v6.07 Beta is running on my 9800GT, initially with just four 'resent lost result' tasks cached prior to reset.

Initial impression is good - work duration estimate seems accurate, and speed good running BOINC v6.4.5 at 4+1 without any external assistance (process lasso or suchlike).

But these resent tasks are guaranteed VLAR-free (I filtered them last time....), so the acid test will come later.

First result back (task 5229551), and validated against CPU stock.

If anything, a little quicker than expected, and the rest of the machine seemed to be less 'laggy' while it was running than I experienced running Raistmer's modified build.

Still the exaggerated FLOP count though (or was v6.03 undercounting?), and no 'elapsed (wall) time' in stderr_out.

...
Still the exaggerated FLOP count though (or was v6.03 undercounting?), and no 'elapsed (wall) time' in stderr_out.

The count logic in 6.03 goes back to setiathome_enhanced 4.12, when there were some changes related to Gaussian fitting. Both Gaussian fitting and Triplet finding have data dependent aspects, and that's difficult to quantify.

The CUDA code doesn't do counting, that's handled by the CPU feeder part. For each call to a CUDA kernel, it tries to add about the same flop count as 6.03 and earlier.

In the case of triplets, the missing factor is an increment based on how many of the data bins were above threshold. That's not a huge amount, but adds up to a noticeable amount when there are millions of Triplet searches. There is a line in the code noting that factor has been left out because it's hard to estimate. I'd expect the effect to show as minor underclaims at VHAR and LAR.

For Gaussians, the situation is even more complex. The first step in setting up involves putting the data in a fixed length array, for 6.03 and earlier that's done by a copy function which doesn't make any flop count increment for many cases, but when bins need to be combined to reduce to the fixed length there are count increments. The equivalent line in 6.05 - 6.07 applies count increments for all cases, that could be easily changed. But 6.03 and earlier have two early exit possibilities in the analysis, both of which depend on the actual data being analyzed. The second of those was an optimization I submitted which was the primary change from 4.11 to 4.12. Both do quick checks on the data and bypass the full check if the data obviously has no Gaussian. Because of those, 6.03 and earlier only do full analysis on a fraction of the arrays. I think it's something like 3% to 5% based on the flop count curves I developed. The data gathered at many angle ranges for that included both a minimum amd maximum which each call to Gaussian fitting (and Triplet finding) might produce, then I chose how much of each to include by checking actual reported flopcounter values. Anyhow, the lines which add flop increments for Gaussian fitting could be scaled back to match. For now, I'd expect the overclaims to be largest toward the lower end of the range over which Gaussian fitting is performed.

Hmm, I didn't mean to steal this thread. I think the flop counting can be made to match much more closely without a lot of adjustment, it shouldn't become a major issue.

                                                                Joe

.... I'd expect the effect to show as minor underclaims at VHAR and LAR.

Yes, I saw that with v6.06, for VHAR at least.

v6.07 has run fast and true overnight. Very nice.

But unfortunately, we've run out of the nice mix of weird ARs we were running before - all I'm getting now is boring mid-range, Great for productivity, lousy for testing.

The other curiosity is that BoincView, which had learned to cope very well with Raistmer's modded v6.06, is now totally confused again - it thinks I have a 35 day cache of work on hand! Not true. Maybe the app_info (lack of?) makes a difference. Anyway, not Berkeley's problem and OT for this thread.

Have tried twice, even detached Beta, but I just keep gettng 6.03 tasks. I've aborted the tasks, but I'd like to run some 6.07.

Had no problem getting 6.05 tasks when they ran, so I'm not sure why I can't get past 6.03 now. driver 180.60 8500GT

Edit: Detached again, same thing just 6.03 tasks, now 3 of them.

Is it possbile for me to just make an app info file, and if so is there a way to download the 6.07 app.

I had the same yesterday,
got for my Vista 32 (Windows/x86) "only" a 6.03 MB application, reset of the project does not help, got the same app again.

Than the maintenance came, so I let it run.

Have crunched CUDA apps on this computer in the past with the same configuration
boinc version 6.5.0 (needed to run 2+1), graphic driver 179.28 for 8600M GS

---

Matthias

I am having same Problem Just MB 6.03.
Have tried resetting use to solve problem just received MB 6.03

Still I surpose it resolve it's self

Michael

Probably it'sserver-side change that limits CUDA app release only to high-performance GPUs - some attempt to get rid of VLAR issue.
Very inefficient attempt IMHO.
If you have low-range GPU (like 8500 for example) and can't recive 6.07 CUDA MB from server you can always return to modded app with app_info.xml - it should work at last.
But try to get stock 6.07 once more - maybe it's not server change but just temporary glitch ...

---

News about SETI opt app releases: https://twitter.com/Raistmer

Thank you for the imfo I have been running MB 6.06 up to Monday will have another try later today after running the MB 6.03 I have,

Was going to reset again but I am 40% into a AP 5.00

Michael

What GPU cards are you guys running?

Joe Segur posted this email at Lunatics:

From: David Anderson <davea@ssl.berkeley.edu>
Cc: boinc_alpha@ssl.berkeley.edu
To: BoincSpy Administrator <boinc_spy@telus.net>
Date: Sun, 11 Jan 2009 21:26:37 -0800
Subject: Re: [boinc_alpha] CUDA Blue Screen of death.
Message: 3

I've seen this also.
It's not related to driver version.
According to NVIDIA engineers,
if a GPU driver request doesn't complete in 2 seconds,
Windows assumes it's hung, and crashes.
Apparently the SETI@home/CUDA app does something that
sometimes takes > 2 secs on slow GPUs.

The NVIDIA people didn't have a fix or workaround for this.
So I'm going to change the server so that it won't send
CUDA jobs to GPUs slower than 60 GFLOPS
(I'm not sure this is the magic number,
but on my machine the GPU is 50 GFLOPS,
computed as clockRate * multiProcessorCount * 2,857).

-- David

I'm not sure exactly what figures to put into that formula, but I'm guessing Shader clock rate (usually the fastest figure) and number of shaders.

For my 9800GT, those are 1500MHz and 112 shaders - so the formula gives 48 GFLOPS, apparently below the threshhold.

Yet I'm continuing to get work issued, returned and validated - so maybe the server limit hasn't been implemented, or a lower cut-off used.

If they really did cut it at 60 GFLOPS (and if I've done the maths right), it would leave out an awful lot of crunchers.

I can't find a clock rating for the card that Joe says David uses - but it has the same number of shaders, and costs six times as much!

8500 GT, very low end card. I guess it's not going to work then, so I guess I'm stuck over at Main running the mod app there. 'll try again as soon as I clear the current cache I put on from Main, but I've already tried two different days with the same thing. I had a feeling it was the card, but I'm not ready to put a better card in the machine just yet, so I'm stuck with what I got.

It seems strange to me that even though my card will run the app that it's to be excluded, but that's life.

Having retired and only using it because the card is there. it's back to watching television again. I'll just run the MB 6.03 till something else comes along.


Michael

8500 GT, very low end card. I guess it's not going to work then, so I guess I'm stuck over at Main running the mod app there. 'll try again as soon as I clear the current cache I put on from Main, but I've already tried two different days with the same thing. I had a feeling it was the card, but I'm not ready to put a better card in the machine just yet, so I'm stuck with what I got.

It seems strange to me that even though my card will run the app that it's to be excluded, but that's life.

Try the mod app here first: it's only speculation that the server limit has been decreased, and even if it has, app_info.xml may foil it.

Edit - it would be interesting to see what messages BOINC got from the server when it first attempted the app download (say after a project reset), and what it says with an app_info.xml

8500 GT, very low end card. I guess it's not going to work then, so I guess I'm stuck over at Main running the mod app there. 'll try again as soon as I clear the current cache I put on from Main, but I've already tried two different days with the same thing. I had a feeling it was the card, but I'm not ready to put a better card in the machine just yet, so I'm stuck with what I got.

It seems strange to me that even though my card will run the app that it's to be excluded, but that's life.

Try the mod app here first: it's only speculation that the server limit has been decreased, and even if it has, app_info.xml may foil it.

Edit - it would be interesting to see what messages BOINC got from the server when it first attempted the app download (say after a project reset), and what it says with an app_info.xml

I put the app info file and mod app in and it deleted it and downloaded 6.03 again. I already know from a task at main that I still can't run VLAR but I was under the impression most can't.

Message from BOINC log
1/14/2009 2:56:16 AM|SETI@home Beta Test|Started download of setiathome_6.03_windows_intelx86.exe
1/14/2009 2:56:16 AM|SETI@home Beta Test|Started download of setigraphics_6.03_windows_intelx86.exe
1/14/2009 2:56:17 AM|SETI@home Beta Test|Finished download of setigraphics_6.03_windows_intelx86.exe
1/14/2009 2:56:17 AM|SETI@home Beta Test|Started download of libfftw3f-3-1-1a_upx.dll
1/14/2009 2:56:18 AM|SETI@home Beta Test|Finished download of setiathome_6.03_windows_intelx86.exe
1/14/2009 2:56:18 AM|SETI@home Beta Test|Started download of setiathome-6.03_AUTHORS
1/14/2009 2:56:19 AM|SETI@home Beta Test|Finished download of libfftw3f-3-1-1a_upx.dll
1/14/2009 2:56:19 AM|SETI@home Beta Test|Finished download of setiathome-6.03_AUTHORS
1/14/2009 2:56:19 AM|SETI@home Beta Test|Started download of setiathome-6.03_COPYING
1/14/2009 2:56:19 AM|SETI@home Beta Test|Started download of setiathome-6.03_COPYRIGHT
1/14/2009 2:56:20 AM|SETI@home Beta Test|Finished download of setiathome-6.03_COPYING
1/14/2009 2:56:20 AM|SETI@home Beta Test|Finished download of setiathome-6.03_COPYRIGHT
1/14/2009 2:56:20 AM|SETI@home Beta Test|Started download of setiathome-6.03_README
1/14/2009 2:56:20 AM|SETI@home Beta Test|Started download of seti_603.jpg
1/14/2009 2:56:21 AM|SETI@home Beta Test|Finished download of setiathome-6.03_README
1/14/2009 2:56:21 AM|SETI@home Beta Test|Finished download of seti_603.jpg
1/14/2009 2:56:21 AM|SETI@home Beta Test|Started download of 03no08aa.22996.8252.7.11.13
1/14/2009 2:56:23 AM|SETI@home Beta Test|Finished download of 03no08aa.22996.8252.7.11.13
1/14/2009 2:56:24 AM|SETI@home Beta Test|Starting 03no08aa.22996.8252.7.11.13_1

trying to download the app off the server gets the same thing.

I put the app info file and mod app in and it deleted it and downloaded 6.03 again.

Then you didn't do it right. If there's a recognised app_info.xml for the project ("[SETI@home Beta Test] Found app_info.xml; using anonymous platform" in BOINC startup messages), then BOINC should never delete apps listed in the app_info.xml, or download stock apps.

Don't worry about VLAR here - I haven't seen anything except mid-range since v6.07 came out. And the latest mod app autokills them anyway.

I put the app info file and mod app in and it deleted it and downloaded 6.03 again.

Then you didn't do it right. If there's a recognised app_info.xml for the project ("[SETI@home Beta Test] Found app_info.xml; using anonymous platform" in BOINC startup messages), then BOINC should never delete apps listed in the app_info.xml, or download stock apps.

Don't worry about VLAR here - I haven't seen anything except mid-range since v6.07 came out. And the latest mod app autokills them anyway.

I guess you're right, though I don't know what I did different this time. I do now have two tasks and the first one is running CUDA.

I wasn't really worried about VLAR tasks, they never did effect me the way they did most, they just would get stuck. I just mentioned it since the lower end cards maybe were excluded because of their inability to run VLAR and I was saying it's true mine still can't. Now if I could just get ahold of the real 6.07 beta app.

Thanks very much for the help Richard and Raistmer

Then (if our speculation is right) David Anderson has mis-described his own workround (or at least, described it ambiguously).

It isn't that the server won't send jobs (== Tasks?), it won't send the application - i.e. it won't automatically determine that the host is CUDA-capable. Shame there is no debug flag we can set to see the server's decision-making process.

I wonder why it decided to allow my card to opt-in?

Device exclusion to cover software bug is unacceptable as long-term solution IMO.and for short-term solution there are other ways exist too.

To all who can't get stock 6.07 here.
Please, try next solution and report if you success or not:

1) download my modded package.
2) take app_info.xml from it.
3) replace all MB_rev400mod_VLARkill_idle_CUDA.exe (for V5b) to setiathome_6.07_windows_intelx86__cuda.exe
4) download setiathome_6.07_windows_intelx86__cuda.exe from elsewhere (there was attached to one of Richard's posts on Lunatics for example - hope he could provide that link again)
5) put setiathome_6.07_windows_intelx86__cuda.exe along with edited app_info.xml (maybe it's worth to delete all AP-related stuff from app_info too) in SETI beta project folder
6) restart BOINC.

And post results (BOINC logs, what happened).

My 9600GSO has only 96 shaders and 1700MHz shader clock freq, but it worth 2 Q9450 cores !!! I will not allow to exclude it !

---

News about SETI opt app releases: https://twitter.com/Raistmer

I'll work up a complete downloadable Beta v6.07 package with app_info.xml and all required DLLs from my stock download - give me half an hour or so.

Am I really still the only person who has got it running?

I'll work up a complete downloadable Beta v6.07 package with app_info.xml and all required DLLs from my stock download - give me half an hour or so.

Am I really still the only person who has got it running?

Oh, fine, thanks!
I still finish pre-loaded cache with my V5b build, then will try to reset beta project and get stock 6.07. Need to get online VLAR crash with stock app to complete formal 6.07 testing :) (or, if some miracle possible - it will not crash although it crashed in standalone run? ... ;) ;) ;) )

---

News about SETI opt app releases: https://twitter.com/Raistmer