The "active thread percentage" is apparently a "time slice" for each GPU task. Both 40 and 70 work for what I am using it for. But 40% is faster for me.
It will run on Titan V's and rtx 3080 ti's for sure. And probably anything from the Titan generation on up.
I have to re-boot to disable the MPS server because I am not a Linux expert.
Respectfully,
Tom M
1. it's not a time slice (MPS might timeslice processes, but active thread percentage does not). it segregates off however much of the GPU to be used for each process. so if you set it to 40%, that means it's using 40% of the total SMs (probably some rounding happening to get it close to your setpoint) for each process.
2. MPS works on Kepler and up. however, the implementation from Volta and up is significantly better than the earlier version. you cannot run both on the same system. if you have both pascal and turing on the same host, you will be stuck with the earlier version of MPS even on the turing card.
3. to disable MPS without a reboot:
drop to a root shell with: sudo su
then: echo quit | nvidia-cuda-mps-control
if MPS is hung up, you might have to manually kill the MPS process with (using the process ID): sudo kill -9 <PID>
The most disappointing result from going from 3 Titan V's to 4 appears to be a larger number of slower running 1.14 (GPU only version) tasks.
Currently I am running the CPU at 50 percent via the boincmgr.
The 50 percent load was suggested by Ian&SteveC in conjunction with asteroids at home CPU tasks.
After I review the latest tasks timings to confirm my supposition and am going to constrain tasks at the project level while running the boincmgr at 90 percent.
I am running Rossetti at home not asteroids.
===edit====
The tasks are now averaging over 1,500 seconds. With 3 gpus they were averaging around 1,200 seconds.
===another edit====
When in doubt do a clean restart. Once you apply gpu power limits it apparently takes a reboot to fully clear the performance constraints.
Looking good. Adding CPU tasks back in.
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Tom M wrote: The "active
)
1. it's not a time slice (MPS might timeslice processes, but active thread percentage does not). it segregates off however much of the GPU to be used for each process. so if you set it to 40%, that means it's using 40% of the total SMs (probably some rounding happening to get it close to your setpoint) for each process.
2. MPS works on Kepler and up. however, the implementation from Volta and up is significantly better than the earlier version. you cannot run both on the same system. if you have both pascal and turing on the same host, you will be stuck with the earlier version of MPS even on the turing card.
3. to disable MPS without a reboot:
drop to a root shell with:
sudo suthen:
echo quit | nvidia-cuda-mps-controlif MPS is hung up, you might have to manually kill the MPS process with (using the process ID):
sudo kill -9 <PID>_________________________________________________________________________
Interesting. I am stuck on
)
Interesting. I am stuck on Pascal, so that will have to do unfortunately.
If I am running 3 tasks per GPU, it would make sense to me to set it to 33%?
taketwicedailey
)
I would experiment. I tried 33 percent and don't remember getting anything much different. 40 percent seemed to be faster in my case
But who knows your system might be faster at 70 percent. Run at least a day at any setting. Then take a look at the task times it settles on.
Tom M
###edit###
Make sure you have written record of your baseline 3x without MPS.
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
The most disappointing result
)
The most disappointing result from going from 3 Titan V's to 4 appears to be a larger number of slower running 1.14 (GPU only version) tasks.
Currently I am running the CPU at 50 percent via the boincmgr.
The 50 percent load was suggested by Ian&SteveC in conjunction with asteroids at home CPU tasks.
After I review the latest tasks timings to confirm my supposition and am going to constrain tasks at the project level while running the boincmgr at 90 percent.
I am running Rossetti at home not asteroids.
===edit====
The tasks are now averaging over 1,500 seconds. With 3 gpus they were averaging around 1,200 seconds.
===another edit====
When in doubt do a clean restart. Once you apply gpu power limits it apparently takes a reboot to fully clear the performance constraints.
Looking good. Adding CPU tasks back in.
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Tom M wrote: Once you apply
)
it doesn't. but you may have had something else hung up.
_________________________________________________________________________
Tom M wrote: The most
)
I have transitioned this system to 40%/v1.08 (aka: 1.07/1.15beta)/5x running on the NVIDIA MPS server.
This gives me 20 GPU threads. Yes, the fastest run times have slowed. But total production seems to be up.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
https://einsteinathome.org/co
)
https://einsteinathome.org/content/new-single-slot-gpus#comment-226415
Anyone want to drop some comments on him? He is wondering around in the wilderness of the Einstein Cafe.
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Hi. My brother just forced
)
Hi.
My brother just forced me to install Mint instead of Windows for our Mother and problems started.
Hardware: 1650super
System Linux Mint 21.3
Drivers 535 and 470
Problem:
Sudo nvidia-smi -pl 70 changes power limit (it gives proper message), but GPU always works with default power 100W
nvidia-smi shows 100/70W
I have edited Xwrapper.config as per message https://einsteinathome.org/pl/goto/comment/219716
Also enabled coolbits
I can control frequency offset and fans rpm so it works.
I'm noob. I can only follow forums hints, but on 2 other systems with UBUNTU other GPUs can reduce power limit without problem.
Can it be related to Mint?
Help please as 100W is to much
tito wrote: Hi. My brother
)
Two questions. Are you using 535 or 470?
And WHY did your brother force you to install Linux instead of Windows?
Btw, I like Ubuntu.
Just for giggles try to raise the PL to an unreasonable number. My result always comes back with the "allowed" power range.
I am away from my Ubuntu box. So I need to check my notes on successfully running the pl option.
Respectfully,
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Now 470, previously
)
Now 470, previously 535.
Range 70 - 120W
Why Mint instead Win? Don't Ask. So difficult to say :/