Mac Seti Enhanced Optimized

...
I also wish that I could get the performance increase that crunch3r got out of his!

...
I also wish that I could get the performance increase that crunch3r got out of his!

If you can get your hand on the intel compiler and the ipp libraries for OSX on intel macs... That'll surely boost the performance on intel macs.

Regarding the G4/G5 macs... well i know that alex kan is looking at it but i wouldn't expect much speed improvements.

P.S.

Did you compile fftw with fma enabled ? at least g5 are capable of fma and it should give a little performance increase.

I compiled fftw with altivec and fma enabled for both the g4 and g5, along with other options (took lots of research and trial and error).

And on the intel Macs, I wish I could afford to get one right now, might be a bit before I can get the $$ up to get one of the intel iMacs, or mini for that matter.

IMHO

selling the G4 and G5 mac as fast as you can is the best you can do atm to get the most money out of them.

PPC apps are dead and won't be able to compete in any further applications.
(remember the change from 68k to ppc ? But this time the transition will go even faster while apple allready developed darin x86 along to the ppc version)

Just think of what will happen if adobe decides to compile photoshop with the intel compiler...

Maybee only a g5 quad is capable of compete with a dual core intel.

Maybe I could name my computer "Charity" and take donations to get an intel mini and the intel compiler? lol

Thanks Alex!

Incredible stuff. For me, gains run somewhere between 50 and 60 percent (this is on a G4 MacMini) - all I can say is "Wow".

Great work on the optimization, I'm not getting any client errors or crashes so far. However, the fan's on way more than with the default client ;D That can only mean it's doing more real computation as opposed to wasting clocks.
____________
Donate to SETI@Home via PayPal!

Optimized SETI@Home apps + Information

Nice performance from my puny 1.42 GHz G4...one of the last results included a 3.73 GHz P4 EE..

3.73 GHz P4 EE Crunch time:
19,928.30 seconds

1.42 GHz MacMini Crunch time:
20.796.55 seconds

I can only imagine what a nice dual G5 can do. Of course, that P4EE can do up to 4 units at once...

PPC is dead - long live PPC! ;D
____________
Donate to SETI@Home via PayPal!

Optimized SETI@Home apps + Information

FYI.

Alex posted a new G4-specifc compile of his v5 worker at:

http://forums.macnn.com/showthread.php?t=295708

It will make a G4 sizzle!
On a 1.33 G4, v5-G4 tests @ 96%!!! faster than the standard worker against a reference WU.

http://tbp.berkeley.edu/~alexkan/seti/seti_enhanced-ppc-v5-g4.zip

1) Suspend Seti-Enhanced project
2) Close Boinc
3) Navigate to system Library\\Application Support\\BOINC Data\\projects\\setiathome.berkeley.edu
4) Trash (or save in folder elsewhere) stock seti worker & app_info.xml
5) unzip file & drag the seti_enhanced-ppc-v5 worker and new app_info.xml into folder
6) restart BOINC, resume project

You do not have to reduce cache or wait until the WU completes before install. You can install w/ a WU already in process and the new worker will pick-up from the point the old worker stopped.
You can also verify that it recognized and is crunching w/ the new worker by looking at the messages tab of BOINC Mgr. and looking for this: "Found app_info.xml; using anonymous platform"
In the WU result, the stderr out will show the worker version as well as the other customary info.

http://setiathome.berkeley.edu/result.php?resultid=338837248

If you like it, please be sure to drop a line and let Alex know. He's workin' hard to keep our PPCs competitive.
Regards.


____________

Well,

the G5 version already worked wonders, even on my G4. Now if this version can do even more remains to be seen :)
Will post more stats as results finish.

In any case, really good work, thanks again Alex!
____________
Donate to SETI@Home via PayPal!

Optimized SETI@Home apps + Information

Should be @ 5%-7% faster than the G5 compile on G4.
I've benched each revision against a Ref WU on my G4 going back to the stock worker.
Take a look:

AlexKan v5-G5 compile (Test4 6.6.06)
real 276m12.600s ( =16572.600s)
user 275m38.475s
sys 0m32.771s
wu_cpu_time = ????

AlexKan v5-G4 compile (Test5 6.9.06)
real 258m57.563s (= 15537.563s)
user 258m24.204s
sys 0m31.797s
wu_cpu_time = 15298.139857

& stock worker to compare:

Stock 5.13 Worker (Test1 5.24.06)
real 499m12.054s ( = 29952.054s)
user 498m11.342s
sys 0m57.878s
wu_cpu_time = 29915.352618

____________

FYI.

Alex posted a new G4-specifc compile of his v5 worker at:

http://forums.macnn.com/showthread.php?t=295708

It will make a G4 sizzle!
On a 1.33 G4, v5-G4 tests @ 96%!!! faster than the standard worker against a reference WU.

http://tbp.berkeley.edu/~alexkan/seti/seti_enhanced-ppc-v5-g4.zip

1) Suspend Seti-Enhanced project
2) Close Boinc
3) Navigate to system Library\\Application Support\\BOINC Data\\projects\\setiathome.berkeley.edu
4) Trash (or save in folder elsewhere) stock seti worker & app_info.xml
5) unzip file & drag the seti_enhanced-ppc-v5 worker and new app_info.xml into folder
6) restart BOINC, resume project

You do not have to reduce cache or wait until the WU completes before install. You can install w/ a WU already in process and the new worker will pick-up from the point the old worker stopped.
You can also verify that it recognized and is crunching w/ the new worker by looking at the messages tab of BOINC Mgr. and looking for this: "Found app_info.xml; using anonymous platform"
In the WU result, the stderr out will show the worker version as well as the other customary info.

http://setiathome.berkeley.edu/result.php?resultid=338837248

If you like it, please be sure to drop a line and let Alex know. He's workin' hard to keep our PPCs competitive.
Regards.

I just installed it on my [edit] imac [/edit] PPC G3 350 and it seems to be working fine so far. I'll keep you guys posted as to successful completion and times. Current estimated time for the work unit is 41 hours.

I just installed it on my [edit] imac [/edit] PPC G3 350 and it seems to be working fine so far. I'll keep you guys posted as to successful completion and times. Current estimated time for the work unit is 41 hours.

I haven't released any compiles for G3, so you should revert to the stock client. In fact, I can guarantee that you'll run into problems as soon as you hit the parts of the code that use Altivec, which are all over the place.

Thank you all for your messages of support, although I recognize a bunch of you from the MacNN forums... :)

So far, it seems that for results with the same credit claim, the G4 compile is really around 8-10% faster than the G5 one (on my G4 MacMini). Great work Alex!

Using your optimized client results in claimed/granted credit being perfectly the same. No computation or client errors with either G5 or G4 client, either (plus, they all validate). Very happy with it so far.
____________
Donate to SETI@Home via PayPal!

Optimized SETI@Home apps + Information

Hi Alex,

Let me know when you've got your v5 code tarball on your web site. I'd like to see what you've done to see if any of it can be folded back into the main client...

Eric
____________

Let me know when you've got your v5 code tarball on your web site. I'd like to see what you've done to see if any of it can be folded back into the main client...

OK, can anyone walk me through running fft_test3 (wisdom) in single user mode? I know how to do it in Terminal, I'm just looking for the most accurate wisdom file I can get.

G4, running 10.3.9 if it makes a difference.

Thanks Gecko, just what I needed. I'm starting to get a tiny bit famaliar with the CLI, but I'm still a LONG way from knowing what the fsck I am doing.

Didn't Alex just post a corrected/updated fft_test3 yesterday?

(still need to join the MacNN forums so I can post there)

Bump!
Needs to be on the front page, it's just too good to hide.

Regards Alex,
Simon.
____________
Donate to SETI@Home via PayPal!

Optimized SETI@Home apps + Information

Needs to be on the front page, it's just too good to hide.

@Alex Kan: Could you post links to your current G4 & G5 versions in the New Optimized Apps thread? That way they’ll be easier for people to find. Or do you still consider them to be ‘in beta’? (FWIW they’re working just fine so far on three G4s here, all running OS 10.3.9.)

I was wondering why you hadn't done that yet, Alex, since the other posts there were just OT anyway.

Your client is the first new *released* optimized client, so it should really be there.

"Alex Kan" wrote:

I wish there was something like your MKL libraries for FFTs on PowerPC, though. FFTW is much faster than vDSP (Apple's performance library) at the FFT sizes that matter to us, so it's really my only option. Writing a faster FFT is well outside my area of expertise. :/

Have you considered getting in touch with the FFTW folks and asking them about further optimizations to FFTW?

At least one of the original authors is now a Prof at MIT and the =all= are "performance hot-rodders". They want FFTW to be the fastest FFT available on any HW and under any OS.

Given your facility with assembly hacking and your experience trying to make the s@hE app faster, talking to the FFTW folks could be very fruitful.

Perhaps this might interest Akos Fekete? He appears to thrive on interesting challenges, though he does have a full plate these days w/ Einstein's S5 rollover. Still, he's been very accessible on the Einstein forum.

Have you considered getting in touch with the FFTW folks and asking them about further optimizations to FFTW?

At least one of the original authors is now a Prof at MIT and the =all= are "performance hot-rodders". They want FFTW to be the fastest FFT available on any HW and under any OS.

Given your facility with assembly hacking and your experience trying to make the s@hE app faster, talking to the FFTW folks could be very fruitful.

Just installed v6 and watching it crunch :o) Too soon for results, but we shall see.
Hats off to you, Alex!

Regards,
Simon
____________
Donate to SETI@Home via PayPal!

Optimized SETI@Home apps + Information

If you don't use graphics (or can live without them), go to the optimized apps links thread and snag the newest version for yet another speedup! Thanks to Simon Zadra of KWSN for pointing out that SETI runs slower when compiled with graphics, regardless of whether or not they're actually used.

You sure are welcome, Alex.

I've installed the graphics-less version on my Mac Mini (which has been running all your builds up to now). It's a G4 1.42 GHz/1GB model.

Thanks again :o)

Regards,
Simon.
____________
Donate to SETI@Home via PayPal!

Optimized SETI@Home apps + Information

If you don't use graphics (or can live without them), go to the optimized apps links thread and snag the newest version for yet another speedup! Thanks to Simon Zadra of KWSN for pointing out that SETI runs slower when compiled with graphics, regardless of whether or not they're actually used.

If you don't use graphics (or can live without them), go to the optimized apps links thread and snag the newest version for yet another speedup! Thanks to Simon Zadra of KWSN for pointing out that SETI runs slower when compiled with graphics, regardless of whether or not they're actually used.

Hi Alex,

my MacMini G4 1.42 GHz is slowly levelling out at 200 RAC.

Amazing work. It's up to 3x quicker than a 3.0 GHz P4 using an unoptimized client, and only about 25-35% slower than a Pentium-D 2.66 GHz using my optimized client (although that machine is running single channel RAM until tomorrow). It's also consistently pulling about 25% more RAC than an Athlon XP 2400+ (2 GHz) running an SSE optimized client.

You've done amazing things :o) I believe the speedup vs. the standard Mac app is larger than with the PC ones.

Is there data how much RAC a G5 can pull per core/total?

Regards,
Simon.
____________
Donate to SETI@Home via PayPal!

Optimized SETI@Home apps + Information

You've done amazing things :o) I believe the speedup vs. the standard Mac app is larger than with the PC ones.

Is there data how much RAC a G5 can pull per core/total?

Niiiiice numbers.

About what I expected, around 7000 seconds for a general 60 cobblestone WU and about 7500-8000 for a VLAR. I still have to test the P-D with dual channel, and I do expect a sizable speedup too.

Still, my first test running a linux cruncher in a VM on Windows has yielded by far the quickest runs for me so far - I'll be testing that hypothesis (that a Linux cruncher in a VM is quicker than either a native Win32 cruncher or a native Linux one) further on the P-D as well as my Athlon64 3500+.

There, I was in the 6800-7300 for 60-63 credit units, as well (with the A64). Nowhere close to it on Windows, haven't tried native Linux on that Machine yet.

And yeah, that Woodcrest is out of this world.
However, this monday Intel started selling them! So for anyone who's interested, you can grab one now. Just be prepared to look around for suitable platforms, not too many right now.

Their prices (same as Core 2 Duo) will stay stable for the remainder of 2006, so you might as well grab one now ;o)
____________
Donate to SETI@Home via PayPal!

Optimized SETI@Home apps + Information

I've posted the source code to v6. Sorry about the delay, if anyone was looking to get their hands on it.

Tetsuji, this code (and every release since v3) contains a working transposed-PoT implementation. I looked through a recent tarball and didn't see anything obviously wrong with the parts using the transposed PoT matrix, but I haven't checked to make sure that the matrix transposes work as advertised. Correctness ought to be easy to verify, since you should be able to compare the PoT arrays generated with and without the transposed code. However, I'll be out of the country for a few weeks and won't be able to put in fixes myself.

Before I leave for a while and forget about them, here's those benchmark numbers on a Dual 1.8 GHz G5 from a while back. (I had to manually adjust the chirp limits to 1/2.5, since IIRC in the package I downloaded, they were set to 4/10.)

testWU-1
stock 5.13 (-nographics): 21:34.26 (1294.26s)
v6 (-nographics): 7:46.93 (466.93s) - 63.9% time reduction

testWU-2
stock 5.13 (-nographics): 18:06.75 (1086.75s)
v6 (-nographics): 8:58.35 (538.35s) - 50.5% time reduction

testWU-4
stock 5.13 (-nographics): 7:33.24 (453.24s)
v6 (-nographics): 3:20.61 (200.61s) - 55.7% time reduction

I was going to add this as an edit, but an hour has already gone by since my last post. :P

Reviving an old thread to post some news...

An initial version of my optimized clients for Intel Macs is almost done. I currently don't have access to the Intel Compiler or IPP libraries, so you'll have to settle for what I can do without touching the FFTs. This means that your performance on those shorter (12-13 credit) WUs might not stack up against Windows/Linux clients. Longer WUs should be just fine. ;)

Keep your eyes peeled for an update to the optimized apps thread!