Thread Subject:

Intel Core Duo & MATLAB

Hi,
I just bought a new notebook computer from Dell. My new computer
sports a Intel Core Duo processor @ 1.67GHz (each core) with 1GB in
RAM.
I currenly work with lengthy computations in the area of Nonlinear
Dynamic Structural Analysis and need a lot of number crunching power.
Thats why I prefered a dual core processor in the first place.
But then I came to realize that when performing my simulations
MATLAB uses only one processor core instead of both. I soon came to
realize that MATLAB is one-threaded.
This is not too bad beacause my computer still ranks high in MATLAB
bench, but it makes for less crunching power than a 3.2 Gig
processor, for instance, which is totally not what I was expecting.
There are a few projects out there for incorporating
multiple-processor capability on Matlab (like Cornell University's
Multi-MATLAB, which is not currently supporting PCs).
I guess the Mathworks should take into account that due to the
incapability of creating faster processors at current
state-of-the-art, the natural way of increasing processor power is
actually having more processor cores and taking advantage of
multi-threaded programming. But MATLAB, if continued as a
single-threaded program, will be unable to untap this power even if
you have a zillion processor computer at home.
This is a call into caution for users. Dual-core processing is nice,
but useless if you are looking for extreme processing power in
MATLAB. Right now you should prefer a faster single-core processor.
One advantage, and one which I've come to find very usefull, is that
you can actually have two (2!!) MATLAB processes running on one
computer with virtually no drop in performance. Or you can open other
programs (like serious performance demanding games) and use them as
if MATLAB wasn't even there. That's a serious plus.
Just be warned that when thinking of a Dual-Core processor for
running MATLAB you just might not get what you expect.
Everyday I find there is much to learn about MATLAB's powers and
limitations.
(So, if anyone out there knows another way of getting matlab into
using both of my cores it would be very nice!)

Hi Jose,

Matlab is single-threaded but BLAS doesn't have to be.
Since your using an intel chip, and therefore the intel MKL for your
BLAS library (the library that handles your matrix operations) you
can set it to use multiple threads.

You need to create an enviornment variable called "OMP_NUM_THREADS"
and set it equal to 2.

In windows this is done via the control panel --> system -->
enivornment variables

In linux its "export OMP_NUM_THREADS 2"

You'll notice that both processors will be used for something like
say inv(rand(5000)), and that it will happen much faster then using
one.

Please let me know if this speeds up your work and by how much.

Cheers,

Chris

What exactly is BLAS? Where can I get it? Is it something like an
'addon' for MATLAB?
The thing is I use a toolbox for my structural dynamics projects, so
I am highly dependant on using MATLAB.

Thanks for your reply.
Jose Abell

BLAS is Basic Linear Algebra Subprograms.

It is some set of built-in functions (in our case in MATLAB) that is
programmed in, say, C, for better performance. This is a part of MATLAB. The
only thing you should do to enable your 2nd core support is, as Chris said,

In windows this is done via the control panel --> system -->
enivornment variables --> create variable --> OMP_NUM_THREADS and set the
value of this variable to 2.

Can you post here the results of bench(5) before and after this procedure?

P.S. Command bench takes into account not only purely "computational power"
whatever it means, but also 2D & 3D performance. As a result, since
notebooks have weaker graphical system, you can see "worse average
performance", even though your CPU is better...

Arthur


"Jose Abell" <jaabell@ing.puc.cl> wrote in message
news:ef3eb7a.1@webcrossing.raydaftYaTP...
> What exactly is BLAS? Where can I get it? Is it something like an
> 'addon' for MATLAB?
> The thing is I use a toolbox for my structural dynamics projects, so
> I am highly dependant on using MATLAB.
>
> Thanks for your reply.
> Jose Abell

That worked out quite fine. I could see both of my cores working. The
only 'but' is that my 2d graphics got worse. Here are the results
from bench(5).

Before Activating 2nd Core

ans =

 0.5098 0.2960 0.1974 0.5950 0.6142 0.7051
 0.5012 0.2926 0.1947 0.5951 0.5597 0.7299
 0.4989 0.2926 0.1986 0.5937 0.6165 0.7197
 0.5043 0.2930 0.1940 0.5949 0.5393 0.7177
 0.5004 0.2927 0.1944 0.5945 0.5428 0.7279

Mean
 0.5029 0.2934 0.1958 0.5946 0.5745 0.7201

After...

ans =

 0.3110 0.2891 0.1919 0.5879 1.0045 0.7477
 0.2897 0.2891 0.1919 0.5885 1.0038 0.7382
 0.3031 0.2880 0.1919 0.5862 1.0030 0.7510
 0.2939 0.2879 0.1922 0.5875 1.0192 0.7242
 0.2879 0.2880 0.1908 0.5880 1.0051 0.7523

Mean
 0.2971 0.2884 0.1918 0.5876 1.0071 0.7427

In average everything dropped a few Thanks a lot, to both of you.
This helps a lot!
Are there other 'tweaks' that could help increase my performance?

JA

José Abell wrote:
>
>
> That worked out quite fine. I could see both of my cores working.
> The
> only 'but' is that my 2d graphics got worse. Here are the results
> from bench(5).
>
> Before Activating 2nd Core
>
> ans =
>
> 0.5098 0.2960 0.1974 0.5950 0.6142 0.7051
> 0.5012 0.2926 0.1947 0.5951 0.5597 0.7299
> 0.4989 0.2926 0.1986 0.5937 0.6165 0.7197
> 0.5043 0.2930 0.1940 0.5949 0.5393 0.7177
> 0.5004 0.2927 0.1944 0.5945 0.5428 0.7279
>
> Mean
> 0.5029 0.2934 0.1958 0.5946 0.5745 0.7201
>
> After...
>
> ans =
>
> 0.3110 0.2891 0.1919 0.5879 1.0045 0.7477
> 0.2897 0.2891 0.1919 0.5885 1.0038 0.7382
> 0.3031 0.2880 0.1919 0.5862 1.0030 0.7510
> 0.2939 0.2879 0.1922 0.5875 1.0192 0.7242
> 0.2879 0.2880 0.1908 0.5880 1.0051 0.7523
>
> Mean
> 0.2971 0.2884 0.1918 0.5876 1.0071 0.7427
>
> In average everything dropped a few Thanks a lot, to both of you.
> This helps a lot!
> Are there other 'tweaks' that could help increase my performance?
>
> JA

Hello group.

If you had a wish list of the best hardware configuration for
computing MatLab. We are using tools to draw on CT images and have
some lag time in the drawing tools presumably due to computation.

Any recommendations would be great

Josh
jasper@ikoemed.com

Chris McIntosh wrote:
>
>
> Hi Jose,
>
> Matlab is single-threaded but BLAS doesn't have to be.
> Since your using an intel chip, and therefore the intel MKL for
> your
> BLAS library (the library that handles your matrix operations) you
> can set it to use multiple threads.
>
> You need to create an enviornment variable called "OMP_NUM_THREADS"
> and set it equal to 2.
>
> In windows this is done via the control panel --> system -->
> enivornment variables
>
> In linux its "export OMP_NUM_THREADS 2"

Hello group.
 
If you had a wish list of the best hardware configuration for
computing MatLab. We are using tools to draw on CT images and have
some lag time in the drawing tools presumably due to computation.
 
Any recommendations would be great. Tried your environmental changes
this helped.
 
Josh
jasper@ikoemed.com
>
> You'll notice that both processors will be used for something like
> say inv(rand(5000)), and that it will happen much faster then using
> one.
>
> Please let me know if this speeds up your work and by how much.
>
> Cheers,
>
> Chris

Arthur wrote:
>
>
> BLAS is Basic Linear Algebra Subprograms.
>
> It is some set of built-in functions (in our case in MATLAB) that
> is
> programmed in, say, C, for better performance. This is a part of
> MATLAB. The
> only thing you should do to enable your 2nd core support is, as
> Chris said,
>
> In windows this is done via the control panel --> system -->
> enivornment variables --> create variable --> OMP_NUM_THREADS
and
> set the
> value of this variable to 2.
>

I am not getting multithreaded BLAS behavior using this method on a
Turion X2 TL-60.

I also tried using the Goto BLAS and that uses a variable called
GOTO_NUM_THREADS. Still only using one core.

I am able to see in the MATLAB session that the environment variables
are set the way I expect. They just do not seem to use the second
core.

Any suggestions?

I am no expert in this area but Chris said when he posted "Since your
using an intel chip...". The Turion is an AMD chip so this could be a
reason. It might be worth investigating other BLAS modifications that
are more suited for your hardware. I know that people have had
favorable reports of the gotoBlas with Turion processors

Perhaps this thread will help you more
 <http://groups.google.co.uk/group/comp.soft-sys.matlab/browse_thread/thread/3a98d613a27c56b6/ee31bb7a7711d964?lnk=gst&q=BLAS+AMD&rnum=8#ee31bb7a7711d964>

Dear all,

I also bought a new Dell inspiron 9400 Laptop (Intel Core 2 Duo
T7200[2GHz] with 2GB RAM)

Primary reason for this purchase was to spead up my research. I do not
see a major increase in computational power (Compared to Pentium 4
2.4GHz)

PLEASE NOTE that I run Matlab version 6.5 !!!!

Correct me if I'm wrong but Matlab(at least 6.5) appears to be purely
one threaded program. How do I know that? Well, I run bench(5) on
Matlab:

0.391 0.406 0.266 0.406 0.39 0.172
0.406 0.406 0.25 0.406 0.39 0.172
0.406 0.391 0.25 0.407 0.391 0.172
0.406 0.406 0.266 0.406 0.39 0.172
0.406 0.406 0.25 0.391 0.375 0.172

Now comes the funny part... I start to Matlab processes and let them
run bench at the same time (concurently) !!!
Start bench(8) on the first proces than start bench(5) on the other.
Running concurently:

Here is the first process:

   0.4060 0.4060 0.2660 0.4380 0.3750 0.1560
    0.4220 0.3910 0.2810 0.4060 0.2650 0
    0.4220 0.4070 0.2500 0.3910 0.2500 0
    0.4060 0.4060 0.2660 0.3900 0.2340 0
    0.4060 0.4060 0.2650 0.4060 0.2660 0
    0.4220 0.4060 0.2660 0.4060 0.2650 0
    0.4060 0.4220 0.2500 0.3900 0.2500 0.0150
    0.4060 0.4060 0.2660 0.4060 0.2500 0.0160

Here is the second process:

0.406 0.407 0.266 0.406 0.438 0.187
0.407 0.406 0.266 0.406 0.406 0.172
0.406 0.391 0.265 0.406 0.391 0.172
0.407 0.39 0.25 0.422 0.391 0.157
0.406 0.391 0.266 0.438 0.375 0.172

Funny how the first process captures the graph produceed by the second
process in 0 secs (while the second process was running)

But forget 3D, look at LU and FTT: same sh..t!

I tried the setting OMP_NUM_THREADS to 2, nothing changes. NOTHING!

Last remark - I'm running on Windows Media Center 2005.

Any ideas how to better utilize this beast I have on my desk?

Or I will be forced to use one core only? Maybe this is Matlab version
issue?

Cheers!

Dan Pearce wrote:
> I am no expert in this area but Chris said when he posted "Since your
> using an intel chip...". The Turion is an AMD chip so this could be a
> reason. It might be worth investigating other BLAS modifications that
> are more suited for your hardware. I know that people have had
> favorable reports of the gotoBlas with Turion processors
>
> Perhaps this thread will help you more
> <http://groups.google.co.uk/group/comp.soft-sys.matlab/browse_thread/thread/3a98d613a27c56b6/ee31bb7a7711d964?lnk=gst&q=BLAS+AMD&rnum=8#ee31bb7a7711d964>

On Wed, 20 Sep 2006 18:30:37 -0700, marius.jurgilas wrote:

> Dear all,
>
> I also bought a new Dell inspiron 9400 Laptop (Intel Core 2 Duo
> T7200[2GHz] with 2GB RAM)
>
> Primary reason for this purchase was to spead up my research. I do not
> see a major increase in computational power (Compared to Pentium 4
> 2.4GHz)
>
> PLEASE NOTE that I run Matlab version 6.5 !!!!
>
> Correct me if I'm wrong but Matlab(at least 6.5) appears to be purely
> one threaded program. How do I know that? Well, I run bench(5) on
> Matlab:


Yes, Matlab was and still is single threaded. This is well known.


> 0.391 0.406 0.266 0.406 0.39 0.172
> 0.406 0.406 0.25 0.406 0.39 0.172
> 0.406 0.391 0.25 0.407 0.391 0.172
> 0.406 0.406 0.266 0.406 0.39 0.172
> 0.406 0.406 0.25 0.391 0.375 0.172
>
> Now comes the funny part... I start to Matlab processes and let them
> run bench at the same time (concurently) !!!
> Start bench(8) on the first proces than start bench(5) on the other.
> Running concurently:
>
> Here is the first process:
>
> 0.4060 0.4060 0.2660 0.4380 0.3750 0.1560
> 0.4220 0.3910 0.2810 0.4060 0.2650 0
> 0.4220 0.4070 0.2500 0.3910 0.2500 0
> 0.4060 0.4060 0.2660 0.3900 0.2340 0
> 0.4060 0.4060 0.2650 0.4060 0.2660 0
> 0.4220 0.4060 0.2660 0.4060 0.2650 0
> 0.4060 0.4220 0.2500 0.3900 0.2500 0.0150
> 0.4060 0.4060 0.2660 0.4060 0.2500 0.0160
>
> Here is the second process:
>
> 0.406 0.407 0.266 0.406 0.438 0.187
> 0.407 0.406 0.266 0.406 0.406 0.172
> 0.406 0.391 0.265 0.406 0.391 0.172
> 0.407 0.39 0.25 0.422 0.391 0.157
> 0.406 0.391 0.266 0.438 0.375 0.172
>
> Funny how the first process captures the graph produceed by the second
> process in 0 secs (while the second process was running)


This is strange. I don't know why this is.


> But forget 3D, look at LU and FTT: same sh..t!


Not surprising at all. Each Matlab process gets 1 CPU, and uses it fully.


> I tried the setting OMP_NUM_THREADS to 2, nothing changes. NOTHING!


This may not work with such an old version of Matlab (BLAS). Not sure
when this came into effect. It definitely does work with R2006a.

> Last remark - I'm running on Windows Media Center 2005.
>
> Any ideas how to better utilize this beast I have on my desk?
>
> Or I will be forced to use one core only? Maybe this is Matlab version
> issue?


You're stuck with one core at a time. I guess you could try running 2
separate Matlab sessions to parallelize your work if it's possible. Other
than that, you could upgrade to the latest Matlab and use the environment
variable to make BLAS multi-threaded.

Dan

Thanks Dan,

Upgraded to 2006a (academic) + BLAS multi-threaded

MAJOR difference:

    0.1855 0.2775 0.1733 0.5147 0.3866 0.2277
    0.1724 0.2626 0.1762 0.5128 0.3780 0.2226
    0.1806 0.2630 0.1758 0.5072 0.3778 0.2266
    0.1625 0.2599 0.1737 0.5069 0.3965 0.2270
    0.1651 0.2577 0.1741 0.4985 0.3815 0.2153

Dan Hensley wrote:
> On Wed, 20 Sep 2006 18:30:37 -0700, marius.jurgilas wrote:
>
> > Dear all,
> >
> > I also bought a new Dell inspiron 9400 Laptop (Intel Core 2 Duo
> > T7200[2GHz] with 2GB RAM)
> >
> > Primary reason for this purchase was to spead up my research. I do not
> > see a major increase in computational power (Compared to Pentium 4
> > 2.4GHz)
> >
> > PLEASE NOTE that I run Matlab version 6.5 !!!!
> >
> > Correct me if I'm wrong but Matlab(at least 6.5) appears to be purely
> > one threaded program. How do I know that? Well, I run bench(5) on
> > Matlab:
>
>
> Yes, Matlab was and still is single threaded. This is well known.
>
>
> > 0.391 0.406 0.266 0.406 0.39 0.172
> > 0.406 0.406 0.25 0.406 0.39 0.172
> > 0.406 0.391 0.25 0.407 0.391 0.172
> > 0.406 0.406 0.266 0.406 0.39 0.172
> > 0.406 0.406 0.25 0.391 0.375 0.172
> >
> > Now comes the funny part... I start to Matlab processes and let them
> > run bench at the same time (concurently) !!!
> > Start bench(8) on the first proces than start bench(5) on the other.
> > Running concurently:
> >
> > Here is the first process:
> >
> > 0.4060 0.4060 0.2660 0.4380 0.3750 0.1560
> > 0.4220 0.3910 0.2810 0.4060 0.2650 0
> > 0.4220 0.4070 0.2500 0.3910 0.2500 0
> > 0.4060 0.4060 0.2660 0.3900 0.2340 0
> > 0.4060 0.4060 0.2650 0.4060 0.2660 0
> > 0.4220 0.4060 0.2660 0.4060 0.2650 0
> > 0.4060 0.4220 0.2500 0.3900 0.2500 0.0150
> > 0.4060 0.4060 0.2660 0.4060 0.2500 0.0160
> >
> > Here is the second process:
> >
> > 0.406 0.407 0.266 0.406 0.438 0.187
> > 0.407 0.406 0.266 0.406 0.406 0.172
> > 0.406 0.391 0.265 0.406 0.391 0.172
> > 0.407 0.39 0.25 0.422 0.391 0.157
> > 0.406 0.391 0.266 0.438 0.375 0.172
> >
> > Funny how the first process captures the graph produceed by the second
> > process in 0 secs (while the second process was running)
>
>
> This is strange. I don't know why this is.
>
>
> > But forget 3D, look at LU and FTT: same sh..t!
>
>
> Not surprising at all. Each Matlab process gets 1 CPU, and uses it fully.
>
>
> > I tried the setting OMP_NUM_THREADS to 2, nothing changes. NOTHING!
>
>
> This may not work with such an old version of Matlab (BLAS). Not sure
> when this came into effect. It definitely does work with R2006a.
>
> > Last remark - I'm running on Windows Media Center 2005.
> >
> > Any ideas how to better utilize this beast I have on my desk?
> >
> > Or I will be forced to use one core only? Maybe this is Matlab version
> > issue?
>
>
> You're stuck with one core at a time. I guess you could try running 2
> separate Matlab sessions to parallelize your work if it's possible. Other
> than that, you could upgrade to the latest Matlab and use the environment
> variable to make BLAS multi-threaded.
>
> Dan

One more time: two matlab processes running bench simultaneously:

First: bench(8):

    0.1382 0.3290 0.2525 0.5065 0.4231 0.2592
    0.1346 0.2679 0.2029 0.5103 0.3862 0.2127
    0.1331 0.2609 0.1837 0.4985 0.3791 0.2124
    0.1430 0.2691 0.1760 0.5124 0.3686 0.2200
    0.1380 0.2624 0.1775 0.5095 0.3737 0.2270
    0.1360 0.2667 0.1796 0.5191 0.3567 0.2099
    0.1348 0.2656 0.1790 0.4999 0.3389 0.2149
    0.1373 0.2655 0.1767 0.5040 0.4648 0.2129

Second: bench(5):


    0.1334 0.2821 0.1857 0.5043 0.3480 0.2519
    0.1351 0.2620 0.1913 0.5175 0.3572 0.2456
    0.1333 0.2613 0.1858 0.5014 0.3384 0.2397
    0.1367 0.2576 0.1851 0.5125 0.3329 0.2754
    0.1340 0.2664 0.1927 0.5079 0.3418 0.2526

*********************************

Single process running: bench(5)

    0.1349 0.2671 0.1764 0.5011 0.3659 0.2096
    0.1356 0.2623 0.1749 0.5078 0.3374 0.2082
    0.1366 0.2684 0.1749 0.5048 0.3481 0.2110
    0.1361 0.2642 0.1762 0.5070 0.3578 0.2075
    0.1352 0.2654 0.1755 0.5057 0.3598 0.2097

(Ignore numbers from previous post since machine there was running in
"minimal power" state)

Pleasing... but confusing...

MJ

marius.jurgilas@gmail.com wrote:
> Thanks Dan,
>
> Upgraded to 2006a (academic) + BLAS multi-threaded
>
> MAJOR difference:
>
> 0.1855 0.2775 0.1733 0.5147 0.3866 0.2277
> 0.1724 0.2626 0.1762 0.5128 0.3780 0.2226
> 0.1806 0.2630 0.1758 0.5072 0.3778 0.2266
> 0.1625 0.2599 0.1737 0.5069 0.3965 0.2270
> 0.1651 0.2577 0.1741 0.4985 0.3815 0.2153
>
> Dan Hensley wrote:
> > On Wed, 20 Sep 2006 18:30:37 -0700, marius.jurgilas wrote:
> >
> > > Dear all,
> > >
> > > I also bought a new Dell inspiron 9400 Laptop (Intel Core 2 Duo
> > > T7200[2GHz] with 2GB RAM)
> > >
> > > Primary reason for this purchase was to spead up my research. I do not
> > > see a major increase in computational power (Compared to Pentium 4
> > > 2.4GHz)
> > >
> > > PLEASE NOTE that I run Matlab version 6.5 !!!!
> > >
> > > Correct me if I'm wrong but Matlab(at least 6.5) appears to be purely
> > > one threaded program. How do I know that? Well, I run bench(5) on
> > > Matlab:
> >
> >
> > Yes, Matlab was and still is single threaded. This is well known.
> >
> >
> > > 0.391 0.406 0.266 0.406 0.39 0.172
> > > 0.406 0.406 0.25 0.406 0.39 0.172
> > > 0.406 0.391 0.25 0.407 0.391 0.172
> > > 0.406 0.406 0.266 0.406 0.39 0.172
> > > 0.406 0.406 0.25 0.391 0.375 0.172
> > >
> > > Now comes the funny part... I start to Matlab processes and let them
> > > run bench at the same time (concurently) !!!
> > > Start bench(8) on the first proces than start bench(5) on the other.
> > > Running concurently:
> > >
> > > Here is the first process:
> > >
> > > 0.4060 0.4060 0.2660 0.4380 0.3750 0.1560
> > > 0.4220 0.3910 0.2810 0.4060 0.2650 0
> > > 0.4220 0.4070 0.2500 0.3910 0.2500 0
> > > 0.4060 0.4060 0.2660 0.3900 0.2340 0
> > > 0.4060 0.4060 0.2650 0.4060 0.2660 0
> > > 0.4220 0.4060 0.2660 0.4060 0.2650 0
> > > 0.4060 0.4220 0.2500 0.3900 0.2500 0.0150
> > > 0.4060 0.4060 0.2660 0.4060 0.2500 0.0160
> > >
> > > Here is the second process:
> > >
> > > 0.406 0.407 0.266 0.406 0.438 0.187
> > > 0.407 0.406 0.266 0.406 0.406 0.172
> > > 0.406 0.391 0.265 0.406 0.391 0.172
> > > 0.407 0.39 0.25 0.422 0.391 0.157
> > > 0.406 0.391 0.266 0.438 0.375 0.172
> > >
> > > Funny how the first process captures the graph produceed by the second
> > > process in 0 secs (while the second process was running)
> >
> >
> > This is strange. I don't know why this is.
> >
> >
> > > But forget 3D, look at LU and FTT: same sh..t!
> >
> >
> > Not surprising at all. Each Matlab process gets 1 CPU, and uses it fully.
> >
> >
> > > I tried the setting OMP_NUM_THREADS to 2, nothing changes. NOTHING!
> >
> >
> > This may not work with such an old version of Matlab (BLAS). Not sure
> > when this came into effect. It definitely does work with R2006a.
> >
> > > Last remark - I'm running on Windows Media Center 2005.
> > >
> > > Any ideas how to better utilize this beast I have on my desk?
> > >
> > > Or I will be forced to use one core only? Maybe this is Matlab version
> > > issue?
> >
> >
> > You're stuck with one core at a time. I guess you could try running 2
> > separate Matlab sessions to parallelize your work if it's possible. Other
> > than that, you could upgrade to the latest Matlab and use the environment
> > variable to make BLAS multi-threaded.
> >
> > Dan