| Re: IDL 6.3 for Mac OS X on Intel Now Available [message #49304] |
Fri, 14 July 2006 01:00 |
Paolo Grigis
Messages: 171
Registered: December 2003
|
Senior Member |
|
|
I think it would be very interesting to compare performance
with a macbook with the same specification running windows
via bootcamp... will this give a further increase in
performance due to windows optimizations? Or at least
somebody with a native windows core duo laptop could
report.
Otherwise I cannot understand why a windows laptop with
a pentium M clocked at 1.6 GHz performs 1.27 in TT3...
or are core duo processors *expected* to be less
efficient than similarly clocked pentium Ms for single
processor task?
(For comparison, on that system JD_TEST returns 0.73,
a bit more than a factor 2 worse than dual core, as
expected).
Ciao,
Paolo
JD Smith wrote:
> On Thu, 13 Jul 2006 09:24:56 -0700, bokubo wrote:
>
>
>> I am pleased to announce the release of IDL 6.3 for Mac OS X on Intel.
>> This new IDL release runs as a native application on all Mac Intel
>> supported machines and offers significant performance benefits. We have
>> seen a growing popularity of Mac OS X for scientific and analysis
>> applications, and this release represents our ongoing commitment to
>> this growing base of IDL users.
>
>
> And....
>
> it's *FAST*. Here's the breakdown for some portable OSX systems,
> following along the tests at:
>
> http://idl.tamu.edu/mac_bench.php
>
> System TT3 (AVG) TT3 (GEOM) JD_TEST
> ============================================================ =================
> PB (G4 1.67GHz, 2GB, PPC IDL native) 3.20 .13 1.86
> MBP(CoreDuo 2GHz, 1GB, PPC IDL via Rosetta) 3.45 .13 3.02
> MBP(CoreDuo 2GHz, 1GB, i386 IDL native) 1.69 .06 0.32
>
> All times in seconds.
>
> PB == PowerBook
> MBP == MacBook Pro (thanks to Jason Harris for a temporary loan)
> TT3 == Time Test 3, run under IDL 6.3, demo mode.
> AVG == average
> GEOM == geometric mean
> JD_TEST ==
> IDL> a=randomu(sd,100L*!CPU.TPOOL_MIN_ELTS)
> IDL> t=systime(1) & a=sqrt(a)/(a>0.5) & print,systime(1)-t
>
> So, for things which are limited by a single processor (TIME_TEST3),
> we're roughly ~2x faster than a G4 PB (which was a slow IDL system, to
> be fair), with only 20% more clock speed.
>
> But, the real fun comes when running big array manipulations, like
> JD_TEST, where the dual Core Duo processors can flex their muscles.
> Here the speedup is closer to 6x, which is almost too good to believe.
> This will of course depend on which operations you use, but testing a
> variety of arithmetic ones, I found speedups of anywhere from 2-7 on
> arrays large enough to make multi-threading effective.
>
> The MBP (a portable system) is now comparable in speed to a quad-processor
> G5, and (I suspect) similar dual-processor Linux/Windows desktops. Very
> respectable. Woeful OSX/IDL performance, R.I.P.
>
> JD
>
|
|
|
|
| Re: IDL 6.3 for Mac OS X on Intel Now Available [message #49311 is a reply to message #49304] |
Thu, 13 July 2006 13:10  |
FL
Messages: 17
Registered: April 2006
|
Junior Member |
|
|
Hi,
FL can be started as 'fl --gui', which will present a very primitive
interface, one window for standard input, and another for standard output
(this is the default on Windows). However, the event loop is running in
this case, and the results for TT3 and JD_TEST are the same.
I use gcc 3.4.x and gcc 4.1.x for developing FL, I think ITTVIS/RSI uses
the same compiler (on linux, Intel compiler on Windows). I have written
MMX/SSE/SSE2 array arithmetics routines in inline assembly, this can be
accounted for some speedup. I have written my own memory allocators (more
than one, different allocators for different usage patterns.) But the real
difference comes from overall design, from the very first day performance
was my main concern. Implementation is very important. Take eg. matrix
multiplication in IDL and Matlab: for 1000x1000 matrices, Matlab is 6-7
times faster, although both are using the same BLAS API. The difference is
in the implementation: IDL uses a simple C BLAS, while Matlab uses ATLAS
(http://math-atlas.sf.net).
My plans:
- very short term: some more vacation :-)
- short term: widgets have absolute priority. Some basic functionality
already work, usable widgets will be available, let's say, by end of
September.
- long term (1-2 years):
DFL (Distributed FL): similar to the IDL-IDL bridge, but slave processes
can be invoked on different machines, communicating through TCP/IP
(FL's of the world, unite! :-)
Matlab front-end: I guess 90% of the present code can be reused in a
Matlab-like interpreter. So, why not create a 2-in-1 monster?
The bad news is that FL will remain a one man show. I used to work in
small groups earlier, but I have very bad memories. Countless hours were
spent for meetings and communication, instead of productive work.
regards,
lajos
On Thu, 13 Jul 2006, JD Smith wrote:
> Impressive. How much of FL's performance gain comes from compiler
> optimizations? I am suspect of speedups in operations limited by
> looping, since FL and GDL don't yet implement widgets or other even
> processing which slows IDL's loops down. However, just taking the
> square root of a bunch of numbers.... that's a different question.
>
> What are the plans for FL? Any hope of combining the two efforts
> (FL/GDL) into one open source project?
>
> JD
>
|
|
|
|
| Re: IDL 6.3 for Mac OS X on Intel Now Available [message #49312 is a reply to message #49311] |
Thu, 13 July 2006 13:03 |
Karl Schultz
Messages: 341
Registered: October 1999
|
Senior Member |
|
|
On Thu, 13 Jul 2006 11:15:15 -0700, JD Smith wrote:
> On Thu, 13 Jul 2006 09:24:56 -0700, bokubo wrote:
>
>> [quoted text muted]
>
> And....
>
> it's *FAST*. Here's the breakdown for some portable OSX systems,
> following along the tests at:
>
> http://idl.tamu.edu/mac_bench.php
>
> System TT3 (AVG) TT3 (GEOM) JD_TEST
> ============================================================ =================
> PB (G4 1.67GHz, 2GB, PPC IDL native) 3.20 .13 1.86
> MBP(CoreDuo 2GHz, 1GB, PPC IDL via Rosetta) 3.45 .13 3.02
> MBP(CoreDuo 2GHz, 1GB, i386 IDL native) 1.69 .06 0.32
>
> All times in seconds.
>
> PB == PowerBook
> MBP == MacBook Pro (thanks to Jason Harris for a temporary loan)
> TT3 == Time Test 3, run under IDL 6.3, demo mode.
> AVG == average
> GEOM == geometric mean
> JD_TEST ==
> IDL> a=randomu(sd,100L*!CPU.TPOOL_MIN_ELTS)
> IDL> t=systime(1) & a=sqrt(a)/(a>0.5) & print,systime(1)-t
>
A few more datapoints (all for TT3 (AVG)):
iMac - i386 CoreDuo 2.0 GHz : 1.2 (ppc binaries with Rosetta: 2.3)
Pentium 4 (Windows) 2.8 GHz : 1.5
G4 ppc 0.7 GHz : 5.6
G5 dual ppc 2.0 GHz : 1.9
Xeon dual 3.4 GHz Linux : 1.1
Karl
|
|
|
|
| Re: IDL 6.3 for Mac OS X on Intel Now Available [message #49315 is a reply to message #49312] |
Thu, 13 July 2006 12:09 |
JD Smith
Messages: 850
Registered: December 1999
|
Senior Member |
|
|
On Thu, 13 Jul 2006 21:00:44 +0200, F�LDY Lajos wrote:
> Hi,
>
> a little FL promotion :-)
>
>
> On Thu, 13 Jul 2006, JD Smith wrote:
>
>> it's *FAST*. Here's the breakdown for some portable OSX systems,
>> following along the tests at:
>>
>> http://idl.tamu.edu/mac_bench.php
>>
>> System TT3 (AVG) TT3 (GEOM) JD_TEST
>> ============================================================ =================
>> PB (G4 1.67GHz, 2GB, PPC IDL native) 3.20 .13 1.86
>> MBP(CoreDuo 2GHz, 1GB, PPC IDL via Rosetta) 3.45 .13 3.02
>> MBP(CoreDuo 2GHz, 1GB, i386 IDL native) 1.69 .06 0.32
>
> FL (64 bit linux) on a dual 1.8GHz Opteron 0.57 .02 0.13
>
> not bad :-)
>
> for FL, TT3 includes assoc test (0.02 s), and running JD_TEST as
>
> a=randomu(sd,100L*100000l)
> t=systime(1) & a=sqrt(a)/(a>0.5) & print,systime(1)-t
>
> (!CPU.TPOOL_MIN_ELTS is 0 by default in FL, because FL uses different
> limits for different operations (eg 100000 for addition, 25000 for sin))
>
> and another good news: the new Intel Core 2 processors will have more
> SSE/SSE2 execution units, so FL will be even faster.
Impressive. How much of FL's performance gain comes from compiler
optimizations? I am suspect of speedups in operations limited by
looping, since FL and GDL don't yet implement widgets or other even
processing which slows IDL's loops down. However, just taking the
square root of a bunch of numbers.... that's a different question.
What are the plans for FL? Any hope of combining the two efforts
(FL/GDL) into one open source project?
JD
|
|
|
|
| Re: IDL 6.3 for Mac OS X on Intel Now Available [message #49316 is a reply to message #49315] |
Thu, 13 July 2006 12:00 |
Foldy Lajos
Messages: 268
Registered: October 2001
|
Senior Member |
|
|
Hi,
a little FL promotion :-)
On Thu, 13 Jul 2006, JD Smith wrote:
> it's *FAST*. Here's the breakdown for some portable OSX systems,
> following along the tests at:
>
> http://idl.tamu.edu/mac_bench.php
>
> System TT3 (AVG) TT3 (GEOM) JD_TEST
> ============================================================ =================
> PB (G4 1.67GHz, 2GB, PPC IDL native) 3.20 .13 1.86
> MBP(CoreDuo 2GHz, 1GB, PPC IDL via Rosetta) 3.45 .13 3.02
> MBP(CoreDuo 2GHz, 1GB, i386 IDL native) 1.69 .06 0.32
FL (64 bit linux) on a dual 1.8GHz Opteron 0.57 .02 0.13
not bad :-)
for FL, TT3 includes assoc test (0.02 s), and running JD_TEST as
a=randomu(sd,100L*100000l)
t=systime(1) & a=sqrt(a)/(a>0.5) & print,systime(1)-t
(!CPU.TPOOL_MIN_ELTS is 0 by default in FL, because FL uses different
limits for different operations (eg 100000 for addition, 25000 for sin))
and another good news: the new Intel Core 2 processors will have more
SSE/SSE2 execution units, so FL will be even faster.
regards,
lajos
> All times in seconds.
>
> PB == PowerBook
> MBP == MacBook Pro (thanks to Jason Harris for a temporary loan)
> TT3 == Time Test 3, run under IDL 6.3, demo mode.
> AVG == average
> GEOM == geometric mean
> JD_TEST ==
> IDL> a=randomu(sd,100L*!CPU.TPOOL_MIN_ELTS)
> IDL> t=systime(1) & a=sqrt(a)/(a>0.5) & print,systime(1)-t
>
> So, for things which are limited by a single processor (TIME_TEST3),
> we're roughly ~2x faster than a G4 PB (which was a slow IDL system, to
> be fair), with only 20% more clock speed.
>
> But, the real fun comes when running big array manipulations, like
> JD_TEST, where the dual Core Duo processors can flex their muscles.
> Here the speedup is closer to 6x, which is almost too good to believe.
> This will of course depend on which operations you use, but testing a
> variety of arithmetic ones, I found speedups of anywhere from 2-7 on
> arrays large enough to make multi-threading effective.
>
> The MBP (a portable system) is now comparable in speed to a quad-processor
> G5, and (I suspect) similar dual-processor Linux/Windows desktops. Very
> respectable. Woeful OSX/IDL performance, R.I.P.
>
> JD
>
>
|
|
|
|
| Re: IDL 6.3 for Mac OS X on Intel Now Available [message #49319 is a reply to message #49316] |
Thu, 13 July 2006 11:15 |
JD Smith
Messages: 850
Registered: December 1999
|
Senior Member |
|
|
On Thu, 13 Jul 2006 09:24:56 -0700, bokubo wrote:
> I am pleased to announce the release of IDL 6.3 for Mac OS X on Intel.
> This new IDL release runs as a native application on all Mac Intel
> supported machines and offers significant performance benefits. We have
> seen a growing popularity of Mac OS X for scientific and analysis
> applications, and this release represents our ongoing commitment to
> this growing base of IDL users.
And....
it's *FAST*. Here's the breakdown for some portable OSX systems,
following along the tests at:
http://idl.tamu.edu/mac_bench.php
System TT3 (AVG) TT3 (GEOM) JD_TEST
============================================================ =================
PB (G4 1.67GHz, 2GB, PPC IDL native) 3.20 .13 1.86
MBP(CoreDuo 2GHz, 1GB, PPC IDL via Rosetta) 3.45 .13 3.02
MBP(CoreDuo 2GHz, 1GB, i386 IDL native) 1.69 .06 0.32
All times in seconds.
PB == PowerBook
MBP == MacBook Pro (thanks to Jason Harris for a temporary loan)
TT3 == Time Test 3, run under IDL 6.3, demo mode.
AVG == average
GEOM == geometric mean
JD_TEST ==
IDL> a=randomu(sd,100L*!CPU.TPOOL_MIN_ELTS)
IDL> t=systime(1) & a=sqrt(a)/(a>0.5) & print,systime(1)-t
So, for things which are limited by a single processor (TIME_TEST3),
we're roughly ~2x faster than a G4 PB (which was a slow IDL system, to
be fair), with only 20% more clock speed.
But, the real fun comes when running big array manipulations, like
JD_TEST, where the dual Core Duo processors can flex their muscles.
Here the speedup is closer to 6x, which is almost too good to believe.
This will of course depend on which operations you use, but testing a
variety of arithmetic ones, I found speedups of anywhere from 2-7 on
arrays large enough to make multi-threading effective.
The MBP (a portable system) is now comparable in speed to a quad-processor
G5, and (I suspect) similar dual-processor Linux/Windows desktops. Very
respectable. Woeful OSX/IDL performance, R.I.P.
JD
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
