An Einstein Schizoid Embolism?

Keith Myers
Keith Myers
Joined: 11 Feb 11
Posts: 4751
Credit: 17675793310
RAC: 5782397

Looked at your new 7.19 host

Looked at your new 7.19 host with the 3080 Ti.

Have you considered trying to run doubles or triples on the card?

I see a benefit running 2X on my 3080. Run_time per task ~ 110 seconds. 

My teammate Ian does them 3X on his 3080 Ti and averages 90 seconds per task.

Just a FYI comment.

Let us know whether the #4592 bug is squashed.

 

Gandolph1
Gandolph1
Joined: 20 Feb 05
Posts: 180
Credit: 389400076
RAC: 14023

At least when I look at the

At least when I look at the scheduler log I now see that it is no longer requesting more work!  Looking good so far...

 

 

Gandolph1
Gandolph1
Joined: 20 Feb 05
Posts: 180
Credit: 389400076
RAC: 14023

Keith Myers wrote: Looked at

Keith Myers wrote:

Looked at your new 7.19 host with the 3080 Ti.

Have you considered trying to run doubles or triples on the card?

I see a benefit running 2X on my 3080. Run_time per task ~ 110 seconds. 

My teammate Ian does them 3X on his 3080 Ti and averages 90 seconds per task.

Just a FYI comment.

Let us know whether the #4592 bug is squashed.

 

I'm still playing with it, I may try to run 2 per GPU, but last time I tried that it didn't improve throughput, in fact it slowed things down slightly.  I'd love to see 90 seconds a task, but I'm not sure running multiples is the answer.

 

 

Ian&Steve C.
Ian&Steve C.
Joined: 19 Jan 20
Posts: 3709
Credit: 34637790038
RAC: 42579209

90s per task is effective.

90s per task is effective. the runtime number will slow down when you run multiples, you just have to realize that running 2 tasks @ 180s is faster than running one at a time task at 100s :).

 

also if you tested this previously with 4000 series tasks, then yes you see no benefit. but with these 3000 series tasks, you will see a benefit since they use less less GPU utilization overall and have an extended toplist portion with low utilization from 90-100%. running multiples fills in the gaps and runs more efficiently.

_________________________________________________________________________

Gandolph1
Gandolph1
Joined: 20 Feb 05
Posts: 180
Credit: 389400076
RAC: 14023

Ian&Steve C. wrote: 90s per

Ian&Steve C. wrote:

90s per task is effective. the runtime number will slow down when you run multiples, you just have to realize that running 2 tasks @ 180s is faster than running one at a time task at 100s :).

 

also if you tested this previously with 4000 series tasks, then yes you see no benefit. but with these 3000 series tasks, you will see a benefit since they use less less GPU utilization overall and have an extended toplist portion with low utilization from 90-100%. running multiples fills in the gaps and runs more efficiently.

I will definitely give it a try again, unfortunately I will have to remember to add to the app_config file for my 2080ti machine because I MAY not want to run multiples on that machine.  Honestly though, now that I have a fully functional 3080ti Hybrid my next order of business is to rebuild this machine in a new case that will provide better airflow and reduced noise.

 

 

Ian&Steve C.
Ian&Steve C.
Joined: 19 Jan 20
Posts: 3709
Credit: 34637790038
RAC: 42579209

i run multiples on all my

i run multiples on all my nvidia systems now. 3x on all of them. my weakest GPU is an RTX 2080, and that works fine at 3x also.

 

definitely works well on the 2080Ti too.

 

when/if 4000-seies tasks come back, I'll switch back to 1x because those definitely do better with 1x.

_________________________________________________________________________

Gandolph1
Gandolph1
Joined: 20 Feb 05
Posts: 180
Credit: 389400076
RAC: 14023

Ian&Steve C. wrote: i run

Ian&Steve C. wrote:

i run multiples on all my nvidia systems now. 3x on all of them. my weakest GPU is an RTX 2080, and that works fine at 3x also.

 

definitely works well on the 2080Ti too.

 

when/if 4000-seies tasks come back, I'll switch back to 1x because those definitely do better with 1x.

Decided since it really doesn't take much effort I'd....

1: Upgrade the 2080TI to the pre-release software because it IS in fact downloading more tasks than it can do

2: Modify the App_config to only run singles on the 2080ti for now

3: Change the project preference to run doubles so the 3080TI can stretch it legs for a bit, at least for today.

So far all looks good....

 

 

Gandolph1
Gandolph1
Joined: 20 Feb 05
Posts: 180
Credit: 389400076
RAC: 14023

Another oddity: I made the

Another oddity:

I made the exact same changes to the 2080ti machine and it refuses to add any additional tasks to the GPU, just keeps plodding along one task at a time.  I stopped and restarted BOINC, didn't work, so I rebooted, still no joy.  I guess I'm going to let it run for a while to see if there's a buffer somewhere that I'm unaware of...

 

 

 

Ian&Steve C.
Ian&Steve C.
Joined: 19 Jan 20
Posts: 3709
Credit: 34637790038
RAC: 42579209

do you have a max_concurrent

do you have a max_concurrent or project_max_concurrent statement that's limiting it to one task?

_________________________________________________________________________

Gandolph1
Gandolph1
Joined: 20 Feb 05
Posts: 180
Credit: 389400076
RAC: 14023

The Max_Concurrent is

The project_max_concurrent is currently set to four on the 2080ti machine because it only has the I7 Quad Core CPU, on the other machine I set it to 10.  (Leaving 6 physical cores for other work. Like World of Warships, which runs quite well alongside BOINC. LOL)

 

Changed it to 8, no difference...

I seem to recall reading something about any changes to the xml file requiring a project reset.  IS that still true?  I hope not.

 

 

 

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.