SETI at HOME Enhanced V6.09 (CUDA 23)

[Jason G]

Attached is a grab of the the client state from my p4 (XP w/sp3).  It is a bit 'unusual' (different again to the installer), since I experimented to find that installing both plan classes was needed for installer migration purposes, but left the version numbers on that machine at 6.09 (since it doesn't matter & I use experimental builds that don't require rescheduling)

Hope that helps, (Sorry can't get to my email right now...Can email later if you need)

Jason

[Later:] I guess you have it now from download count, removing to save space... Let me know if you need something different.  I'll do my best to recreate any kindof scenario you might need.

[Even Later:] If you're having trouble locating a problem, After some thought I can walk through a scenario that *might* demonstrate one reasonably likely/common potential issue, that I could test our combined theories about how it *should* react on that same p4.  I've come to think the newest installers & the client state I gave might be red herrings in this situation, since the app infos have been somewhat 'robustified' to minimise the potential for task loss.  Stock 6.09 & Boinc without anonymous platform use, however, IIRC aren't likely to be so forgiving.

[Marius]

Hello Jason,

Thanks indeed Its a neat example where the reschedule ignores the 609/cuda23 application. In this particulary situation there are only 603 and 609/cuda workunits so i was not able to test all. If you have a combination with 603, 608/cuda, 609/cuda and 609/cuda23 that would be very nice, the more difficult the better and even multiple projects are welcome to see if it can crack that..

I think that as long people don't use the slider to move units from the cpu to the gpu everything should be just fine AFAICS,  VLAR's (and optional the VHAR's) will be moved to the cpu. The other way from cpu to gpu is kind of a problem because in this combination of 609/cuda and 609/cuda23 i don't know which cuda application is active and which one i may use (mayby i just need to remove the slider and stick with the basics of moving VLAR and VHAR to the cpu)

Is it btw also possible to have multiple 603 applications for the cpu?

Greetings,
Marius

[Jason G]

...The other way from cpu to gpu is kind of a problem ...

hehehe, well you sussed out my example scenario  .     The 6.03->6.09 problem is common, because some will attempt to install & run a shiny new cuda card (Very Excited!  )  on a full cache, get 'project has no work' and immediately through impatience attempt to shift CPU work over, with no 6.08/cuda matching app/version/planclass on the host.  So that's why I said 'not compatible with 6.09' which as you point out isn't strictly the complete truth, but something we need to look at.

Now the problems get multiplied a little with upcoming possible further plan classes and platforms/devices etc, and I feel recognition should be configurable for specific app/version/planclass, and the optimal rescheduing requirements may change with development in the near future.  Aware that you have mentioned being quite busy elsewhere, Are you amenable to providing source so that we can maintain (giving you svn access as well) a codebase such that we can consider integration with the installer & better assist users while you're not about ?

Cheers, Jason

[Edit:]  multiple 6.03 apps ? 'Naturally' fom the project at this time No, but Yes I've actually tried that myself by creating different plan classes.  I did that to compare different apps under live conditions.  Given that Boinc has been improved to schedule by resources, it is quite feasible to have, for example, mutithreeaded and single threaded variants.  Really a particular application executable should be distinguished by application name, version number and plan class.

[Questor]

Hello Jason,

Thanks indeed Its a neat example where the reschedule ignores the 609/cuda23 application. In this particulary situation there are only 603 and 609/cuda workunits so i was not able to test all. If you have a combination with 603, 608/cuda, 609/cuda and 609/cuda23 that would be very nice, the more difficult the better and even multiple projects are welcome to see if it can crack that..

I think that as long people don't use the slider to move units from the cpu to the gpu everything should be just fine AFAICS,  VLAR's (and optional the VHAR's) will be moved to the cpu. The other way from cpu to gpu is kind of a problem because in this combination of 609/cuda and 609/cuda23 i don't know which cuda application is active and which one i may use (mayby i just need to remove the slider and stick with the basics of moving VLAR and VHAR to the cpu)

Is it btw also possible to have multiple 603 applications for the cpu?

Greetings,
Marius

Hi Marius - glad to see you're around again!

One observation on removing the slider. Occasionally we get VLAR storms where I end up with nothing but VLAR tasks (I think the big crunchers who are still running VLAR killer apps probably dont help this situation as the tasks have got to end up somewhere!). In this situation, I would prefer not to move all VLAR tasks to CPU and leave the GPU idle but take the hit on slow processing. I am probably in the monority on this but I would like the ability to rebrand VLARs back to the GPU if it goes idle. The norm for me is to set the slider to the 'normal' running ratio of tasks. (However I reaslise the difficulty of extracting the right cuda information to construct a plan_class when moving from CPU to GPU).

With the new Fermi cards (only just got a couple of GTX470s so still experimanting) I havent yet worked out how 'bad' the hit is. Also just to make things more complicated the consensus I have seen so far is to introduce a new plan_class of cuda_fermi.

I discovered in testing that Reschedule didnt work for cuda_fermi plan class but still use it to check the current ratio of tasks - unfortunately I stupidly hit the Run button by mistake but only lost a few tasks when I think it rebranded them back to GPU with the wrong plan_class entry and they errored out when it couldn't find the app.



John.

[Marius]

With the new Fermi cards (only just got a couple of GTX470s so still experimanting) I havent yet worked out how 'bad' the hit is. Also just to make things more complicated the consensus I have seen so far is to introduce a new plan_class of cuda_fermi.

I discovered in testing that Reschedule didnt work for cuda_fermi plan class but still use it to check the current ratio of tasks - unfortunately I stupidly hit the Run button by mistake but only lost a few tasks when I think it rebranded them back to GPU with the wrong plan_class entry and they errored out when it couldn't find the app.

Why i'm i not suprised, another variation in the plan_class definitions . Could you please email me a client_state.xml example at marius66 at home dot nl and i'll have a look at it.

plan_class=cuda (608)
plan_class=cuda23 (609)
plan_class=cudafermi (6???)

Thanks in advance,
Marius

[Richard Haselgrove]

It is not yet certain what the new plan class will be - it hasn't been installed yet. The definitive answer will become apparent on the applications page in due course.

But I strongly suspect they will copy the Beta application plan. That would make the plan_class

cuda_fermi

as Questor wrote, and the app_version 610: everything else is exactly analogous to the earlier plan_classes.

[Questor]

It is not yet certain what the new plan class will be - it hasn't been installed yet. The definitive answer will become apparent on the applications page in due course.

But I strongly suspect they will copy the Beta application plan. That would make the plan_class

cuda_fermi

as Questor wrote, and the app_version 610: everything else is exactly analogous to the earlier plan_classes.

Marius,
Apologies for not replying to your post - I have been away on holiday for a week.

I have now emailed you a copy of my fermi client state file. However, the machine is using an app_info.xml so is only picking up the app entries from there. However, the fermi app has now been released to live setiathome so if it helps I can temporarily remove the opt apps and get a clean client state file using stock apps if it will help. It might have to wait until the current issues with the seti site are resolved as getting the apps and new wus will probaly be an issue.

John.