V10/11 of modified SETI MB CUDA + opt AP package for full multi-GPU+CPU use

[FalconFly]

I have one nagging question concerning the V10 mod :

Does the CUDA application read the proxy settings from the BOINC Manager preferences?

If not, that would explain why it persistently fails to download and process any Workunits for me.
(after a fresh install, I have (again) 6 AP WorkUnits running on the quad due to the recommended cc_config.xml and 0 CUDA WorkUnits on the idle GPUs, still receive the bogus server messages, Message from Server : No Astropulse WorkUnits available)

-- edit --
Changed Proxy to direct Internet connection but so far nothing changed.

[Josef W. Segur]

I have one nagging question concerning the V10 mod :

Does the CUDA application read the proxy settings from the BOINC Manager preferences?
...

The app does not have anything to do with work fetch, that's BOINC's job. You'll have to figure out why the server thinks you've disabled the other kinds of work. One thing I forgot when I suggested checking your project preferences is that there are 4 venues, the host might be using any one of those so you have to also confirm that setting on the host page or by looking at the BOINC messages on the host at startup.
                                                                               Joe

[FalconFly]

The Host is under venue "Home" as all my Systems.

Good news is : Partial success !

After completely uninstalling and deleting the entire Directories of the former installation, I went for a brand new install with BOINC 6.4.5 Win64, set run mode to "Suspended", then attached SETI as new Project.
After allowing it to download one AP WorkUnit, I quit the BOINC manager.

Then I installed the appropriate V10 Package, edited the Number_of_GPUs to 2 and inserted the cc_config.xml reflecting 6 CPUs (4 CPU + 2 GPU).

However, when I restarted BOINC with all in place, the CPU got fed with 5 AP WorkUnits, while the GPU only runs 1 WorkUnit.

Now I'd have to wait some 20 hours for a CPU to finish its AP job and hope that the 2nd GPU gets a workunit instead.
Worst case, once the non-existent 6th core is free, the CPU may get assigned Job #5 while the GPU remains only half loaded.

Is that normal and unavoidable ?

[FalconFly]

*arg*

My fears were correct, after 1 GPU quickly finishing its job, I now (again) have the Quad CPU running 6 Tasks while the GPU didn't get any work
The GPUs next chance to get assigned work is in.... approx 30 hours, as the QuadCore is now running extremely inefficient on 6 AP jobs.

This stuff is really killing me. I organized some 3000$ worth of GPUs to do some serious crunching but this is just costing me tons of time with no visible success.
I already wish I had build 4-5 Octa Core CPU systems or with max. 1 GPU added instead, at least those would be long crunching at full speed...

I don't get it how this works easy for others, for me this has been the biggest single pain in the arse I ever had with BOINC so far...



-----------------------------------------------------------------

Another attempt :
I created a Work venue, so I can experiment without affecting the already smooth running Hosts.

- set Preferences to run Mainbeam WorkUnits only

The Result for now looks good as far as I can tell, for the first time ever I got both GPUs running it seems :


The question remains, however, will it happen again that the CPU gets 6 applications loaded or will that modified cuda recognize the situation and keep both GPUs busy instead ?

[Jason G]

Hi FalconFly,
   I'm just curious if there's a particular reason you're preferring Boinc 6.4.5 over 6.6.20 (Sorry if you mentioned it already, but I haven't fully read the thread  )  Further, you mention changing the number of GPUs to 2 ? I'm not aware of such a setting! 

I would recommend the following configuration (after backing up your current config) :
- Boinc 6.6.20
- No cc_config n_cpus entry at all
- Install applications using hanfy dandy installer from the Beta Installer thread (Or manually if preferred, Worth a try since you have Tester access? )
- edit/insert flops values in app_info (If desired)
- [EDIT: I notice your installation is Win64, so Installer may or may not work, and you'd want to switch out the CPU AKv8 app with the 64 bit variety.... But still should get you going]

[Raistmer]

"team" packs (V10 included) became obsolete now. To simplify life for users who can't read whole this thread to stay tuned with latest  changes (not too latest already though, BOINC 6.6.20 exists as recommended version more than month already AFAIK) I lock this thread.
All my own support of "teamed" packs now ceased (until I get some example why "teamed" pack should be used instead of BOINC 6.6.20 native GPU scheduling at least)

Will remove V10 archives from first post also.

"teamed" packs were temporary solution and well served their aim during lifespan. But now this period is over.

What should be used instead:
1) BOINC 6.6.20
2) latest CUDA MB build (can be found in this thread and will be posted in new thread devoted optimized CUDA MultiBeam app).
3) appropriate app_info.xml
some examples can be found in this thread: http://setiathome.berkeley.edu/forum_thread.php?id=52589

2) and 3) could be replaced by using Jason's Lunatics installer (currently in beta stage, can be found in beta area of this forum:  http://lunatics.kwsn.net/installer-testing/index.0.html )

What should not be used
1) ncpus field in cc_config.xml
2) app_info.xml supplied with obsolete "team" packs.
3)  "teamed" modification of AK_v8 (was part of "team" packs).