x38g reports

[Pepi]

Hi Jason !
This 39e you posted is fastest app on my system. Can do 4 WU in parallel without any problems, but stuck when first of four is finished and new one need to start
In the other way, work with much less memory usage then any of previous releases. Now , as always I will crunch at least 100 WU to see how this app works.
Good work!!!

[Jason G]

This 39e you posted is fastest app on my system.

That actually surprises me, since I dialled back some things for diagnosis & refinement, and am not focussed on speed at this time.  Gradually refining/fixing things I suppose may help 'real-world', as opposed to laboratory bench, speed as well, so I'll keep that in mind as things go further forward. 

The memory footprint may yet still end up from a little bit, to considerably smaller.  I'm not sure at this stage.  4 at a time is getting pretty eager though 

Jason

[Pepi]

I don't know what are you doing, but you doing well ( whatever you do with this app)

[perryjay]

Going for a fourth  http://setiathome.berkeley.edu/workunit.php?wuid=766762437 

I agreed with another running x38g while a stock 6.03 found an extra gaussian.  The x38g was first, I was third. Shouldn't I have validated him?

But anyway, made it through the night with no problems at all to report.  As to the comment about this being the fastest app yet, could it just be that it seems to load faster and we don't have that snag near the end of the WU anymore. Those two give us about a minutes advantage right there at least.

[Josef W. Segur]

Going for a fourth  http://setiathome.berkeley.edu/workunit.php?wuid=766762437 

I agreed with another running x38g while a stock 6.03 found an extra gaussian.  The x38g was first, I was third. Shouldn't I have validated him?

Yes, x38g on a GTX 460 and x39d on a GTS 450 really ought to be so close that an inconclusive comparison is nearly impossible. IMO the tiny likelihood of one of the reported or "best" signals being at a critical level should be much rarer than necessary to explain the number of inconclusives that are happening even between stock and the x3[8|9] builds.

Edit: Attaching the WU for that particular case. I have no way of comparing x38g to x39e unless someone else tests. I could do a CPU test, but won't unless CUDA testing seems to indicate it's needed.
                                                                   Joe

[Jason G]

As far as I'm concerned, x39d & e are different on those particular cards to x38g, and it's those family of 'newer' cards that brought us into x39 diagnostic builds trying to locate a specific issue with those GPUs (& some others).

My current suspicions are along the lines that x38g & earlier builds, on certain cards & drivers ,can have some silent failures, that while not necessarily manifesting in obvious reportable count differences can certainly lead to differences in the best signals.   

With regard to the likelihood that some such hidden error exists, with x38g it's possible, while with x39d highly unlikely.  In other words while the computation codepaths are basically the same, the driver version & kernel reliability cross GPU is not, which is why we are running 'diagnostic' builds & not optimising for performance at this point.


Jason

[perryjay]

Hey boss, just wanted to let you know, Raistmer, Claggy and Ghost made me do it!!! They ganged up on me!   

Only kidding, but I am running Raistmer's new app for APs on NVidia GPUs. Ghost said it was running okay on his with two MBs running at a time so I guess I will find out if three at a time will work. Haven't got any work for it yet but I'll let you know how it goes.

[Jason G]

OK.  If it pinches all the CPU from the Cuda app, starving it out, blame Raistmer.

[Claggy]

Going for a fourth  http://setiathome.berkeley.edu/workunit.php?wuid=766762437 

I agreed with another running x38g while a stock 6.03 found an extra gaussian.  The x38g was first, I was third. Shouldn't I have validated him?

Yes, x38g on a GTX 460 and x39d on a GTS 450 really ought to be so close that an inconclusive comparison is nearly impossible. IMO the tiny likelihood of one of the reported or "best" signals being at a critical level should be much rarer than necessary to explain the number of inconclusives that are happening even between stock and the x3[8|9] builds.

Edit: Attaching the WU for that particular case. I have no way of comparing x38g to x39e unless someone else tests. I could do a CPU test, but won't unless CUDA testing seems to indicate it's needed.
                                                                   Joe

Here's a benchrun comparing x39e to x32f and x38g,

Edit: x39e was Weakly similar against x32f, but Strongly similar,  Q= 99.96% against x38g

Edit 2: did an x38d run too, x39d was Weakly similar against x32f, but Strongly similar,  Q= 99.97% against x38g

Claggy

[Jason G]

Looks like the chirp difference to me altered the best gaussian, as opposed to more recent x39 changes.

I'll run that one on AKv8b for a double precision CPU chirp reference comparison.

(Barring the mentioned reliability issues we're looking for, x38g & x39d/e should have matched one another in this case)

[Josef W. Segur]

...
(Barring the mentioned reliability issues we're looking for, x38g & x39d/e should have matched one another in this case)

Claggy's x38g and x39e results did agree on the best_gaussian (and everything else) so can't explain why Perryjay's result didn't get strongly similar against Phud's.

I expect the x32f best_gaussian (which was one of the reported gaussians) is more likely to match CPU results, simply because it has a considerable history of few inconclusives.
                                                                Joe

[Jason G]

I expect the x32f best_gaussian (which was one of the reported gaussians) is more likely to match CPU results, simply because it has a considerable history of few inconclusives.

Well,  testing that theory grabbing a AKv8b result to add to the collection (That's taking a while  ). 

As no 'direct'  Gaussian search modifications were made in x32f through 39e, I currently call the x38g chirp & some other kernels 'unstable' under some conditions on certain cards under as yet undetermined conditions. If it turns out something simpler then I'll be happy with that.

I haven't looked at the spikes' proximity to threshold, but given the known 6.03 limitations (which should show in my AKv8b result if a factor) then I think the 3-way circus on the live runs might go something like this:

x38g Vs 6.03 disagrees by spikes, with possible suspect chirp in x38g presenting effects
x39d Vs 6.03 disagrees by spikes
x38g Vs x39d, possible suspect x38g chirp (reliability)

So far we have seen mismatched gaussians between AKv8 & x32f, with a full length test task from your FG set, I'm putting forward that the accuracy of those in the x38g one is repaired to match CPU by the chirp, but that instability created an issue in the live result not seen under bench, and that the majority of the remaining disagreement comes from the spikes.

[Claggy]

I've taken out my GTX460 and fitted my 9800GTX+ and i'm in the process of running a bench comparing x39d and x39e against x32f and x38g,

Claggy

[Jason G]

OK, apart from the fix for the VRAM blowout, x39e is identical to x39d, so you could shorten your test by one build if you wanted, though I suppose the extra run couldn't hurt to see if remaining stability issues show up,  despite that none seem to under bench (the frustrating part  )

[Jason G]

------------
Running app : AK_v8b_win_x64_SSSE3x.exe -verb -nog
with WU     : 27fe11ac.12560.9065.8.10.100.wu
Started at  : 05:48:23.796
Ended at    : 07:24:09.576
   5745.740 secs Elapsed
   5179.405 secs CPU time
Result      : stored as ref for validation.
------------
Running app : Lunatics_x39e_win32_cuda32.exe -verb -nog
with WU     : 27fe11ac.12560.9065.8.10.100.wu
Started at  : 07:24:12.637
Ended at    : 07:30:50.101
    397.415 secs Elapsed
     50.638 secs CPU time
Speedup     : 99.02%
Ratio       : 102.28 x
ref-AK_v8b_win_x64_SSSE3x.exe-27fe11ac.12560.9065.8.10.100.wu.res:-
Result      : Strongly similar,  Q= 99.74%


Attaching bench & result files for manual comparisons.... [Done, analysing]