I thought it be better to have them in one place...

Linux 32bit:

SSE

SSE2

SSE3


Linux 64bit:

SSE3

SSSE3

SSE4.1

All compiled with Intel icpc from the original code.
All tested against the test wus and giving exactly the same result.

Please report any problems here.

____________
mic.

All compiled with Intel icpc from the original code.
All tested against the test wus and giving exactly the same result.

What kind of speedup did you get on these compared to the stock app? And did you use any specific compiler flags?

Due to the v11 release I removed the apps from the server.
New ones soon. :-)
____________
mic.

Stickying this because i think it's something our linux users would like. Also, the linux optimized apps seem to have been returning good results :)

I suppose the question then has to be asked.

Why are the linux opt apps giving good results, but the Win ones are causing a load of trouble? Is it a compiler flag difference or code change?

And I'm sure there are a lot of Win users (and maybe MAC users too) who would like to have similarly optimized apps (that return good results).
____________
Never surrender and never give up. In the darkest hour there is always hope.

I suppose the question then has to be asked.

Why are the linux opt apps giving good results, but the Win ones are causing a load of trouble? Is it a compiler flag difference or code change?

And I'm sure there are a lot of Win users (and maybe MAC users too) who would like to have similarly optimized apps (that return good results).

New recompiled apps for Linux64 on Intel:

SSE3_64


SSSE3_64


SSE41_64

It seems they don't run well on AMD, so it might better for AMD users to try a testfile in standalone mode --> look here

Please report errors (and success) here.
____________
mic.

New recompiled apps for Linux64 on Intel:

SSE3_64


SSSE3_64


SSE41_64

It seems they don't run well on AMD, so it might better for AMD users to try a testfile in standalone mode --> look here

Please report errors (and success) here.

Hi speedimic,

THX again for the great work!
My Linux quad likes the SSSE3 app very much... ;-)

Xubuntu 8.10 64-bit, Intel Q6600 @ 2.4 Ghz

Each WU needs ~18-19 minutes on that host and no errors so far! :-)))
____________
Lovely greetings, Cori

New recompiled apps for Linux32 on Intel:

SSE3_32


SSE2_32


SSE_32

It seems they don't run well on AMD, so it might better for AMD users to try a testfile in standalone mode --> look here

Please report errors (and success) here.
____________
mic.

As I told here, I have some problem with some wus only with SSE4.1 X86_64 app !

I don t know if it's the application or the new assimilator/validator :/

It seems they don't run well on AMD, so it might better for AMD users to try a testfile in standalone mode --> look here

How do you install these clients in Ubuntu64?

Tried replacing the stock opt app in /var/lib/boinc-client/.. but it errored out.
Where to place the app_info.xml?

Guess i'm not following the right steps?

thx

The Intel compiler runs degraded code when run on AMD processors. The highest SSE level that it runs on AMD processors is SSE2.

HTH

How do you install these clients in Ubuntu64?

Tried replacing the stock opt app in /var/lib/boinc-client/.. but it errored out.
Where to place the app_info.xml?

Guess i'm not following the right steps?

thx

The SSE3 code runs very well on AMD LE-1600, like C2D with the same clock.
I use icpc -xO for this.

And here are flags:

CXX_i686 = icpc
CXXFLAGS_i686 = -xSSE3 -O3 -ipo -no-prec-div -static -fp-model fast=2 -fp-speculation=fast -opt-calloc -unroll-aggressive -opt-multi-version-aggressive -fast-transcendentals

CXX_x86_64 = icpc
CXXFLAGS_x86_64 = -xSSE4.1 -O3 -ipo -no-prec-div -static -fp-model fast=2 -fp-speculation=fast -opt-calloc -unroll-aggressive -opt-multi-version-aggressive -fast-transcendentals

And here are flags:

CXX_i686 = icpc
CXXFLAGS_i686 = -xSSE3 -O3 -ipo -no-prec-div -static -fp-model fast=2 -fp-speculation=fast -opt-calloc -unroll-aggressive -opt-multi-version-aggressive -fast-transcendentals

CXX_x86_64 = icpc
CXXFLAGS_x86_64 = -xSSE4.1 -O3 -ipo -no-prec-div -static -fp-model fast=2 -fp-speculation=fast -opt-calloc -unroll-aggressive -opt-multi-version-aggressive -fast-transcendentals

You can omit "fast-transcendentals" as this is the default when specifying "-fp-model fast" (or even fast=2).
I don't use -unroll-aggressive and -opt-multi-version-aggressive, does it help the performance? I would think that it doesn't bring much to the table.

The SSE3 code runs very well on AMD LE-1600, like C2D with the same clock.
I use icpc -xO for this.

At run-time the processor is probed and if it's by AMD, then degraded code is run instead of the SSE3 code.

See http://techreport.com/discussions.x/8547 for a snippet.

HTH

1. This article is very old: 11:58 AM on July 13, 2005
2. I've got a lot better performance with SSE2 (20% boost) than without it, and slightly better performance with SSE3 than SSE2 (another 1% boost) and I'm talking about AMD chip and milkyway app of course.

1 - It's been known in the open source community and Intel's response was that they cannot guarantee their compiler except on their processors, fair enough. Is this new enough for you?

2 - 1% is too close to noise to call a boost.

2. But I've made more tests:

averaged boost in calculation times for 126 runs of milkyway app on idle machine
SSE3 app: 121.86%
SSE2 app: 119.03%
base app: 100.00%

I don't think this is 'noise' only...
I'm confused now...
Maybe this is milkyway specific...

Travis, please take a look at this host, everything coming in after 20:45 UTC is done with the new recompiled v14.
____________
mic.

Travis, please take a look at this host, everything coming in after 20:45 UTC is done with the new recompiled v14.

I'll let you know as soon as I get some more results from it. But it should be OK if it was returning the same results for the test workunits.

SSE3 might be doing some other optimizations (or have some changes in optimizations) which are better than what was in SSE2, because it's newer.

Looks to me like it's generating good results, so I'd go ahead and release it.

I'll take that as a yes :-) I just downloaded the SSE version. Unfortunately only my 1.3 GHz Celerons can take advantage of these apps (much to their pleasure I might add). The .12 version saw an increase from 2:02 to 1:37 CPU time as compared to the stock version.

My AMD 3800+ x2 and AMD 5600+ x2 can't handle them (as expected). I only tried the 64 bit SSE3 app tho. Maybe the 32 bit SSE2 will work? It's much too late tonight to embark on what may be a major task. I'll try it tomorrow unless someone knows it's futile.

The .12 version saw an increase from 2:02 to 1:37 CPU time as compared to the stock version.

Travis, please take a look at this host, everything coming in now is done with the new recompiled v15.

____________
mic.

You people obviously put a lot of hard work in and I thank you for that, but is there a newbies guide to installing these? I have just started playing with linux and have 2 quad cores that I'd love to try these out on.

I have tried searching but to no avail.

Thanks again,

Neal

It looked good until maybe the couple workunits which were bad... However the stock app is STILL doing the same thing ;( I have no clue what's up.

You people obviously put a lot of hard work in and I thank you for that, but is there a newbies guide to installing these? I have just started playing with linux and have 2 quad cores that I'd love to try these out on.

I have tried searching but to no avail.

Thanks again,

Neal

Travis, please take a look at this host, everything coming in now is done with the new recompiled v16.

Feels like "Groundhog Day"... :)
____________
mic.

I'll post the the new apps whe I get back from work, around 17.00 UTC.
...hope that's soon enough to catch the stricter error checking.

____________
mic.

For the AMD users I got two new apps to try:

AMD SSE3_64

For the AMD users I got two new apps to try:

AMD SSE3_64

Just wanted to try this one out, but it gave me a "Not found"-message...

Looks like they're doing well. Haven't had a bad result come back with any compiled v0.16 app yet :)

AMD SSE3_64

Hi
thanks a lot for these linux apps, they are running fine. I have one older AMD SSE-capable and on that I am now trying the v16 "Linux on Intel" SSE, and that seems to be working as well
____________

Since x86-64 guarantees that at least SSE2 is available, did you make sure to enable vectorization through GCC's -ftree-vectorize option (implied by -O3 in versions 4.3 and later)? For that matter, any SSE build could benefit from vectorization.

HTH


____________

Since x86-64 guarantees that at least SSE2 is available, did you make sure to enable vectorization through GCC's -ftree-vectorize option (implied by -O3 in versions 4.3 and later)? For that matter, any SSE build could benefit from vectorization.

HTH

...
Unfortunately it takes about 24 minutes, which is about 2 1/2 times longer than the Windows one.

Gives a good result anyway, but it's not worth the 64-bit and so gives less credits.

Here's apples to apples. SSE_32 and AMD SSE3_64 are both winners. First of all they run! Here's what I've seen so far (remember sample size is 1 for the new app).

SSE_32: Running on two 1.3 GHz Celeron computers running SUSE 10.3. The stock app took about 2:02 hours. The optimized app (both .14 and .16) take about about 1:35 hours.

AMD SSE3_64: My AMD64 5600+X2 running SUSE 10.3 64 bit ran the optimized .07 (SSE2) app in 0:31 hours. The stock apps from 0.12 to 0.16 ran in about 0:53 hours. The new optimized app ran in 0:23 hours.

AMD SSE3_64: My AMD64 3800+X2 OC'd to 2.35 GHz running SUSE 11.1 64 bit ran the 0.16 stock app in 1:01 hours. The new optimized app ran in 0:28 hours.

Thanks Speedimic for the optimized apps. They are working VERY well, but what will you do for an encore? :-)

Rick

Many thanks from me too!

I had the .14 SSE2 running on my Core Duo (Fedora 9). Times went down for about an hour and a half to just under an hour. I found out that PNI = SSE3 so I'm trying the .16 SSE3 version now.

Running well on my Core 2 Duo (Fedora 7) as well. I'm getting ready to upgrade to the .16 release.
____________
Kathryn :o)
The BOINC FAQ Service
The Unofficial BOINC Wiki
The Trac System
More BOINC information than you can shake a stick of RAM at.

How long is the stock app running?

If it gives less credit, you might be running into the 108 credit barrier.

What do mean by 'it's not worth the 64bit'?

Here are some results from this host:

• 0.16: 82min
  
• 0.16: 84min
  
• 0.16: 82min
  
• 0.16: 81min
  
• 0.16: 84min
  
• 0.16: 81min
  
• 0.15: 83min
  

I didn't record the times for 0.14, but IIRC they were roughly in the same ballpark, i.e., no noticeable improvement between versions.

Note that this host is a production server which idles most of the time, therefore Linux power-management runs it at the slowest speed of 1GHz most of the time.
____________

Wow, that's fast!!!
My Linux quad (2.4Ghz) needs at least around 18 minutes with the opti SSSE3 (64bit) app!
What am I doing wrong?
____________
Lovely greetings, Cori

I don't know how Crunch3r managed that but as he made his optimized app way earlier he made also a Linux one. Compared to his Windows app (5-6 minutes on my X2) it took then 4-5 minutes...

What do mean by 'it's not worth the 64bit'?

Was no intended harm, I meant it that way that a 64-Bit app normally should be faster than a 32-Bit app.
And that is not the case. ;-)...

Normally the 64bit (stock) apps are compiled with SSE2 enabled, because a 64bit-capable cpu is also capable of at least SSE2.

I found out that PNI = SSE3

Looks like it is indicated by sse4_1

My Xeon X3350 cat /proc/cpuinfo flags:
fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmovpat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx lm constant_tsc arch_perfmon pebs bts rep_good nopl pni monitor ds_cpl vmx smx est tm2 ssse3 cx16 xtpr sse4_1 lahf_lm

Thank you for your SSE41_64 version speedimic. It is working well on my computer @ 3.3GHz.

cat /proc/cpuinfo flags shows -
fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe nx lm constant_tsc pni monitor ds_cpl est tm2 xtpr

I know pni = SSE3 but how is SSE4.1 indicated?

Hi mic,

I've added these to zslip

I hope that's OK

Looks like they're doing well. Haven't had a bad result come back with any compiled v0.16 app yet :)

Now then, the new recompiled v16 apps for Linuxs:

Linux32 on Intel

SSE3_32
SSE2_32
SSE_32

Linux64 on Intel

SSE3_64
SSSE3_64
SSE41_64

For AMD users:

AMD SSE3_64
AMD SSE2_32

I only had the chance to test the AMD SSE2_32 on my Athlon64 3200+, so the rest of the testing is up to you... Please report!

Hi mic,

I've added these to zslip

I hope that's OK

Hi mic,

I've added these to zslip

I hope that's OK

Sure.
Always good to have everything in one place.

If someone feels like crunching full speed on linux, here are the new v18d apps.
Crunch time is cut to half compared to the prior version (on my quad at least...).

Linux32 Pack
(with new SSSE3-version)

Linux64 Pack

LinuxAMD Pack

Report problems & errors here...

____________
mic.

Thanks speedimic! Just dumped it onto my linux32 laptop.

Yeah, 32bit SSSE3 Linux dropped from 22 to 9 minutes.
64bit SSSE3 runs them in 8.5 minutes compared to 18 with the stock app.
Both on Intel Q6600.
Very nice :)

Thanks mic, I now have it running on my Fedora9_64 boxes

My laptop is now running around 22-25 minutes with the optimised Linux app.
____________

If someone feels like crunching full speed on linux, here are the new v18d apps.
Crunch time is cut to half compared to the prior version (on my quad at least...).

Linux32 Pack
(with new SSSE3-version)

Linux64 Pack

LinuxAMD Pack

Report problems & errors here...

If there's any other Linux32/64 - SSE-level combination needed, just tell me!

____________
mic.

Just got myself the shiny new version of the Intel compiler... and made new apps.

Codebase is still 18d.
Due to the varying WUs I can't really tell if those are any faster - that' up to you to decide... ;)
So I would call updating optional!

What has changed is that everything up to and including sse3 is AMD compatible, and I made all sse-levels the compiler offered me.

Linux32-SSE
Linux32-SSE2
Linux32-SSE3
Linux32-SSSE3
Linux32-SSE4.1
Linux32-SSE4.2

Linux64-SSE3
Linux64-SSSE3
Linux64-SSE4.1
Linux64-SSE4.2



____________
mic.

What has changed is that everything up to and including sse3 is AMD compatible, and I made all sse-levels the compiler offered me.

Linux32-SSE

If that's a ICC 11 version, than SSE is just a synonyme for x87 (no enhanced instruction set). It won't put any SSE instructions in the code.
The difference to the stock app is just the change gcc -> ICC. But maybe that gains some speed for older machines, too.

Just got myself the shiny new version of the Intel compiler... and made new apps.

Codebase is still 18d.
Due to the varying WUs I can't really tell if those are any faster - that' up to you to decide... ;)
So I would call updating optional!

What has changed is that everything up to and including sse3 is AMD compatible, and I made all sse-levels the compiler offered me.

What about Intel Linux 64 Bit SSE2 ? I have a couple of Q6600's on SSE2 PC's
____________

Proud Founder and member of BOINC@AUSTRALIA

Q6600 has SSE3 and SSSE3, doesn't it? SSE3 is denoted by the code pni.

PNI (Prescott New Instructions) was the original engineering code name for SSE3.

OK on one Linux PC it just said SSE2 but on the windows using CPU-Z it says SSE3 so I guess I am fine and do not need a SSE2 version. I am not up on all this stuff :)

Thanks
____________

Proud Founder and member of BOINC@AUSTRALIA

No worries, judging by your stats for many projects you're struggling along quite well.
A few suitable ATI cards in some of those boxes and you'll be in orbit. :)

In case I didn't make myself clear before, Linux will not report SSE3 but pni instead. pni=SSE3.

I don't know what CPU-Z shows, as I haven't got a working SSE3 capable computer. With perfect timing my X3350 computer with a HD3850 failed a few days before the ATI 32-bit Windows teaser application was released.

Turns out that Ubuntu (cat /proc/cpuinfo) shows it as ssse3, and the optimized ssse3 works fine, well so far anyway.
____________

Proud Founder and member of BOINC@AUSTRALIA

Yes that's fine, SSSE3 is different to SSE3 though, it has some extra instructions and an extra "S". :)
SSE3=pni and SSSE3=SSSE3.

As mentioned in an earlier post in this thread an older Linux kernel may not report the most recent codes. This may explain why cat /proc/cpuinfo will report the same CPU differently on different boxes if you are using older and newer versions of Linux on them.

Not sure of the difference in speed between SSE3 and SSSE3 versions, I've only ever tried the Linux64 SSE4.1.
The main thing is you are crunching OK.

Not sure of the difference in speed between SSE3 and SSSE3 versions...

Not sure of the difference in speed between SSE3 and SSSE3 versions...

It should be zilch, since it's unlikely that the compiler will find opportunities in MW code to fit the multimedia-like SSSE3 instructions.

HTH

Is there any plan for a linux app for ATI/GPU in the pipeline quad or i7

Yes, release ATI app for Linux, please. CUDA for all systems is under construction now, but it makes sense to make app for stronger cards first and ATI is the one. I hate running my machine under Win just to be able to run Milky on GPU. :-( And I do not think there are more guys with NVidia GTS/GTX 200 cards than Linux ones with ATI 38xx/48xx.

Still a little disappointed with the linux apps.

I have SSE3 running on an e6400@2.4 test machine ...

-rw-r--r-- 1 root root 594 2009-02-18 21:33 app_info.xml
-rw-r--r-- 1 root root 35127 2009-01-19 20:07 GPL.txt
-rwxr-xr-x 1 root root 1222624 2009-02-18 21:25 milkyway_0.18_sse3_i686-pc-linux-gnu

This yield a wu time of 35minutes

Whereas the windows op app yields a time of 20minutes on a slower processor.(1.8ghz T2390)

The difference seems a lot to me or is this typical ?



____________

After reading that Intel compiler flag "-fp-model fast=2" was being used on these applications, I figured I would run a test, and results are as expected.

The Linux 64bit SSSE3 0.18d mkII app downloaded from zslip.com gives invalid results. I suspect that the same flag was used on the other apps, too. Flag "-fp-model precise" needs to be used to get the fitness in line with expected output.

This app's output:

The Linux 64bit SSSE3 0.18d mkII app downloaded from zslip.com gives invalid results.
[..]
This app's output:
fitness: -3.027909854710229

Correct output:
fitness: -3.027909854710189

If that is the case, then why does Travis have a thread about testing custom apps, saying "The results should look like the following:" instead of "should be within a deviation of the following:"?

I'm sure others would like to know what this allowed deviation is. The result posted above has error as a result of about a 50% speed increase. That's pretty major.

He did post awhile ago that the results must be to the x place, I believe it was the 10th. Those results are equal to the 12th.
____________
Doesn't expecting the unexpected make the unexpected the expected?
If it makes sense, DON'T do it.

Still a little disappointed with the linux apps...

Whereas the windows op app yields a time of 20minutes on a slower processor.(1.8ghz T2390)

The difference seems a lot to me or is this typical ?

It might be because Linux manages power differently from Windows, running BOINC applications at a slow CPU frequency in order to save energy. See more details here.

HTH

The linux client is simply inefficient compared to win32.

The linux client is simply inefficient compared to win32.

The linux client is simply inefficient compared to win32.

I thought the fastest version of Speedimic is quite close.

Just looked it up, speedimics own Q9550 (45nm, 2.83GHz) is taking about 1030 seconds for a 27.77 credit WU under Linux.

He has another host (Q6600, 65nm, appears to be overclocked to 2.7-2.8GHz from the benchmark values) completing the same tasks in about 1150 seconds under Windows. It is a somehow bad comparison because of the different clockspeed and that the 45nm CPUs should be faster per clock than their 65nm counterparts, but nonetheless it is clear that his Linux version can't be that bad.

I thought the fastest version of Speedimic is quite close.

Just looked it up, speedimics own Q9550 (45nm, 2.83GHz) is taking about 1030 seconds for a 27.77 credit WU under Linux.

He has another host (Q6600, 65nm, appears to be overclocked to 2.7-2.8GHz from the benchmark values) completing the same tasks in about 1150 seconds under Windows. It is a somehow bad comparison because of the different clockspeed and that the 45nm CPUs should be faster per clock than their 65nm counterparts, but nonetheless it is clear that his Linux version can't be that bad.

Here are my tasks ....

Laptop win32 : http://milkyway.cs.rpi.edu/milkyway/result.php?resultid=39511370

Linux : http://milkyway.cs.rpi.edu/milkyway/result.php?resultid=39523720

Thanks for the pointers I'll recheck if I'm using the wrong client

You could use the SSSE3 version (instead of only SSE3) on your Core2. The SSE3 binary retains compatibilty to AMD CPUs and may be a bit slower.

The linux client is simply inefficient compared to win32.

I thought the fastest version of Speedimic is quite close.

Just looked it up, speedimics own Q9550 (45nm, 2.83GHz) is taking about 1030 seconds for a 27.77 credit WU under Linux.

He has another host (Q6600, 65nm, appears to be overclocked to 2.7-2.8GHz from the benchmark values) completing the same tasks in about 1150 seconds under Windows. It is a somehow bad comparison because of the different clockspeed and that the 45nm CPUs should be faster per clock than their 65nm counterparts, but nonetheless it is clear that his Linux version can't be that bad.

[@mfl0p:
The compiler flag "-fp-model fast=2" doesn't make any difference - nighter in speed nor in accuracy (at least on my test box).

Yes, release ATI app for Linux, please. CUDA for all systems is under construction now, but it makes sense to make app for stronger cards first and ATI is the one. I hate running my machine under Win just to be able to run Milky on GPU. :-( And I do not think there are more guys with NVidia GTS/GTX 200 cards than Linux ones with ATI 38xx/48xx.

PLEASE, PLEASE, PLEASE!!! make ATI GPU 0.20 app for linux x64. I'm pissed off using windows... Right now I'm gonna build a rig for MW, but if there is a chance to avoid this - it will perfect :-)

And furthermore - there is app for linux in Collatz. What's the problem to make the same trick for MW?
____________

PLEASE, PLEASE, PLEASE!!! make ATI GPU 0.20 app for linux x64. I'm pissed off using windows... Right now I'm gonna build a rig for MW, but if there is a chance to avoid this - it will perfect :-)

And furthermore - there is app for linux in Collatz. What's the problem to make the same trick for MW?

WOW, that's great news :-) Hopefully u'll get code soon

I'm not a geek in Linux (I'm using Ubuntu 3 years only), but u if need any help - feel free to contact me :-)
____________

how about a linux ati app.
____________
Blah blah blah you know the rest

how about a linux ati app.

I'll add my request in for a Linux ATI GPU application. Hasn't the source code been available for a while now?

The Linux x86-64 SSE4.1 app consistently crashed on me about 30% of the way through, with a SIGSEGV (segfault) error. I have a C2Q 9400 that should support it.

looks no1 really care :( that's weird...
____________

i also have a hd5870 just asking to crunch under linux64... is this being worked on or is my GPU doomed to only crunch collatz?

PLEASE, PLEASE, PLEASE!!! make ATI GPU 0.20 app for linux x64. I'm pissed off using windows... Right now I'm gonna build a rig for MW, but if there is a chance to avoid this - it will perfect :-)

And furthermore - there is app for linux in Collatz. What's the problem to make the same trick for MW?

Please see http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=1475, once the code is available I plan on attempting to compile a version for GNU/Linux.

7 months later ... any change?

I'm working on a new OpenCL one which should work everywhere(Linux/OS X/Windows/Nvidia/ATI) right now. I hope to have it ready to release in 1-2 weeks.

Hi Matt,

Nice to hear that :-)

R u talking about MilkyWay@Home Version 3 apps? if so, when we can expect them?
____________

I am thinking of going Linux with Intel and ATI/AMD GPU farm. How is this project shaping up? Any advice for set up?