[Ola_Arwidsson]

I've finally got one of the newer seti workers running! It's the "setiathome-4.18-applevdsp-panther-G4-alpha-4" from the noted webpage. The "seti@home-G4-a5" from the thread above gives client error.

Which seti workers are supposed to work on 10.3.9?

Also some of you give the time for "the reference unit", where do I find it, and how do I trick boinc to use it (or any other unit I want to reuse)?

[beadman]

alexkan:
Here's another set of results for you.

PowerBook G4 1.67GHz, 1MB RAM, 512KB L2, OS 10.3.9, BOINC 4.44 Superbench, SETI G4-alpha-4, hostid=280526
116668088 27701654 22 Sep 2005 4:30:09 UTC 22 Sep 2005 9:23:47 UTC Over Success Done 6,087.64 32.95 pending
116553788 27673686 21 Sep 2005 23:40:38 UTC 22 Sep 2005 4:40:16 UTC Over Success Done 5,626.90 30.46 29.87
116415259 27640203 21 Sep 2005 19:56:46 UTC 22 Sep 2005 1:00:07 UTC Over Success Done 6,075.20 33.08 pending
116290981 27609785 21 Sep 2005 12:23:45 UTC 21 Sep 2005 20:50:18 UTC Over Success Done 9,284.66 50.55 26.55
116194901 27586310 21 Sep 2005 8:31:07 UTC 21 Sep 2005 12:33:53 UTC Over Success Done 5,589.81 30.43 14.53
116079034 27558039 21 Sep 2005 3:46:08 UTC 21 Sep 2005 8:31:06 UTC Over Success Done 6,097.22 33.2 pending
115970474 27532316 20 Sep 2005 23:05:37 UTC 21 Sep 2005 3:56:15 UTC Over Success Done 5,612.87 30.56 30.56
115923904 27521068 20 Sep 2005 21:10:16 UTC 21 Sep 2005 1:10:06 UTC Over Success Done 223.15 1.21 14.26

Average time is 6,339 sec., low is 5,589, high is 9,284 - Interestingly, the single result with the extremely low 223 seconds, is valid, and the other two who ran it appear to have "normal" times with it. Sep. 20 and earlier, this machine was running the alpha-1 and averaging 15,354 seconds, with high of 21,763 and low of 9,053.

beadman

[alexkan]

Originally Posted by Ola_Arwidsson

I've finally got one of the newer seti workers running! It's the "setiathome-4.18-applevdsp-panther-G4-alpha-4" from the noted webpage. The "seti@home-G4-a5" from the thread above gives client error.

Which seti workers are supposed to work on 10.3.9?

Also some of you give the time for "the reference unit", where do I find it, and how do I trick boinc to use it (or any other unit I want to reuse)?

alpha-4 is the most recent version of the client that will run on Panther. I haven't compiled alpha-5 for Tiger yet, since the fallback vDSP functions aren't in place yet.

The reference unit and (mostly applicable) instructions for benchmarking with the reference unit can be found at http://marisan.nl/seti/reference.htm. In general, what you want to do is put the reference unit (renamed to work_unit.sah) and the client in a directory by itself, then run the client in a terminal with no other programs running. Once the client exits, look in init_data.xml for a key that mentions CPU time. You should recognize the number as being around the range of CPU times that you're already getting. In general, you can rerun a work unit by replacing work_unit.sah in the benchmarking directory with the work unit in question.

[Karl Schimanek]

A friend of mine had problems with a couple of WUs:

http://setiathome.berkeley.edu/resul...ltid=115232298
http://setiathome.berkeley.edu/resul...ltid=115063796
http://setiathome.berkeley.edu/resul...ltid=114814957
http://setiathome.berkeley.edu/resul...ltid=114119898
http://setiathome.berkeley.edu/resul...ltid=113454009
http://setiathome.berkeley.edu/resul...ltid=113453977
"Validate state Checked, but no consensus yet"

http://setiathome.berkeley.edu/resul...ltid=114815187
"Validate state Invalid"

He is running alpha5 on a G5 Dual 2.5GHz (Tiger)

My results:
http://setiathome.berkeley.edu/resul...userid=8169359
All valid

Karl

[BTBlomberg]

Originally Posted by beadman

Interestingly, the single result with the extremely low 223 seconds, is valid, and the other two who ran it appear to have "normal" times with it.

These are nothing new and have been around since the start of the SETI@Home Project (Classic). I am not sure, but I think they are small test or control WU that may provide some benchmarking or test that things are working properly for the SETI Team.

[rick]

Originally Posted by Karl Schimanek

A friend of mine had problems with a couple of WUs:

Can you keep us updated about the ones which are "Validate state Checked, but no consensus yet". If you or anyone else receives any more invalid work units, can you please post links to them so I can ask for copies from the SETI@home people. Or even mail them to me if you've still got the work units.

Until then I guess the only thing to do is to switch back to alpha 4 until we figure this out and release a fix. It might be something dodgy in the new Gaussian code which I haven't had chance to look over yet.

As far as I know we've only had 2 invalid results (both on G5s) but I'm unable to find the one that Snake Doctor posted (broken stats page...? they come up empty).

Anyone getting invalid results on G4s?

[electrician1972]

Rick, Check this one out. It happened early on and hasn't happened since. I'm pretty sure this was after I moved all my machines to G4 A5.

Result ID 114453468
Name 09no03aa.23825.2481.54842.127_0
Workunit 27161722
Created 17 Sep 2005 2:18:57 UTC
Sent 18 Sep 2005 7:00:47 UTC
Received 18 Sep 2005 10:34:31 UTC
Server state Over
Outcome Client error
Client state Computing
Exit status 131 (0x83)
Computer ID 1416352
Report deadline 2 Oct 2005 7:00:47 UTC
CPU time 6537.409379
stderr out
<core_client_version>4.44</core_client_version>
<message>process exited with code 131 (0x83)
</message>
<stderr_txt>
SIGSEGV: segmentation violation
Exiting...

</stderr_txt>
Validate state Invalid
Claimed credit 15.298809553121
Granted credit 0
application version 4.18

[BTBlomberg]

Here is feedabck on how my machines are doing with Alpha5. I have seen a 7.3 to 8.5 % increase over Alpha-4 between the two machines.

Code:
PowerMac G4 Dual 866
Mac OS 10.4.2

Alpha-5 G4 Opt Seti 4.18 __________ Increase Multiple __ vs JL
Results Returned/Verified: ____ 12/3
Average CPU time (sec): ____ 8125.13 ____ 1.085 ________ 3.72
Average claimed credit: ______ 22.96
Average granted credit: ______ 22.26

Alpha-4 G4 Opt Seti 4.18
Results Returned/Verified: ___ 20/19
Average CPU time (sec): ____ 8817.82
Average claimed credit: ______ 24.92
Average granted credit: ______ 24.29

Now teh results for the PowerBook are not as complete as the PowerMac yet, so once more WUs are processed I will update.

Code:
PowerBook G4 500
Mac OS 10.4.2

Alpha-5 G4 Opt Seti 4.18 __________ Increase Multiple __ vs JL
Results Returned/Verified: _____ 2/0
Average CPU time (sec): ____ 9346.21 ____ 1.073 ________ 4.32
Average claimed credit: ______ 18.29
Average granted credit: ________ N/A

Alpha-4 G4 Opt Seti 4.18
Results Returned/Verified: ___ 11/10
Average CPU time (sec): ___ 10032.34
Average claimed credit: ______ 19.61
Average granted credit: ______ 20.53

Unfortunately I have no Stock SETI@Home WU Times for these machines to compare the overall increase above that base.

Also, I noticed one thing that did not happen when replacing SETI Client SW that has when I have changed between Worker Versions and that is changing out the Worker software (changing from Alpha-4 to Alpha-5) did not wipe out all WU including the one in progress. THis has happened to me anytime I have switched in teh past, but likely becasue the Worker info is the same between compiles it looks like the same Worker to Boinc and the WU in progress. That is nice.

[amigoivo]

Originally Posted by rick

Can you keep us updated about the ones which are "Validate state Checked, but no consensus yet". If you or anyone else receives any more invalid work units, can you please post links to them so I can ask for copies from the SETI@home people. Or even mail them to me if you've still got the work units.

Until then I guess the only thing to do is to switch back to alpha 4 until we figure this out and release a fix. It might be something dodgy in the new Gaussian code which I haven't had chance to look over yet.

As far as I know we've only had 2 invalid results (both on G5s) but I'm unable to find the one that Snake Doctor posted (broken stats page...? they come up empty).

Anyone getting invalid results on G4s?

Hello rick,

i am the mysteriously friend of Karl.
I have a DUAL 2.5GHz PowerMac with 3GB Ram, and i run Seti and Einstein.
Today my Mac have to crunch the whole day, because a´m not at home.

I will search tomorrow in the about 30 new completet WUs for any similarly problem and post it here.

Have a nice day,
Thanks, Ivo

p.s.: Thanks to the Translate-Widget of Tiger.

[Shaktai]

Okay, I am kind of working backwords here but here are results with the reference work unit on an iMac G5 1.6 ghz. This is with the G5 App.

a5 - 5678 seconds
a4 - 5667 seconds
a3 - 5637 seconds

I wasn't able to download alpha 1 and will run alpha 2 and maybe javalizards client this weekend when I have a little more time. Interestingly, there is less then a minutes difference, and that could be due to as little as me checking the state.sah file periodically to view the progress. Of course, anyway you look at it, it is still a huge improvement, around 4x from the stock app.

Originally Posted by rick

Can you keep us updated about the ones which are "Validate state Checked, but no consensus yet". If you or anyone else receives any more invalid work units, can you please post links to them so I can ask for copies from the SETI@home people. Or even mail them to me if you've still got the work units.

Until then I guess the only thing to do is to switch back to alpha 4 until we figure this out and release a fix. It might be something dodgy in the new Gaussian code which I haven't had chance to look over yet.

As far as I know we've only had 2 invalid results (both on G5s) but I'm unable to find the one that Snake Doctor posted (broken stats page...? they come up empty).

Anyone getting invalid results on G4s?

No problems on my PowerMac G4 single-1.25. @60 units, no validation errors. @10% faster than Alpha-4 and averaging @ 5400sec per WU on my machine. Seems the 7450 series G4, 4.44 + Alpha-5 combo really works well! Keep the revisions coming!

Originally Posted by GeckoR7

No problems on my PowerMac G4 single-1.25. @60 units, no validation errors. @10% faster than Alpha-4 and averaging @ 5400sec per WU on my machine. Seems the 7450 series G4, 4.44 + Alpha-5 combo really works well! Keep the revisions coming!

By the way, 10.4.2

[rick]

Originally Posted by electrician1972

SIGSEGV: segmentation violation
Exiting...

Damn it. This means that the code is wrong and addressing memory it shouldn't do. Could someone within 20 miles of Berkeley go and throttle Alex. Cheers.

I've had some help from the guys on the boinc_opt mailing list. It is possible to download work units after completion, but only for a short time (I don't know how long).

I've managed to download some of those which are flagged "Validate state Checked, but no consensus yet" but none of those which are invalid (presumably because they're much older). It will take me about 4 hours to test each one.

If you get an invalid result can you mail me (I check my mail more often than the forum) the work unit name (e.g. 09no03aa.23825.2481.54842.127_0) so I can download it. If you want to get it yourself you can run this Python script:
http://writhe.org.uk/seti@home/setiURL
Run it in the Terminal using something like:

Code:
python setiURL 09no03aa.23825.2481.54842.127_0

If the work unit is no longer available it will say "Unable to get work unit", otherwise it will save it to a file with the appropriate name.

[Lateralus]

Okay, am I doing something wrong?

With A5, I am getting 10,000-11,000/sec per WU? These results are from last night, when I was running BOINC Menubar as stock. This morning I added the AltiVec BOINC worker to the BOINC app. Will this make a difference?

[BTBlomberg]

I have been using BOINC Menubar 4.44 from:
http://members.dslextreme.com/~reade...eam/boinc.html
But I have taken the BOINC 4.44 G4 SuperBench client from that page and used it to replace the Boinc app in the package by coping it into the app package and changing the name to just "boinc" after renameing or removing the origional.

Has worked well for me. I am not sure if it will do what you need or will give you anything different.

[chboss]

Hi Rick

My 3 G4 based machines have not turned in any invalid results so far.
I am running Alpha-5 under 10.4.2 on all three.

One iMac 1.6GHz G5 is running under 10.3.9 also there no invalid results.

[BTBlomberg]

Likewise, on the 2 G4 mqchines I have been commenting on with Alpha-5 no invalid results so far. Several are still pending but none have been bad.

[amigoivo]

Originally Posted by rick

Can you keep us updated about the ones which are "Validate state Checked, but no consensus yet". If you or anyone else receives any more invalid work units, can you please post links to them so I can ask for copies from the SETI@home people. Or even mail them to me if you've still got the work units.

Until then I guess the only thing to do is to switch back to alpha 4 until we figure this out and release a fix. It might be something dodgy in the new Gaussian code which I haven't had chance to look over yet.

As far as I know we've only had 2 invalid results (both on G5s) but I'm unable to find the one that Snake Doctor posted (broken stats page...? they come up empty).

Anyone getting invalid results on G4s?

Hello rick,

in the last results was only one with the status "Checked, but no consensus yet"
Link: http://setiathome.berkeley.edu/resul...ltid=116847938

and one invalid result was in the results that were checked from Karl:
Link: http://setiathome.berkeley.edu/resul...ltid=115063698

There were only 8 WUs with staus "pending" that i will check later.

Have a nice weekend,
Greetz Amigoivo

[rick]

After some more investigation, work units with "Checked, but no consensus yet" just mean that the full set of results (from other users) have not been returned yet, so there's no way to find the "canonical" result and therefore allocate credit. The ones that I've tested have validated fine.

One last thing, if anyone's using the Python script then you have to remove the trailing underscore and all the characters after it. So:

Code:
python setiURL 09no03aa.23825.2481.54842.127_0

should actually be

Code:
python setiURL 09no03aa.23825.2481.54842.127

Originally Posted by amigoivo

and one invalid result was in the results that were checked from Karl:
Link: http://setiathome.berkeley.edu/resul...ltid=115063698

Cool, this is the first invalid result I've been able to get the work unit from. I'll give this a run overnight. It should give a good indication of where we've gone wrong.

[Lateralus]

Originally Posted by BTBlomberg

I have been using BOINC Menubar 4.44 from:
http://members.dslextreme.com/~reade...eam/boinc.html
But I have taken the BOINC 4.44 G4 SuperBench client from that page and used it to replace the Boinc app in the package by coping it into the app package and changing the name to just "boinc" after renameing or removing the origional.

Has worked well for me. I am not sure if it will do what you need or will give you anything different.

Downloaded the BOINC Menubar app. Replaced the BOINC file with the Superbench version. Added the AltiVec SETI worker, Alpha 5, to the proper folder.

These are the times I am getting now;

http://members.arstechnica.com/x/powermacman/BOINC.png

Still crap.

[Karl Schimanek]

Here is another invalid result from his computer:
http://setiathome.berkeley.edu/resul...ltid=114815187

It's a older one.

Karl

[BTBlomberg]

Originally Posted by Lateralus

Downloaded the BOINC Menubar app. Replaced the BOINC file with the Superbench version. Added the AltiVec SETI worker, Alpha 5, to the proper folder.

These are the times I am getting now;

http://members.arstechnica.com/x/powermacman/BOINC.png

Still crap.

I have to say that looks more like the Javalizard Client. I would check to be sure it isn't hidden in a Library folder somewhere and the newly installed version are not the ones running. One reason I say that is the Javalizard Client has horrible overestimates for Claimed Points that your picture shows.

If you look in the "Activity Monitor" -> "All Processes" and see "setiathome_4.3_..." then it is the Javalizard one. If you see "seti@home-G5-a5" you have the right one. If the 4.3 one search for Seti or Boinc for the bad folder after you kill the offending Seti in the "Activity Monitor" (found in Applications->Utilities).

[Lateralus]

It's "seti@home-G4-a5".

[alexkan]

Originally Posted by Lateralus

It's "seti@home-G5-a5".

What is Processor Performance set to in Energy Saver? (If it's on Automatic, try setting it to highest. I'm not sure what the criteria are for DFS running the chip at half speed, but it sure sounds like you're getting about half the performance that I would expect from your particular model.)

[Shaktai]

Okay, here are the Reference unit times for Alpha 2 through Alpha 5 for the G5. Looks like the big gain came from 2 to 3, and nothing since. The variances between a3 and a5 are probably no more then other minor computer activity. Or in other words, normal variances.

These times are for the G5 app only and do not represent the G4.

iMac G5 1.6 ghz. 1 gig RAM,

a5 - 5678 seconds
a4 - 5667 seconds
a3 - 5637 seconds
a2 - 6307.63 seconds

[alexkan]

Originally Posted by Shaktai

Okay, here are the Reference unit times for Alpha 2 through Alpha 5 for the G5. Looks like the big gain came from 2 to 3, and nothing since. The variances between a3 and a5 are probably no more then other minor computer activity. Or in other words, normal variances.

These times are for the G5 app only and do not represent the G4.

iMac G5 1.6 ghz. 1 gig RAM,

a5 - 5678 seconds
a4 - 5667 seconds
a3 - 5637 seconds
a2 - 6307.63 seconds

Interesting...looks like I may have to take another look at where the G5's bottlenecks are. Incidentally, these binaries, unlike alpha-3 and earlier, include debug symbols, so if anyone wants to profile the reference work unit in Shark on a G5 for us, the results should actually be meaningful. (Shaktai, I know we discussed this a while back, so if you could give it another try, that'd be great.) Ideally, sampling at 125 ms with no time limit and counting dL2 misses should give us most of the information we need to tweak.

An alternate solution is to buy G5s for Rick and me, but I suspect you all would prefer just taking Shark traces instead.

[Karl Schimanek]

Yes, a G4 at 2.2GHz should outperform a G5 at 2.5GHz on SETI.
A 7448 more than ever
Apples Intel switch – brilliant strategy

MPC 7450 733MHz:

A1: 19.000 seconds
A3: 12.200 seconds
A4: 11.150 seconds
A5: 10.900 seconds

No ref. unit, only the longest run on different units.

[Shaktai]

Originally Posted by alexkan

Interesting...looks like I may have to take another look at where the G5's bottlenecks are. Incidentally, these binaries, unlike alpha-3 and earlier, include debug symbols, so if anyone wants to profile the reference work unit in Shark on a G5 for us, the results should actually be meaningful. (Shaktai, I know we discussed this a while back, so if you could give it another try, that'd be great.) Ideally, sampling at 125 ms with no time limit and counting dL2 misses should give us most of the information we need to tweak.

An alternate solution is to buy G5s for Rick and me, but I suspect you all would prefer just taking Shark traces instead.

Okay, I am running G5-a5 with Shark now. How do I get you the results?

[alexkan]

Originally Posted by Shaktai

Okay, I am running G5-a5 with Shark now. How do I get you the results?

There should be an option to save a completed Shark trace from the File menu. Embed source in the trace where possible (it's probably safe to click 'Ignore' for the files that Shark can't find). Then just email it to me, or do a YouSendIt thing or something.

[Lateralus]

Originally Posted by alexkan

What is Processor Performance set to in Energy Saver? (If it's on Automatic, try setting it to highest. I'm not sure what the criteria are for DFS running the chip at half speed, but it sure sounds like you're getting about half the performance that I would expect from your particular model.)

I'm on a G4, no Energy Saver performance settings.

And I was mistaken, I just copy-pasted what you had said. I am indeed running the G4 version of the client.

[alexkan]

Originally Posted by Lateralus

I'm on a G4, no Energy Saver performance settings.

And I was mistaken, I just copy-pasted what you had said. I am indeed running the G4 version of the client.

Hmm...in that case, I'm tempted to ask you to run the client on the reference unit from a terminal to see what times you get in that case. Also, if you're running a PowerLogix CPU upgrade, do you have CPU Director installed? (It seems like 7447As are supposed to have some amount of frequency switching. I was inspired to ask by the discussions on the main SETI boards, since apparently running SETI at a low priority is not enough to get AMD CPUs with Cool'n'Quiet to ramp up to their full clock speed.)

[BTBlomberg]

Sorry, Lateralus. That was my mistake. I see you are on G4 but my Brain lusts for a G5 I guess.

Still odd results.

[Lateralus]

Running the client through Terminal doens't seem to make much difference. The upgrade is a GigaDesigns, set clock speed. No throttling.

I'm remembering why I left SETI BOINC for RC5 again...

[Shaktai]

Originally Posted by Lateralus

Running the client through Terminal doens't seem to make much difference. The upgrade is a GigaDesigns, set clock speed. No throttling.

I'm remembering why I left SETI BOINC for RC5 again...

Lateralus, your times do seem to be the exception rather then the rule. For some reason your box isn't making good use of the altivec instructions. Have you considered trying Einstein@Home? The default app is altivec enabled (faster then the Intel or AMD boxes), and might make real good use of your gigadesign upgrade. This would be a good way to test if it is a design issue with the gigadesign, or a possible code issue with the altivec SETI apps. You could also try running the reference SETI unit with SHARK and send the results to Alex. That might help trace any code issues.

To test it with Einstein:
- Set a very small cache while attached to SETI. no more then .1 or .2 This should download only one or two work units. when you attach to Einstein.
- Run the units. As a reference point, my iMac G5 1.6 ghz averages 19,000 to 20,000 seconds per unit.
- set "No new work" after your first download will prevent new work, until you have tried it out.

Your decision though. It is always frustrating to not get the most out of your computer.

[alexkan]

Originally Posted by Lateralus

Running the client through Terminal doens't seem to make much difference. The upgrade is a GigaDesigns, set clock speed. No throttling.

I'm remembering why I left SETI BOINC for RC5 again...

Sorry this isn't working out for you...I might be just as frustrated about this as you are. Still trying to figure out why this might be the case, though. Considering that the limiting factor for the client on most systems is how quickly we can stream the work unit data into memory during processing, it seems like something is holding you back in terms of memory bandwidth. I suppose the best thing you could do for us as this point would really be to generate some Shark traces and/or provide us with more detailed system specs.

Still, if this is already too much work for too little reward and you want to switch to Einstein to help the team, I totally understand.

Also, Shaktai, with regard to the timings for the alphas, I've come to the realization that the timing numbers in init_data.xml tend to overestimate the amount of CPU time that the client uses. If that's where you got the numbers from, try timing the client just using "time ./seti@home-G5-a5". Of course, since timing the alphas would take on the order of 10 hours, no hurry. I could actually post my own timings, but I suspect that as one of the developers, I wouldn't be able to help being at least a little biased.

[BTBlomberg]

Here is an update on results from Alpha-5 on my two G4s. The increase over Alpha-4 looks to be more around 2.6% to 4.2% with this set of WUs. Undestandibly this all has to do with the WU you draw but the avarage should be a good measure. Still the increase over the Javalizard client remains impressive at 3.57 to 4.12 times the speed.

Still no Bad WUs, just some with outstanding verification which is normal.

As of 9/25/2005

Code:
PowerMac G4 Dual 866
Mac OS 10.4.2

Alpha-5 G4 Opt Seti 4.18 ___________ Increase Multiple __ vs JL
Results Returned/Verified: ___ 67/45
Average CPU time (sec): ____ 8466.02 ______ 1.042 ________ 3.57
Average claimed credit: ______ 23.91
Average granted credit: ______ 22.22

Alpha-4 G4 Opt Seti 4.18
Results Returned/Verified: ___ 20/19
Average CPU time (sec): ____ 8817.82
Average claimed credit: ______ 24.92
Average granted credit: ______ 24.29


PowerBook G4 500
Mac OS 10.4.2

Alpha-5 G4 Opt Seti 4.18 ___________ Increase Multiple __ vs JL
Results Returned/Verified: ___ 21/18
Average CPU time (sec): ____ 9778.85 ______ 1.026 ________ 4.12
Average claimed credit: ______ 19.16
Average granted credit: ______ 20.34

Alpha-4 G4 Opt Seti 4.18
Results Returned/Verified: ___ 11/10
Average CPU time (sec): ___ 10032.35
Average claimed credit: ______ 19.61
Average granted credit: ______ 20.53

on my B&W upgraded to a G4 1GHz the alpha-5 client is working without any errors with average timings around 15200. from the alpha-1 to the alpha-5 the timings have went down from 35000 to this value of 15200. the last release was lower improvement with a gain of about 2000 for each wu. good job keep on that way guys

[Snake_doctor]

Ok, so I have been busy with other stuff lately, and I had not been keeping up. Last night I upgraded my G4 Powerbook (1 GHz, OS 10.3.9) from the A2 app to the A4 app. I also went from the MacNN 4.44 client, to the MacNN 4.44 superbench client. Now all I can think is that I must have done something wrong when I installed A2, or the MacNN 4.44 client before, because my WU times have gone from about 5 hours to right at 2 hours. Moreover the the results seem to be verifying.

So I thought why not try this on the G4 Dual? So I installed it there as well. The Dual was already in the 2.5 to 3 hour range for a WU. Now it is more like a little under 2 hours per.

So I had to figure this was a fluke right? Maybe I was just getting small WUs or something. Well the improvements seem to be holding. Between the two systems there are now 7 WUs that show CPU times almost half of what I had before.

So I have to ask Alex, what did you leave out? When I first started running the release version of Boinc SETI a WU took about 7 hours on my G4 Dual, and about 9 on the Powerbook. Now they "squirt" through the systems like P@H WUs! Outstanding work. As I get enough Stats to be useful I will post actual CPU times rather than clock times.

Regards
Phil

[Lateralus]

Eh... I give. Back to RC5-72.

[alexkan]

Originally Posted by Lateralus

Eh... I give. Back to RC5-72.

Sorry it didn't work out for you, Lateralus. I know RC5-72 is hard to beat in terms of sheer optimization for G4s. I also get the feeling that it's not quite as dependent on memory bandwidth as SETI BOINC seems to be.

In other news, Rick reports that he's gotten fat binaries working, so alpha-6 will be released as a fat binary...when it comes out. As it turns out, the chirp function should be able to get a bit of easy speedup on G5s. I won't overstate the performance gains this time, since I clearly can't account for all the hardware combinations (especially CPU upgrades) that these clients will be run on.

Also, I'm not surprised that you've seen such a huge jump on your 10.3.9 machines, Snake_doctor, since 10.3.9 has technically only seen releases of alpha-1 and alpha-4. I should really be less lazy about writing fallback code for vDSP functions.

[Snake_doctor]

Well the results are mounting, but I have two that are very interesting.
http://setiathome.berkeley.edu/worku...?wuid=28348506
and
http://setiathome.berkeley.edu/worku...?wuid=28320105

I have no idea how these two could process in less that 90 seconds and still validate, when the other systems that crunched them took normal times. I even was awarded credit on the first one, so it must be ok.

The numbers are still holding. Both systems are averaging in the area of 5600 seconds consistantly. I can't believe it. So far none of the WUs have errored out and the only two that even look strange are the two above. Both from system id 1359569 which is the G4 Dual.

Regards
Phil

[Snake_doctor]

Originally Posted by alexkan

Also, I'm not surprised that you've seen such a huge jump on your 10.3.9 machines, Snake_doctor, since 10.3.9 has technically only seen releases of alpha-1 and alpha-4. I should really be less lazy about writing fallback code for vDSP functions.

Well for the first time all is right with the universe. My Macs are eating the Windoze machines for lunch.

[BTBlomberg]

Snake_Doctor,

These are normal as they are likely SETI test Units ad not you normal WU. I have seen them sent out as frequently as with just a few real ones in between or it's gone with 35 in between, but they send them out frequently and randomly. When I average I throw them out as strays. No need to worry about them and they have been part of SETI@ Home since the beginning (Classic).

think this is te second time I have pointed this out in this tread.

[Snake_doctor]

Originally Posted by BTBlomberg

These are normal as they are likely SETI test Units ad not you normal WU.

I would agree except that the other systems have normal processing times. I am begining to suspect something more like the E@H and Rosetta@H problem, where the system would lose track of the total time on the WU, and report a bad time but a good overall result. Although I have no reason that this would happen. I have not turned off Boinc or done any of the things that caused this at E@H and Rosetta.

I too remove them from any averaging I do. I just thought Alex or Rick might want to see them.

Regards
Phil

[BTBlomberg]

Forgive my typing breakdown (misspelling).

I have to say that I have see this since 1999 with the Graphic Classic SETI and I have see this happen with no interuptions or interference. Thinking about it a little more it could be that they have an auto splitter program for the recordings and it may cut them into peices of equal length. These files may be the odd remainders of these files and small enough that the clients easily figure out there are no peaks and rush them through.

Otherwise the test file thing works too. Hard to know without a SETI insider, but I think they have nothing to do with an error anywhere. But that is just my take.

[dtsang]

Running the alpha-4 G4 client on a Power Mac G4/466 (640 MB CL3 memory), Mac OS X 10.3.9.

I've done five units so far, one of them already verified and credited. My average time is 9489 seconds (range 8651-10211 sec) - a most excellent result, considering this machine is more than four and a half years old! I'm looking forward to even faster processing with the soon-to-be-released alpha-6 client. Good job and keep up the great work.

[Shaktai]

Originally Posted by BTBlomberg

I have to say that I have see this since 1999 with the Graphic Classic SETI and I have see this happen with no interuptions or interference. Thinking about it a little more it could be that they have an auto splitter program for the recordings and it may cut them into peices of equal length. These files may be the odd remainders of these files and small enough that the clients easily figure out there are no peaks and rush them through.

Otherwise the test file thing works too. Hard to know without a SETI insider, but I think they have nothing to do with an error anywhere. But that is just my take.

There is also the noise factor. Some SETI work units are just noisy. (Terrestrial interference usually) The app will scan through the noise quickly, but will never be able to find any useful data to process and will finish the work unit in a very short time. This is a fairly common occurance. More so then the frequency of "test" units. Noisy units tend to come in small batches, so it wouldn't be unusual to get more then one at a time. The unit still validates, and you still get credit for the time spent on it.

[BTBlomberg]

Originally Posted by Shaktai

There is also the noise factor. Some SETI work units are just noisy. (Terrestrial interference usually) The app will scan through the noise quickly, but will never be able to find any useful data to process and will finish the work unit in a very short time. This is a fairly common occurance. More so then the frequency of "test" units. Noisy units tend to come in small batches, so it wouldn't be unusual to get more then one at a time. The unit still validates, and you still get credit for the time spent on it.

That's a great point and may be the best answer.

although I have seen some of these units to have less time or files size to them way beck when, but have not payed attention lately to that detail. Using the download commands in this tread one could compare them though. I will take a look just for the sake of it.

[Snake_doctor]

Originally Posted by BTBlomberg

Forgive my typing breakdown (misspelling).

I have to say that I have see this since 1999 with the Graphic Classic SETI and I have see this happen with no interuptions or interference. Thinking about it a little more it could be that they have an auto splitter program for the recordings and it may cut them into peices of equal length. These files may be the odd remainders of these files and small enough that the clients easily figure out there are no peaks and rush them through.

Otherwise the test file thing works too. Hard to know without a SETI insider, but I think they have nothing to do with an error anywhere. But that is just my take.

I still don't see how the other systems in the same quorum could process them in normal times if it is something in the WU. I have seen many of these short WUs myself both here and in classic, but never when other systems took over 5000 CPU seconds and I only took about 80. Unless they are testing certain system types and there is something in the WU thas says to my system "not you" so it dumps out. The other quorum machines were all Windoze systems.

In any case I have just done some averages (without the two WUs in question) and it looks like this-

G4 Powerbook 1GHz, OS 10.3.9, MacNN Client 4.44 Superbench, AlexKan A4 app. -
5,986.73 CPU Seconds/ WU

G4 Powerbook 1GHz, OS 10.3.9, MacNN Client GUI 4.44, AlexKan A1 app. -
18,030.83 CPU Seconds/ WU

G4 Mirror 1.4GHz, Dual CPU, OS 10.3.9, MacNN Client 4.44 Superbench, AlexKan A4 app-
5,487.96 CPU Seconds/ WU

G4 Mirror 1.4GHz, Dual CPU, OS 10.3.9, MacNN GUI Client 4.44, AlexKan A1 app-
11,410.12 CPU Seconds/ WU

Now In anybodys book that is an impressive performance improvement. I do not have any WU stats from the Javalizard MacNN app but I remember that A1 was at least twice as fast as that.

When I started crunching Boinc SETI the WUs were taking between 7 and 9 hours running all standard Boinc release software. So overall the performance increase is on the order of 450% because I am running a WU in under two hours now. (and in some cases 70 seconds )

I have noticed something else that you guys probably already knew. The Menubar client uses less CPU cycles that the GUI client, There are periods where the MB client goes completely quiet but the GUI is always using from 3 to 10% CPU. While this does not effect the CPU time for the WU it has an effect on overall throughput. I have also discovered that you can start the MB client and while it is crunching you can use the GUI client to ocassionally "Tweek" and "Peek" at things, or control other systems remotely. I wish the MB version had all the same functions as the GUI for controling the WUs.

By the way forget the typing, we all do it especially at night.

Most people do not realize that if the first and last letters of the words are correct and the number of letters in the words are correct, your brain will read the words correctly no matter what letters are in the middle. You were close enough.

[BTBlomberg]

Yes the Command Line and MenuBar are best to run. GUI just slows it all down. Have not Run GUI since Classic. The MenuBar is really nice becasue it does not take too much but can keep you on top of the WUs whereas the Command Line leaves you blind. MenuBar is just the Command Line with an interface.

Only graphics I need is the little progress line next to the dish.