GPU/Bigadv... AGAIN...

[Norcalsteve]

SO,
From a other post (want to keep it organized to post topic... so i made a new one)... I have my GPU's affinity lock disabled, (YES) and i added an environmet variable to windows with the NV_FAH_CPU_AFFINITY and its value set at 15...

First. when my GPU clients were done, i ctrl-c' the clients and set the values. i also exited my VM ware Bigadv so i can restart my rig. Well... my dumba$$ hit the "up arrow" in terminal, after i got everything back up, and used pre-typed commands by scrolling back to them... I TOTALLY SCREWED THAT UP... I used my -smp 8 flag, instead of the new flag i had just started with, -smp 7... that lost my whole 10% of work!!!! ugh!!!! 5.5hrs or so wasted...

I was planning to change to -smp 8, but AFTER the -smp 7 run was done since i already stared it. But in my run of the settings... i notice a TOTAL PPD DECREASE with -smp 8 over -smp 7. I was getting 43000 PPD with the VM Bigadv set to -smp 7 (GPU clients running in Windows)... now, I am back to the -smp 8 flag, i am getting 38000-36000 PPD... and my Windows GPU clients went from a solid 9200-9500 PPD rate, to 5000-7500 ppd rate.

Please explain this? I am sure the Bonus will over take the PPD loss... right? so i should not worry?

 

[ChasR]

First issue is priority. THe VM, if you leave it open, will be running at normal priority on all cores and will steal all the cpu cycles the gpus need since they are running on the same cores. To fix this, add to the VMware Configuration file:

priority.grabbed="idle"
priority.ungrabbed="idle"

Now the VM will always be running lower priority than the GPUs, whether open and grabbed or minimized and ungrabbed, provided you set the GPUs to low priority in configuration (slightly higher in systray). Anyone running ATI GPUs may be better off using-smp 7. THe current nVidia drivers appear to use a lot more cpu cycles in Win 7 than they do in XP x64, probably reducing the effectiveness of my recommendation.

You also need to look at the WUs run on the GPUs, They are going to affect output of the SMP client when both SMP and GPU clients are run on all cores with GPU priority higher than SMP priority . If you have 1888 point WUs on the GPUs, the SMP client will run nearly full speed. When you have 353 pointers on the GPUs, the SMP client will slow by about 1 minute/frame, perhaps more.

 

[Norcalsteve]

i only had priority.ungrabbed="idle" in the VM client. Now i got both in. will report back if any issues.

Thanks again!