAS library: OpenBLAS
Number of processors: 4
OpenMP enabled: yes
Performance summary:
vanilla -
dominates outright in 0 out of 6 tests
openmp -
dominates outright in 0 out of 6 tests
dominates in 1 test(s) for mnk >= 5120000
dominates in 1 test(s) for mnk < 4194304
openblas -
dominates outright in 4 out of 6 tests
dominates in 1 test(s) for mnk >= 4194304
dominates in 1 test(s) for mnk < 5120000
PS
-----Original Message-----
From: gretl-devel-bounces(a)lists.wfu.edu [mailto:gretl-devel-bounces@lists.wfu.edu] On
Behalf Of Riccardo (Jack) Lucchetti
Sent: Wednesday, June 18, 2014 12:13 PM
To: Gretl development
Subject: Re: [Gretl-devel] matrix_perf results
On Tue, 17 Jun 2014, Allin Cottrell wrote:
Thanks to all of you who have run the matrix_perf tests. This will be
helpful in setting gretl's (internal, default) parameters for using
the system BLAS versus OpenMP (where available), versus our own
single-threaded matrix multiplication code.
Sorry I'm late. A few more interesting results here: two machines with same operating
system (64-bit debian). One is a low-end dual core, the other is a modern machine with avx
and two physical processors, each with
8 hyperthreaded cores.
The results follow, but I believe the moral of the story (also, having seen the results
others posted earlier) is quite evident: the "right"
software setup depends heavily on what your hardware/software combination is.
Machine #1:
? matrix_perf(1234)
dgemm experiment 1, variant 1, speed in Gflops
m n k vanilla openmp netlib
128 128 128 1.1414 3.8052 4.4472
128 128 256 1.6901 3.9216 9.6005
128 128 512 1.7049 4.0992 12.939
128 128 1024 1.7066 4.0703 14.629
128 128 2048 1.6609 3.0855 14.598
result: netlib dominates
dgemm experiment 1, variant 2, speed in Gflops
m n k vanilla openmp netlib
128 128 128 1.6039 2.8291 12.559
256 256 128 1.6551 3.7319 13.157
512 512 128 1.5689 3.1081 12.157
1024 1024 128 1.7065 3.2241 13.810
2048 2048 128 1.4343 3.1975 12.901
result: netlib dominates
dgemm experiment 1, variant 3, speed in Gflops
m n k vanilla openmp netlib
128 128 128 1.6506 3.6948 9.4904
256 256 256 1.6570 3.6680 12.913
512 512 512 1.4917 3.4395 16.028
1024 1024 1024 0.70373 1.4884 18.831
2048 2048 2048 0.78776 1.5937 17.032
result: netlib dominates
dgemm experiment 2, variant 1, speed in Gflops
m n k vanilla openmp netlib
8 8 8 0.46703 0.37920 0.32284
16 8 8 0.63081 0.60029 0.54729
32 8 8 0.73601 0.89405 0.92477
64 8 8 0.90807 1.1178 1.3149
128 8 8 0.99528 1.4653 1.6911
256 8 8 1.0726 1.5460 1.5603
512 8 8 1.0186 1.8071 2.1102
1024 8 8 1.1191 1.6092 2.1806
2048 8 8 1.0810 1.6515 2.2472
4096 8 8 1.1118 1.6283 2.2618
result: netlib dominates for mnk >= 2048
vanilla dominates for mnk < 2048
dgemm experiment 2, variant 2, speed in Gflops
m n k vanilla openmp netlib
10 2 1000 1.4020 1.0541 2.9305
20 2 1000 1.2824 1.3566 3.2125
40 2 1000 1.4981 1.9113 2.3364
80 2 1000 1.6409 3.3097 3.0373
160 2 1000 1.6510 3.4486 2.8229
320 2 1000 1.2217 2.5190 2.2135
640 2 1000 0.80308 1.6429 1.9382
1280 2 1000 0.78715 1.5664 1.8491
2560 2 1000 0.71995 1.5641 1.8383
5120 2 1000 0.80175 1.5103 1.7516
result: netlib dominates for mnk >= 1280000
dgemm experiment 2, variant 3, speed in Gflops
m n k vanilla openmp netlib
10 10 1000 1.3822 2.9129 7.9036
20 10 1000 1.2999 3.3758 9.7947
40 10 1000 1.5159 3.2116 10.159
80 10 1000 1.6433 3.8072 11.100
160 10 1000 1.4395 4.1740 11.270
320 10 1000 1.1835 2.6255 9.5855
result: netlib dominates
Operating system: Linux (64-bit)
BLAS library: Netlib
Number of processors: 2
OpenMP enabled: yes
Performance summary:
vanilla -
dominates outright in 0 out of 6 tests
dominates in 1 test(s) for mnk < 2048
openmp -
dominates outright in 0 out of 6 tests
netlib -
dominates outright in 4 out of 6 tests
dominates in 2 test(s) for mnk >= (2048, 1280000)
Machine #2:
? matrix_perf(1234)
dgemm experiment 1, variant 1, speed in Gflops
m n k vanilla openmp netlib
128 128 128 0.90944 1.6489 3.6727
128 128 256 1.0361 12.066 4.0342
128 128 512 1.0363 13.360 3.2194
128 128 1024 2.1998 14.647 4.3699
128 128 2048 2.2040 15.157 4.4580
result: openmp dominates for mnk >= 4194304
netlib dominates for mnk < 4194304
dgemm experiment 1, variant 2, speed in Gflops
m n k vanilla openmp netlib
128 128 128 0.99176 7.2687 3.3558
256 256 128 1.0369 10.092 4.5736
512 512 128 1.0251 10.393 5.5707
1024 1024 128 1.0528 12.044 5.8882
2048 2048 128 1.5099 12.170 5.7919
result: openmp dominates
dgemm experiment 1, variant 3, speed in Gflops
m n k vanilla openmp netlib
128 128 128 1.0068 8.0279 3.1235
256 256 256 1.0678 12.979 5.2844
512 512 512 1.2698 14.865 6.8351
1024 1024 1024 2.3056 15.893 7.5533
2048 2048 2048 1.8804 19.930 10.979
result: openmp dominates
dgemm experiment 2, variant 1, speed in Gflops
m n k vanilla openmp netlib
8 8 8 0.31635 0.060822 0.40654
16 8 8 0.37316 0.13059 0.64134
32 8 8 0.48452 0.25061 0.89951
64 8 8 1.0088 0.35585 1.1348
128 8 8 1.2879 0.56582 1.3032
256 8 8 1.3405 0.83658 1.3995
512 8 8 1.3665 1.0493 0.61674
1024 8 8 1.3656 1.2329 0.87132
2048 8 8 1.3579 1.3256 1.0189
4096 8 8 1.3372 1.3590 0.59412
result: openmp dominates for mnk >= 262144
dgemm experiment 2, variant 2, speed in Gflops
m n k vanilla openmp netlib
10 2 1000 0.93204 0.27760 0.39253
20 2 1000 1.8316 0.51884 0.71238
40 2 1000 1.6916 0.60586 0.85907
80 2 1000 1.9516 1.0208 1.2188
160 2 1000 2.1520 1.1731 1.4970
320 2 1000 2.2201 1.4638 1.6964
640 2 1000 2.3148 1.8720 1.9165
1280 2 1000 2.2985 1.9854 2.3177
2560 2 1000 2.1321 2.4638 2.2796
5120 2 1000 1.8681 1.6833 1.5077
result: vanilla dominates for mnk >= 10240000
dgemm experiment 2, variant 3, speed in Gflops
m n k vanilla openmp netlib
10 10 1000 0.75487 0.64920 2.2521
20 10 1000 1.4513 0.99118 3.2809
40 10 1000 1.7493 1.5317 4.3326
80 10 1000 2.0249 2.1957 4.8481
160 10 1000 2.1919 3.3981 5.6694
320 10 1000 2.2979 5.4260 6.2356
result: netlib dominates
Operating system: Linux (64-bit)
BLAS library: Netlib
Number of processors: 32
OpenMP enabled: yes
Performance summary:
vanilla -
dominates outright in 0 out of 6 tests
dominates in 1 test(s) for mnk >= 10240000
openmp -
dominates outright in 2 out of 6 tests
dominates in 2 test(s) for mnk >= (4194304, 262144)
netlib -
dominates outright in 1 out of 6 tests
dominates in 1 test(s) for mnk < 4194304
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
2:43 p.m.
On Wed, 18 Jun 2014, Summers, Peter wrote:
I'm also late to this party. Here are my results from 64-bit
Windows 7:
Thanks, Peter, this is encouraging with regard to the new Windows
build. Maybe this is worth trying for 32-bit Windows too.
Small comment: that poor result from openblas on the very first test
is probably just a matter of warming up the machine's vector
registers; if you were to run it a second time you'd probably see
openblas dominating from the get-go.
? matrix_perf(0)
dgemm experiment 1, variant 1, speed in Gflops
m n k vanilla openmp openblas
128 128 128 0.51479 2.4324 0.49444
128 128 256 1.3617 2.8694 7.5620
128 128 512 1.3178 3.2178 7.3135
128 128 1024 1.4547 3.2327 6.8889
128 128 2048 1.3938 3.2093 7.5701
result: openblas dominates for mnk >= 4194304
openmp dominates for mnk < 4194304
[...]
Allin
2:54 p.m.
Small comment: that poor result from openblas on the very first test
is probably just a matter of warming up the machine's vector registers; if you were to
run it a second time you'd
probably see openblas dominating from the get-go.
Right you are, Allin. The first test was pretty much the first thing I did after booting
up my computer. 2 subsequent runs show near-total openblas domination: 6/6 cases in one
run, 5/6 in the other. The lone exception was
dgemm experiment 2, variant 2, speed in Gflops
m n k vanilla openmp openblas
10 2 1000 1.1782 0.77966 2.1771
20 2 1000 1.3308 0.99745 2.6331
40 2 1000 1.3111 1.3642 2.7879
80 2 1000 1.0477 1.4362 2.8529
160 2 1000 1.2724 1.8858 2.6338
320 2 1000 1.1903 1.8180 2.4244
640 2 1000 1.1162 1.6754 1.6777
1280 2 1000 1.1157 1.6019 1.5981
2560 2 1000 1.0857 1.7744 1.7199
5120 2 1000 1.1078 1.6811 1.7112
result: openblas dominates for mnk >= 10240000
PS
-----Original Message-----
From: gretl-devel-bounces(a)lists.wfu.edu [mailto:gretl-devel-bounces@lists.wfu.edu] On
Behalf Of Allin Cottrell
Sent: Wednesday, June 18, 2014 2:44 PM
To: Gretl development
Subject: Re: [Gretl-devel] matrix_perf results
On Wed, 18 Jun 2014, Summers, Peter wrote:
I'm also late to this party. Here are my results from 64-bit
Windows
7:
Thanks, Peter, this is encouraging with regard to the new Windows build. Maybe this is
worth trying for 32-bit Windows too.
Small comment: that poor result from openblas on the very first test is probably just a
matter of warming up the machine's vector registers; if you were to run it a second
time you'd probably see openblas dominating from the get-go.
? matrix_perf(0)
dgemm experiment 1, variant 1, speed in Gflops
m n k vanilla openmp openblas
128 128 128 0.51479 2.4324 0.49444
128 128 256 1.3617 2.8694 7.5620
128 128 512 1.3178 3.2178 7.3135
128 128 1024 1.4547 3.2327 6.8889
128 128 2048 1.3938 3.2093 7.5701
result: openblas dominates for mnk >= 4194304 openmp dominates for
mnk < 4194304
[...]
Allin
_______________________________________________
Gretl-devel mailing list
Gretl-devel(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-devel
3:05 p.m.
On Wed, 18 Jun 2014, Summers, Peter wrote:
> Small comment: that poor result from openblas on the very first
> test is probably just a matter of warming up the machine's vector
> registers; if you were to run it a second time you'd probably see
> openblas dominating from the get-go.
Right you are, Allin. The first test was pretty much the first
thing I did after booting up my computer. 2 subsequent runs show
near-total openblas domination: 6/6 cases in one run, 5/6 in the
other. The lone exception was
dgemm experiment 2, variant 2, speed in Gflops
m n k vanilla openmp openblas
10 2 1000 1.1782 0.77966 2.1771
20 2 1000 1.3308 0.99745 2.6331
40 2 1000 1.3111 1.3642 2.7879
80 2 1000 1.0477 1.4362 2.8529
160 2 1000 1.2724 1.8858 2.6338
320 2 1000 1.1903 1.8180 2.4244
640 2 1000 1.1162 1.6754 1.6777
1280 2 1000 1.1157 1.6019 1.5981
2560 2 1000 1.0857 1.7744 1.7199
5120 2 1000 1.1078 1.6811 1.7112
result: openblas dominates for mnk >= 10240000
Thanks. Yes, this last result is expected: it's the one "known hole"
in openblas performance. By default, openblas does not do
multi-threading if any of the matrix dimensions m, n, or k is less
than 4. In other batches of results (though not really here) I've
seen openmp do significantly better for matrices that are big but
have one small dimension < 4.
Allin
11:29 a.m.
I just want to add another result obtained on Windows 7 (64-bit) but using
gretl's 32-bit version:
<OUTPUT>
dgemm experiment 1, variant 1, speed in Gflops
m n k vanilla openmp netlib
128 128 128 2.1368 3.4852 2.3743
128 128 256 2.5487 4.4178 2.5388
128 128 512 2.7106 5.0125 2.7782
128 128 1024 2.7998 5.4136 2.8940
128 128 2048 2.7670 5.3196 2.9199
result: openmp dominates
dgemm experiment 1, variant 2, speed in Gflops
m n k vanilla openmp netlib
128 128 128 2.3372 3.4591 2.3988
256 256 128 2.3256 4.4618 2.7094
512 512 128 2.1448 4.6779 2.7654
1024 1024 128 2.4723 4.7650 2.7767
2048 2048 128 2.5796 4.6320 2.7367
result: openmp dominates
dgemm experiment 1, variant 3, speed in Gflops
m n k vanilla openmp netlib
128 128 128 2.6201 3.3397 2.3657
256 256 256 2.8571 5.0325 2.9471
512 512 512 3.1156 4.9564 3.1536
1024 1024 1024 2.3233 5.2080 2.3002
2048 2048 2048 2.3192 4.4718 2.3189
result: openmp dominates
dgemm experiment 2, variant 1, speed in Gflops
m n k vanilla openmp netlib
8 8 8 0.44731 0.027292 0.46398
16 8 8 0.56046 0.050733 0.62706
32 8 8 0.66725 0.10275 0.66642
64 8 8 0.73151 0.18160 0.74646
128 8 8 0.78491 0.30364 0.79847
256 8 8 0.80797 0.45798 0.82017
512 8 8 0.79569 0.61589 0.78943
1024 8 8 0.97164 0.75438 0.82914
2048 8 8 0.83483 0.80910 0.71419
4096 8 8 0.84114 0.85660 0.83595
result: openmp dominates for mnk >= 262144
dgemm experiment 2, variant 2, speed in Gflops
m n k vanilla openmp netlib
10 2 1000 1.9133 0.59579 2.2596
20 2 1000 2.4185 0.93620 2.6299
40 2 1000 2.4249 1.5991 2.4208
80 2 1000 2.5991 2.3573 2.7317
160 2 1000 2.8793 3.3007 2.9413
320 2 1000 2.9538 4.5477 2.9906
640 2 1000 2.2554 3.4918 2.2917
1280 2 1000 2.2609 3.7071 2.2745
2560 2 1000 2.2118 3.3263 2.2296
5120 2 1000 2.2168 3.4272 2.2340
result: openmp dominates for mnk >= 320000
dgemm experiment 2, variant 3, speed in Gflops
m n k vanilla openmp netlib
10 10 1000 1.9689 1.9801 2.3104
20 10 1000 2.4662 3.0143 2.7699
40 10 1000 2.4801 3.8146 2.4038
80 10 1000 2.6682 4.2088 2.7531
160 10 1000 2.9370 4.6238 2.9567
320 10 1000 2.9992 4.6294 2.9979
result: openmp dominates for mnk >= 200000
netlib dominates for mnk < 200000
Operating system: Windows (32-bit)
BLAS library: Netlib
Number of processors: 4
OpenMP enabled: yes
Performance summary:
vanilla -
dominates outright in 0 out of 6 tests
openmp -
dominates outright in 3 out of 6 tests
dominates in 3 test(s) for mnk >= (262144, 320000, 200000)
netlib -
dominates outright in 0 out of 6 tests
dominates in 1 test(s) for mnk < 200000
</OUTPUT>
Artur
2014-06-18 21:05 GMT+02:00 Allin Cottrell <cottrell(a)wfu.edu>:
On Wed, 18 Jun 2014, Summers, Peter wrote:
>> Small comment: that poor result from openblas on the very first
>> test is probably just a matter of warming up the machine's vector
>> registers; if you were to run it a second time you'd probably see
>> openblas dominating from the get-go.
>
> Right you are, Allin. The first test was pretty much the first
> thing I did after booting up my computer. 2 subsequent runs show
> near-total openblas domination: 6/6 cases in one run, 5/6 in the
> other. The lone exception was
>
> dgemm experiment 2, variant 2, speed in Gflops
>
> m n k vanilla openmp openblas
> 10 2 1000 1.1782 0.77966 2.1771
> 20 2 1000 1.3308 0.99745 2.6331
> 40 2 1000 1.3111 1.3642 2.7879
> 80 2 1000 1.0477 1.4362 2.8529
> 160 2 1000 1.2724 1.8858 2.6338
> 320 2 1000 1.1903 1.8180 2.4244
> 640 2 1000 1.1162 1.6754 1.6777
> 1280 2 1000 1.1157 1.6019 1.5981
> 2560 2 1000 1.0857 1.7744 1.7199
> 5120 2 1000 1.1078 1.6811 1.7112
>
> result: openblas dominates for mnk >= 10240000
Thanks. Yes, this last result is expected: it's the one "known hole"
in openblas performance. By default, openblas does not do
multi-threading if any of the matrix dimensions m, n, or k is less
than 4. In other batches of results (though not really here) I've
seen openmp do significantly better for matrices that are big but
have one small dimension < 4.
Allin
_______________________________________________
Gretl-devel mailing list
Gretl-devel(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-devel
2:37 p.m.
On Wed, 18 Jun 2014, Riccardo (Jack) Lucchetti wrote:
Sorry I'm late. A few more interesting results here: two machines
with same
operating system (64-bit debian). One is a low-end dual core, the other is a
modern machine with avx and two physical processors, each with 8
hyperthreaded cores.
The results follow [snipped here], but I believe the moral of the
story (also, having seen a the results others posted earlier) is
quite evident: the "right" software setup depends heavily on what
your hardware/software combination is.
Yes, true enough. Though I think we're on safe ground in reckoning
that on the Mac we ought to hand off dgemm to VecLib except perhaps
for very small matrices.
Can I ask -- this applies to Marcin too -- the BLAS that's
identified by gretl as "netlib": is this really netlib and not ATLAS
or OpenBLAS in disguise? Based on Artur's information I can believe
it is in fact plain Netlib BLAS with heavy optimization by the
compiler, but I'd just like to check.
Allin
3800
days inactive
3806
days old
11 comments
6 participants
participants (6)
-
Allin Cottrell -
Artur T. -
Hélio Guilherme -
Marcin Błażejowski -
Riccardo (Jack) Lucchetti -
Summers, Peter