x38g reports

[Jason G]

FYI,  I picked up and ran an AP on my GPU last night. That could have been the problem. Everything is back up to speed after rebooting but my MB tasks on GPU are running high priority. That should settle down after a couple run here soon. I also have a few APs waiting so we shall see what happens.

Something to watch: I have no idea if Raistmer included any boincApi fixes for modern Cuda 4.0 drivers, or if they'd be needed under OpenCL running as they most definitely are under Cuda.  One way to find out would be to repeatedly exit Boinc (shutting down the AP while in progress) [or just snooze/unsnooze etc)  & see if it triggers a sticky downclock or not.  If so, then you'll just have to slap Raistmer around a bit to fix it. [ I mean ask nicely...  ]

Jason

[perryjay]

I'm over in his thread too. Seems it tried to start two APs at once a bit ago. After 10 minutes I had to suspend the second one as it hadn't even started yet. It was strange because even though I changed the count to .5 I had not changed the number of iterations. The only thing I can think of is the fact it was guesstimating the TTC as over 1600 hours and it just over rid everything else. Once I finish this first AP things should drop drastically and I will see what happens.

[Claggy]

FYI,  I picked up and ran an AP on my GPU last night. That could have been the problem. Everything is back up to speed after rebooting but my MB tasks on GPU are running high priority. That should settle down after a couple run here soon. I also have a few APs waiting so we shall see what happens.

Something to watch: I have no idea if Raistmer included any boincApi fixes for modern Cuda 4.0 drivers, or if they'd be needed under OpenCL running as they most definitely are under Cuda.  One way to find out would be to repeatedly exit Boinc (shutting down the AP while in progress) [or just snooze/unsnooze etc)  & see if it triggers a sticky downclock or not.  If so, then you'll just have to slap Raistmer around a bit to fix it. [ I mean ask nicely...  ]

Jason

Raistmer hasn't spoken about doing any api changes on any of his apps for Cuda 4 drivers, his apps also consume large amounts of CPU time when running with Cuda 4 drivers (when running on their own),
if there are other apps running like CPU apps, they manage to claw back some of that CPU time, but i think the elapsed time of the OpenCL tasks suffers,

Claggy

[Jason G]

Raistmer hasn't spoken about doing any api changes on any of his apps for Cuda 4 drivers, his apps also consume large amounts of CPU time when running with Cuda 4 drivers (when running on their own),
if there are other apps running like CPU apps, they manage to claw back some of that CPU time, but i think the elapsed time of the OpenCL tasks suffers,

Hmm, Well most of the Cuda 4 driver issues with actual cuda code, so far, have turned out to be a consequence of the underlying OS/driver changes, and are remedied by changing boincApi to behave a bit more threadsafe [to get back some stability], and taking into consideration the underlying memory model changes [to get back some performance+ a bit extra in some cases]. 

 I'd expect eventually they'd apply to OpenCL as well, but didn't expect that yet.  Oh well, at least before it becomes a real problem for Raistmer on Ati/OpenCL, we should have most of the kinks ironed out & techniques improved on the Cuda side.

[Edit:] do newer Ati drivers 'appear' to be getting slower as well for old-school coding techniques ?

[Mike]

[Edit:] do newer Ati drivers 'appear' to be getting slower as well for old-school coding techniques ?

When the setup is working properly no.

But there are some cicumstances some GPUs dont work well on each setup.
Will ty that lout on my sons PC as soon i find some time.

[Claggy]

Raistmer hasn't spoken about doing any api changes on any of his apps for Cuda 4 drivers, his apps also consume large amounts of CPU time
[Edit:] do newer Ati drivers 'appear' to be getting slower as well for old-school coding techniques ?

Don't know, don't actively run any Native ATI CAL projects, i used to run Collatz, but don't anymore, i've found SDK_2.4 speeds up OpenCL work on my HD5770,
but Raistmer has found slowdowns/instability with Cat 11.6/SDK2.4 on his HD69** when using his OpenCL apps

Claggy

[Jason G]

OK, thanks both, they could be at some intermediate stage with those too, making things look 'weird'.   On the Cuda side I'll test/check a few things to do with the 'unified memory model' while ironing out some of the V7 code.

Jason

[perryjay]

Well, I've decided to give Claggy's idea a try. I've set mine to .51 for the APs and .49 for the MB CUDA tasks. I hope it works for me. Right now I'm running two MB tasks so it will have to wait until I get some more AP work.

[Claggy]

Raistmer hasn't spoken about doing any api changes on any of his apps for Cuda 4 drivers, his apps also consume large amounts of CPU time
[Edit:] do newer Ati drivers 'appear' to be getting slower as well for old-school coding techniques ?

I've grabbed a few Collatz_mini Wu's, with Cat 11.6/SD2.4 my first Wu's completed in 6 min 15 secs & 6 min 12secs, down from average of ~6min 31 secs which would have been done with Cat 10.xx over 6 months ago,

Claggy

[Slavac]

Wish I could be more help, I'm still dealing with downclocking issues.

[perryjay]

Starting the thread on NC should be of help. I've noticed a number of my wingmen with the 560Ti throwing out -9s and other bad results giving them an invalid or error. It would be good to see if it is something in the card itself causing it.

[perryjay]

Here's a strange one http://setiathome.berkeley.edu/workunit.php?wuid=771323155 Check out computer 5257703. He's showing he has two GTS460s but all his GPU work is coming up no GPUs found.

Okay, looked a little closer, he's running driver version 258.96 with the x32f app. I tried to PM him but we will see if he responds.

[Jason G]

Here's a strange one http://setiathome.berkeley.edu/workunit.php?wuid=771323155 Check out computer 5257703. He's showing he has two GTS460s but all his GPU work is coming up no GPUs found.

He probably installed Boinc as a service (protected application)

[perryjay]

That's possible too. I didn't think of that.

I hope it's just the driver. He's running an I7 and cranking out a lot of good work on it. With those two 460s running he'll take off like a house afire.   



another interesting one  http://setiathome.berkeley.edu/workunit.php?wuid=763277874  they tried three times to prove me wrong when I didn't match that first -9 and still refused to give me canonical.   


Another one inconclusive  http://setiathome.berkeley.edu/workunit.php?wuid=782642742  He's running a 470 with v38g. He found 20 spikes I found 12.  Looking at his work I'd say I'll win.

[Jason G]

another interesting one  http://setiathome.berkeley.edu/workunit.php?wuid=763277874  they tried three times to prove me wrong when I didn't match that first -9 and still refused to give me canonical.   

Nice, the Cuda23 ones really dropped the ball with that task, the one that processed successfully missed a reportable gaussian the CPU apps & yourself picked up.  That pretty much correlates with the chirp + gaussian weirdnesses we've been seeing, and need to investigate & understand more deeply.  It is logical that the canonical was chosen as the AKv8 one, as it's likely 'in between' the 6.03 & yours, so representative.  I also take it as a sign that x38g was moving in the right direction to make better matches to CPU apps on some types of results, but there is still work to do.

Another one inconclusive  http://setiathome.berkeley.edu/workunit.php?wuid=782642742  He's running a 470 with v38g. He found 20 spikes I found 12.  Looking at his work I'd say I'll win.

  He could have power or cooling issues or anything, hard to say.  Once some of the more urgent issues are solved, I'll start playing with nVidia APi to see if I can extract thermal / power / clock info .  Decent info along those lines could paint a clearer picture  in some cases, if some info could be printed to stderr.

Jason