Welcome to MilkyWay@home

GPU in non-continuous operation.

Questions and Answers : Windows : GPU in non-continuous operation.
Message board moderation

To post messages, you must log in.

1 · 2 · Next

AuthorMessage
Profile OpusPocus

Send message
Joined: 31 Oct 16
Posts: 6
Credit: 8,519,176
RAC: 0
Message 74773 - Posted: 11 Dec 2022, 14:29:02 UTC
Last modified: 11 Dec 2022, 14:30:01 UTC

Hello,
first of all, I am French so apologize in advance for my English something bad!

I come to you to understand why my GPU is so well used.
I have an RX 6900Xt graphics card.
The processing of a worker instance takes about 25s to this card.
It doesn't seem to be getting work quickly enough from MilkyWay @home.

She usually has work for 10 minutes and sometimes does not work for 30-45 minutes. Have I forgotten any settings?

Thank you in advance for your answers. Wishing you a good day.
ID: 74773 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile mikey
Avatar

Send message
Joined: 8 May 09
Posts: 3333
Credit: 524,010,781
RAC: 3,159
Message 74777 - Posted: 12 Dec 2022, 11:43:06 UTC - in response to Message 74773.  
Last modified: 12 Dec 2022, 11:44:15 UTC

Hello,
first of all, I am French so apologize in advance for my English something bad!

I come to you to understand why my GPU is so well used.
I have an RX 6900Xt graphics card.
The processing of a worker instance takes about 25s to this card.
It doesn't seem to be getting work quickly enough from MilkyWay @home.

She usually has work for 10 minutes and sometimes does not work for 30-45 minutes. Have I forgotten any settings?

Thank you in advance for your answers. Wishing you a good day.


No it's a problem with a Server side setting a previous Admin changed that no one can now find since they left! The backoff to get new tasks is 10 minutes of not asking for any tasks at all then it should send you another batch of 300 tasks, what alot of people do with the really fast gpu's like you have is to setup a zero resource share project as a backup and crunch for it while MilkyWay is idle, then when MilkyWay gets the next batch of tasks your gpu will stop getting tasks from the zero resource share project and run MilkyWay tasks until they are all gone again.

BTW your English is MUCH better than my one year of high school French taken 50 years ago!!!
ID: 74777 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile OpusPocus

Send message
Joined: 31 Oct 16
Posts: 6
Credit: 8,519,176
RAC: 0
Message 74779 - Posted: 12 Dec 2022, 15:53:25 UTC - in response to Message 74777.  

Hey, thank you very much for all your information! :)
I will follow this to the letter and ask for more work, I actually had an average of 100 work spots from MilkyWay.

Have a good evening.
ID: 74779 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Link
Avatar

Send message
Joined: 19 Jul 10
Posts: 609
Credit: 19,236,357
RAC: 871
Message 74780 - Posted: 12 Dec 2022, 16:03:49 UTC

An alternative to backup project would be running more than one BOINC client, your card should be more efficient when running few tasks at the same time anyway. Probably (at least) 4 tasks at once should be OK, you'll have to try yourself what works best.
ID: 74780 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile mikey
Avatar

Send message
Joined: 8 May 09
Posts: 3333
Credit: 524,010,781
RAC: 3,159
Message 74782 - Posted: 12 Dec 2022, 17:10:48 UTC - in response to Message 74780.  

An alternative to backup project would be running more than one BOINC client, your card should be more efficient when running few tasks at the same time anyway. Probably (at least) 4 tasks at once should be OK, you'll have to try yourself what works best.


How would that work Link? If you only have one gpu are you saying that BOTH instances of Boinc would get work for the same gpu?
ID: 74782 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Link
Avatar

Send message
Joined: 19 Jul 10
Posts: 609
Credit: 19,236,357
RAC: 871
Message 74783 - Posted: 12 Dec 2022, 18:35:16 UTC - in response to Message 74782.  
Last modified: 12 Dec 2022, 18:55:03 UTC

If you only have one gpu are you saying that BOTH instances of Boinc would get work for the same gpu?

I think they should. Of course it's better to run more concurrent WUs from one BOINC client via app_config, but with the server side limits here at Milkyway, I don't see other possibility to keep high-end GPUs busy, at least not until they can fix that configuration bug (<request_delay> in scheduler reply must be equal or larger than the limit on the server, currently it's 91 seconds).

But that 10 minutes doesn not seem to require "10 minutes of not asking for any tasks at all", my last successful request was about 8 minutes after the previous one. Still, lots of "Scheduler request completed: got 0 new tasks" in the log, sometimes I don't get anything for 2-3 hours. Not an issue for my GTX 275, but surely for all those AMD DP-beasts.
ID: 74783 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile mikey
Avatar

Send message
Joined: 8 May 09
Posts: 3333
Credit: 524,010,781
RAC: 3,159
Message 74788 - Posted: 13 Dec 2022, 12:04:39 UTC - in response to Message 74783.  

If you only have one gpu are you saying that BOTH instances of Boinc would get work for the same gpu?

I think they should. Of course it's better to run more concurrent WUs from one BOINC client via app_config, but with the server side limits here at Milkyway, I don't see other possibility to keep high-end GPUs busy, at least not until they can fix that configuration bug (<request_delay> in scheduler reply must be equal or larger than the limit on the server, currently it's 91 seconds).

But that 10 minutes doesn not seem to require "10 minutes of not asking for any tasks at all", my last successful request was about 8 minutes after the previous one. Still, lots of "Scheduler request completed: got 0 new tasks" in the log, sometimes I don't get anything for 2-3 hours. Not an issue for my GTX 275, but surely for all those AMD DP-beasts.


I know Petri of the GpuUsersGroup Team did a workaround with their own version of Boinc but unless you are on their Team you can't get it, Keith Myers posted a few hints as to what they are doing someplace on these Forums but I don't remember where as like you none of my gpu's are fast enough to run thru 300 tasks before asking for new ones and I don't run multiple tasks on one gpu, at least right now

That's interesting that the 10 minutes is no longer in effect but the Server STILL can't keep some gpu's fed, I know an Admin years ago said they didn't want just a few gpu's getting all the tasks as if they were 'cheating' then it would ruin alot of work, ie more gpu's means less chance of that
ID: 74788 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Link
Avatar

Send message
Joined: 19 Jul 10
Posts: 609
Credit: 19,236,357
RAC: 871
Message 74791 - Posted: 13 Dec 2022, 15:21:40 UTC - in response to Message 74788.  
Last modified: 13 Dec 2022, 15:22:51 UTC

I know Petri of the GpuUsersGroup Team did a workaround with their own version of Boinc

IIRC that client was just to cheat around server side WU limits for a single client, running multiple instances is supported by the official client via cc_config or cmd parameter.
ID: 74791 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile mikey
Avatar

Send message
Joined: 8 May 09
Posts: 3333
Credit: 524,010,781
RAC: 3,159
Message 74794 - Posted: 13 Dec 2022, 23:10:27 UTC - in response to Message 74791.  

I know Petri of the GpuUsersGroup Team did a workaround with their own version of Boinc


IIRC that client was just to cheat around server side WU limits for a single client, running multiple instances is supported by the official client via cc_config or cmd parameter.


Yes that's true
ID: 74794 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
EDU Enthusiasts of Digital Uni...

Send message
Joined: 7 May 10
Posts: 6
Credit: 67,026,308
RAC: 0
Message 74800 - Posted: 15 Dec 2022, 14:28:05 UTC

Why wouldn't milkyway admins do any changes? Its for them in first place - and more power to them.
Proper fix would be to make scheduler give you new tasks when you report finished work...
(I'm quite sure other projects would help milky owners if they asked - if they don't know how.)

Another point of benefit is to consolidate their size... the tasks are too small. I think making a single task from 4 or 6 of them would make sense (while still being viable on older gpu's). Such a granularity not only affects performance of the client negatively but server/s are likely stressed more due to it too.

All kinds of 'hacks' on user level are bothersome...
ID: 74800 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile mikey
Avatar

Send message
Joined: 8 May 09
Posts: 3333
Credit: 524,010,781
RAC: 3,159
Message 74801 - Posted: 16 Dec 2022, 3:49:20 UTC - in response to Message 74800.  

Why wouldn't milkyway admins do any changes? Its for them in first place - and more power to them.
Proper fix would be to make scheduler give you new tasks when you report finished work...
(I'm quite sure other projects would help milky owners if they asked - if they don't know how.)

Another point of benefit is to consolidate their size... the tasks are too small. I think making a single task from 4 or 6 of them would make sense (while still being viable on older gpu's). Such a granularity not only affects performance of the client negatively but server/s are likely stressed more due to it too.

All kinds of 'hacks' on user level are bothersome...


Because as I said earlier a previous Admin made the change and either failed to write it down or they can't find their notes and the current Admins can't find where to change it back now, so until they upgrade to a brand new version on the Server side without importing all the tweaks etc that have been done over the years then it is what it is. Interestingly there is a Boinc Admin Support Group but no one from MilkyWay seems to be interested in contacting them for help or is waaay too busy as the Admins here are grad students with ALOT of stuff on their plate.
ID: 74801 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile OpusPocus

Send message
Joined: 31 Oct 16
Posts: 6
Credit: 8,519,176
RAC: 0
Message 74803 - Posted: 16 Dec 2022, 5:47:37 UTC - in response to Message 74783.  
Last modified: 16 Dec 2022, 5:53:38 UTC

Duplicate sorry, I won't say why.
ID: 74803 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile OpusPocus

Send message
Joined: 31 Oct 16
Posts: 6
Credit: 8,519,176
RAC: 0
Message 74804 - Posted: 16 Dec 2022, 5:51:50 UTC - in response to Message 74783.  

Hello, very interesting what you say!
How and where can I access the "app-config" you mention?
What is the next step?

I have already tried to launch another Boinc instance per virtual machine but it was inconclusive and the GPU was absent.


the idea of ​​using 2 instances for my GPU interests me because I note that I consume about 85w on a card that has a capacity of 430w "GPU CHip Power Draw"
Which is useful for continuing to play at odds, but at night the GPU could be used more.
ID: 74804 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Link
Avatar

Send message
Joined: 19 Jul 10
Posts: 609
Credit: 19,236,357
RAC: 871
Message 74805 - Posted: 16 Dec 2022, 10:43:13 UTC - in response to Message 74804.  

Try this, no VM needed. You might want to create the additional BOINC data dirs inside the original one and than move them out of there to get the permissions right.
ID: 74805 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile OpusPocus

Send message
Joined: 31 Oct 16
Posts: 6
Credit: 8,519,176
RAC: 0
Message 74807 - Posted: 16 Dec 2022, 14:39:37 UTC - in response to Message 74805.  

Thank you, I just looked, it seems to be a real mess, I already note that there are requested files that I don't have.

As updates have been made a while ago on Boinc, I would not be surprised if the procedure is once again different.
ID: 74807 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Link
Avatar

Send message
Joined: 19 Jul 10
Posts: 609
Credit: 19,236,357
RAC: 871
Message 74808 - Posted: 16 Dec 2022, 15:14:25 UTC - in response to Message 74807.  
Last modified: 16 Dec 2022, 15:17:45 UTC

Well, since you use Windows, you could simply install BOINC as many times as you need it (I suggest 4), everytime with different data dir (but same program directory). The last installation should be a reinstallation of your current data dir I think, but without the option to run at startup/logon. The batch file to start all instances should be still right.
ID: 74808 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile OpusPocus

Send message
Joined: 31 Oct 16
Posts: 6
Credit: 8,519,176
RAC: 0
Message 74809 - Posted: 16 Dec 2022, 17:03:34 UTC - in response to Message 74808.  

I will try when I have some time thank you very much!
ID: 74809 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
EDU Enthusiasts of Digital Uni...

Send message
Joined: 7 May 10
Posts: 6
Credit: 67,026,308
RAC: 0
Message 74810 - Posted: 16 Dec 2022, 18:23:15 UTC - in response to Message 74801.  

no one from MilkyWay seems to be interested in contacting them for help or is waaay too busy as the Admins here are grad students with ALOT of stuff on their plate.

Not very constructive i know but I think its just not that important for them. It works for whatever they are doing.

If admins/devs actually read some of this, consider following:
- whatever optimizations you had made on the server side are already out-of-date,
- newer version will be faster, and more conservative with resources
- help is there for those who ask. (but you have to ask.)
- always make a backup before you do anything. (disk clone / disk img / db dump / executable etc whatever works for you)
- since value is static of 300 tasks, it should be easy finding where its specified in code. (but making tasks longer/bigger would stop-gap fix it without modifying any server side scheduler code)
ID: 74810 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile mikey
Avatar

Send message
Joined: 8 May 09
Posts: 3333
Credit: 524,010,781
RAC: 3,159
Message 74823 - Posted: 16 Dec 2022, 23:04:42 UTC - in response to Message 74810.  
Last modified: 16 Dec 2022, 23:09:24 UTC

no one from MilkyWay seems to be interested in contacting them for help or is waaay too busy as the Admins here are grad students with ALOT of stuff on their plate.

Not very constructive i know but I think its just not that important for them. It works for whatever they are doing.

If admins/devs actually read some of this, consider following:
- whatever optimizations you had made on the server side are already out-of-date,
- newer version will be faster, and more conservative with resources
- help is there for those who ask. (but you have to ask.)
- always make a backup before you do anything. (disk clone / disk img / db dump / executable etc whatever works for you)
- since value is static of 300 tasks, it should be easy finding where its specified in code. (but making tasks longer/bigger would stop-gap fix it without modifying any server side scheduler code)


The Old Admin DID read these posts but I don't know if the new one has gotten that far up to speed yet

Also I was talking about tweaks to the Server Software for Milkyway and how THEY do the drive locations etc etc not the software. Each Project has to test the cpu and gpu drivers as they come out or hope they work with the settings they already have, most just work but sometimes they don't so if you ONLY crunch there's no need to be on the bleeding edge of driver updates.

The 300 task thing could be in over half a dozen different places and several Admins have looked so it could even be a combination of places that's the problem.
ID: 74823 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Keith Myers
Avatar

Send message
Joined: 24 Jan 11
Posts: 709
Credit: 548,887,510
RAC: 58,779
Message 74825 - Posted: 17 Dec 2022, 8:43:50 UTC - in response to Message 74804.  

Hello, very interesting what you say!
How and where can I access the "app-config" you mention?
What is the next step?

I have already tried to launch another Boinc instance per virtual machine but it was inconclusive and the GPU was absent.


the idea of ​​using 2 instances for my GPU interests me because I note that I consume about 85w on a card that has a capacity of 430w "GPU CHip Power Draw"
Which is useful for continuing to play at odds, but at night the GPU could be used more.

https://boinc.berkeley.edu/wiki/Client_configuration#Application_configuration
Has all the information for crafting a app_config.xml file for running multiple tasks per card.
ID: 74825 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
1 · 2 · Next

Questions and Answers : Windows : GPU in non-continuous operation.

©2024 Astroinformatics Group