[sdubz]

OK!!! I think I have gotten somewhere!


I have 3 results now with one of my compiles that seems to be working!


this one has been validated: http://setiathome.berkeley.edu/resul...ltid=326284363


Waiting for this one to validate: http://setiathome.berkeley.edu/resul...ltid=325871709


This one has a SETI@Home Informational message -9 result_overflow
NOTE: The number of results detected exceeds the storage space allocated. error: http://setiathome.berkeley.edu/resul...ltid=326909117 but so does the same wu from a pentium machine! http://setiathome.berkeley.edu/resul...ltid=326909118 so I think it is actually a "noisy" wu.


I'm going to try another build this time before I jump the gun again and try to release it.


I would also like to say that I'm not actually modifying any code, I am just turning on optimization flags that seem to have been built in like -O3 -unroll loops, etc. Altivec is there too but I'm not sure if it is actually built into the code, so if it isn't then I bet Alex and Rick could make a screamer!


I found that the nightly build I was using seemed to be the problem with the last build, using the nightly tarball from 5-11-2006 seemed to fix that.


Now I am going to try to see if it will build against boinc version 5.5.0 instead of 5.3.12 (not sure if that will make a difference).


Then I will make a test release.

[Gecko_r7]

Thanks for all your efforts Boog, you're doing great! Sounds promising. I'm excited as well to see what Alex and Rick can wring-out of this this. Good luck w/ your testing and please let me know if I can help.

[sdubz]

Originally Posted by Gecko_r7

Thanks for all your efforts Boog, you're doing great! Sounds promising. I'm excited as well to see what Alex and Rick can wring-out of this this. Good luck w/ your testing and please let me know if I can help.

Well, now I'm just starting to get frustrated, this last couple builds are actually taking longer than the original client!

eventually I may figure some of this out.

[lepetitmartien]

As I Switched to the official app yesterday evening I can tell you it's not all your fault, they released WU very demanding since yesterday (claiming in the 9-10 hours on my G5).

Also there seems to be quite a lot of problems into processing WU if you look at the other crunchers on the same units as you, we're on a bumpy road…

Thumbs up Boog, thanks for all you efforts.

[sdubz]

Man, I have learned alot trying to take this on! it's not as easy as I had thought just to add little optimize flags here and there!

I have enhanced workers that need tested, they seem to be running and reporting fine, but I still need more checks to make sure.

G4>> http://boog.is-a-geek.org/seti/seti_enhanced_g4_b5.tgz
G5>> http://boog.is-a-geek.org/seti/seti_enhanced_g5_b5.tgz

and as an added test (I have no way of testing an intel binary)
INTEL>> http://boog.is-a-geek.org/seti/seti_...d_intel_b5.tgz


These all seem to support graphics, and the screensaver works.
The universal binary for boinc version 5.5.0 with universal screensaver is here : http://boog.is-a-geek.org/boinc/boin...creensaver.tgz

Please let me know how these do, and remember these "may" not work correctly, and are not as fast as the "old" version of seti. And I still bet Alex and Rick could make this faster than I can!


Thanks!
Boog

[arkayn]

I will see what it does on my eMac.

[Gecko_r7]

Please let me know how these do, and remember these "may" not work correctly, and are not as fast as the "old" version of seti. And I still bet Alex and Rick could make this faster than I can!


Thanks!
Boog[/quote]

I'll give it a go on the G4.

[lepetitmartien]

trying the G5 one, crunching happily for now (about 30% of a WU I had already in the making). It's noticeably faster (it'd better be!) and it crunches. Now will it validate?

[sdubz]

Originally Posted by lepetitmartien

trying the G5 one, crunching happily for now (about 30% of a WU I had already in the making). It's noticeably faster (it'd better be!) and it crunches. Now will it validate?

I sure hope it does!

All of the wu's I have crunched using this build have validated, even the ones that had the -9 error because the other computers that shared that wu had the same error.

http://setiathome.berkeley.edu/resul...ltid=326721031
http://setiathome.berkeley.edu/resul...ltid=327386606

here is the latest one that it just reported http://setiathome.berkeley.edu/resul...ltid=326957162 still waiting on it to be validated (as of 6am eastern).


I'm glad to see it is deffinately going faster! I have been messing with this so much I was starting to have trouble seeing if it is faster or if I was just wishing it is faster! lol

[jedimstr]

Will try it out on my Quad G5, Powerbook G4 1.5, G4 Mini, and eMac over the weekend.\

Crossing fingers...

Of course it "would" be nice to actually have Altivec optimizations included as well...

[sdubz]

Originally Posted by jedimstr

Will try it out on my Quad G5, Powerbook G4 1.5, G4 Mini, and eMac over the weekend.\

Crossing fingers...

Of course it "would" be nice to actually have Altivec optimizations included as well...

If I knew how to check to see if they are actually there I would. Altivec should be built into the ffwt libraries (and I compiled the g4 and g5 versions separetly and for each cpu specifically), then they are linked against the seti executeable.

So, at least that much is altivec, if there can be more done in the actual seti code then these workers could be awesome!

[Knightrider]

Hi Boog,

I tried out the G5 version on my Quad and it seems to be working ok.

Here are the first 3 results

1/ here It initial validated where four others failed.

2/ here initial validated ok and was quickest at 11938.50 cpu time

3/ here fully Valid ok at 2nd quickest with 16991.76. Only just behind a time of 16304.64 for the quickest.

So well done. I notice that the wu's vary a lot in length of TTC (time to completion), don't know if that makes any difference to your optimization.

Maybe Rick and Alex could use this as a base for altivec enhancement, just a thought.



K.

edit 1246

[lepetitmartien]

First 2 WU in the benchmarks

Here and there

I'm "2397079" (I'm not a number, I'm a free man!)

The "there" unit was partially crunched (about 50%) with the official app. I'm doing well for my old G5. If it goes that good later, we'll "only" need some altivec thrown in.

[Gecko_r7]

HI Boog,

Ran a few on my G4

http://setiathome.berkeley.edu/resul...ltid=328311660
http://setiathome.berkeley.edu/resul...ltid=328315049
http://setiathome.berkeley.edu/resul...ltid=328345363
http://setiathome.berkeley.edu/resul...ltid=328268901

Running fine, no validation issues so far.
I'm still trying to get a handle on the speed since it's more difficult to compare WUs as length depends on the AR. My perception is that its faster, just can't tell you how much yet until I get more WUs crunched. I'd say , we've got a winner in the making. Excellent job! Now, "who" do we know that can make Altivec really shine......Humm....

[lepetitmartien]

Another one for me

On the time gain, it did in about 5 hours a WU of "8 hours", note that this is from memory, so not very scientific. On the comparison side, I do roughly the same as non optimized 2.8-3 GHz pentium 4, but the same with I suppose crunch3r app for XP are faster still it seems (nearly twice).

So there's still room for improvement… (altivec… where is my altivec…)

On a side note, we are very very fast when crunching noise. Like here (note, it was with the first Bogg trial this one)

[halimedia]

I'd suggest that a reference-WU be used (chosen by boog, I guess), and performance comparisons on different machines would be run against that WU (just like with the previous optimized workers). That way, we'd know for sure to what extent there are performance improvements in the optimized workers. This appears particularly useful, because the enhanced WUs seem to show much greater variability with regard to CPU time required to completion.

Just a thought...

Ron

[Gecko_r7]

Originally Posted by halimedia

I'd suggest that a reference-WU be used (chosen by boog, I guess), and performance comparisons on different machines would be run against that WU (just like with the previous optimized workers). That way, we'd know for sure to what extent there are performance improvements in the optimized workers. This appears particularly useful, because the enhanced WUs seem to show much greater variability with regard to CPU time required to completion.

Just a thought...

Ron

I agree. Between the WU length (AR) and CPU speed differences, it will be a real pain trying to measure and compare speeds. Need a reference unit.

[tony.escobar]

I started looking at the completion times for the WUs crunched with boog's app. The most interesting comparison was Gecko_r7's 4th WU: It's now fully validated and Gecko had the 3rd slowest time on his PM G4 MDD. The slowest time was from a PowerMac 10,1 (G4 mini)from some guy on Team AOL(!?).

Looking at the other WUs crunched with boog's app:

Knightrider's #2 (on a G5 Quad) has now fully validated and with the fastest time. Others were 2.8GHz & 3GHz P4s.

Knightrider's #3—2nd-fastest, just behind a Pentium M

Letpetitmartien—G5
“Here” WU was a short WU, now fully validated, and lpm was 2nd-fastest (by ~100 and ~200 seconds over 3rd and 4th place machines). Third- and fourth-place machines are Athlon 64-bit FX-60s running WinXP.

“There” WU was 2nd-fastest. Two slower machines were 2.4 and 3.2 P4s, which claimed same credit as lpm.

“Another One” now fully validated. lpm’s was the slowest time of three. The other machines were 3.2 and 3.0 P4s.

Of Gecko_r7 other 3 WUs, only #1 has been fully validated. Basically tied for slowest among 4 machines (other machine with similar time was Athlon 1.6Ghz running XP).

Note that none of the other machines appeared to have used optimized apps.

QS

[lepetitmartien]

Yep, this RA affair is confusing a lot things…

Got another WU coming… (the next after will be a lonnnnng one)

Also, you can name me LPM for short