Advanced search

Message boards : Wish list : Make GPU do different projects

Author Message
Aurum
Avatar
Send message
Joined: 12 Jul 17
Posts: 399
Credit: 13,024,100,382
RAC: 766,238
Level
Trp
Scientific publications
watwatwat
Message 51102 - Posted: 26 Dec 2018 | 19:52:38 UTC

Anyone know a way to make a GPU run two different projects at the same time???
E.g., run GPUgrid & Einstein on the same GPU. Or, GPUgrid & a different GPU project, in case I run out of Einstein it can run Milkyway.

PappaLitto
Send message
Joined: 21 Mar 16
Posts: 511
Credit: 4,617,042,755
RAC: 0
Level
Arg
Scientific publications
watwatwatwatwatwatwatwat
Message 51104 - Posted: 26 Dec 2018 | 20:54:18 UTC - in response to Message 51102.

Anyone know a way to make a GPU run two different projects at the same time???
E.g., run GPUgrid & Einstein on the same GPU. Or, GPUgrid & a different GPU project, in case I run out of Einstein it can run Milkyway.

With SYAN_SYNC enabled on linux, you should be at 100% with no need for another concurrent work unit.

The only way I have been able to have two WUs from two different projects at the same time is to have an app_config.xml file for each project saying 0.5 GPU per WU each

Zalster
Avatar
Send message
Joined: 26 Feb 14
Posts: 211
Credit: 4,496,324,562
RAC: 0
Level
Arg
Scientific publications
watwatwatwatwatwatwatwat
Message 51105 - Posted: 26 Dec 2018 | 21:01:00 UTC - in response to Message 51102.

Could probably do it with a app_config.xml and limit GPU to 0.5 with a full CPU then using a <max_concurrent> following each type of work unit to 1 so only 1 work unit runs on the card.

However, I would advise against it.

You risk causing errors by mixing different projects on the GPUs.

Best case is you slow down the processing on the card by more than 50%.
____________

Aurum
Avatar
Send message
Joined: 12 Jul 17
Posts: 399
Credit: 13,024,100,382
RAC: 766,238
Level
Trp
Scientific publications
watwatwat
Message 51106 - Posted: 26 Dec 2018 | 22:11:59 UTC - in response to Message 51104.

Hi PappaLitto, You answered another question I had. So I have SWAN_SYNC enabled on a Linux computer and you're saying there is no need to say this:

<app>
<name>acemdlong</name>
<gpu_versions>
<cpu_usage>1.0</cpu_usage>
<gpu_usage>0.5</gpu_usage>
</gpu_versions>
</app>

and I should only run one project per GPU and I don't even need an app_config.xml file??? I'll try it and see if I can tell the difference.

Now it's been running either two GPUgrid WUs per GPU or one GPUgrid WU with one Milkyway and I don't see any noticeable slowing or errors.

kksplace
Send message
Joined: 4 Mar 18
Posts: 53
Credit: 1,400,026,749
RAC: 3,603,159
Level
Met
Scientific publications
wat
Message 51107 - Posted: 26 Dec 2018 | 22:28:40 UTC - in response to Message 51106.
Last modified: 26 Dec 2018 | 22:31:03 UTC

I took a look at your computers to find one with a similar setup to one of mine. You have a host with 2 1080s and Linux Mint, similar to mine with Mint and a 1080, both with SWAN_SYNC (SPIN) enabled. I compared a recent set of two PABLO...MOR18 WUs against my most recent of the same type.

http://www.gpugrid.net//results.php?hostid=487152

Your times were about 37000 and 38000 seconds each. Mine was just under 17000. There is no way this is definitive with such a small sample and also not knowing what else was running on your GPUs at the same time. However, it points out that it might be worth your effort to check out only running one GPUGRID WU with SWAN_SYNC enabled and compare overall points/time.

For what-it-is-worth, I do run two MilkyWay WUs at at time -- still better results. I think it has to do with the fact that during a Milkyway run, there are 'down times' on the GPU (watching on a Windows 10 setup with Afterburner) which get filled up when running two with offset start times.

Aurum
Avatar
Send message
Joined: 12 Jul 17
Posts: 399
Credit: 13,024,100,382
RAC: 766,238
Level
Trp
Scientific publications
watwatwat
Message 51108 - Posted: 26 Dec 2018 | 22:42:38 UTC - in response to Message 51105.

I couldn't think of a way to make that work. Say a rig has two GPU cards. I get any and all of these possibilities at some point:
GG+GG and GG+GG
GG+MW and GG+GG
GG+MW and GG+MW
MW+MW and GG+GG
MW+MW and GG+MW
MW+MW and MW+MW

GG = GPUgrid WU
MW = Milkyway WU

I'd like: IF WU1=GG THEN WU2≠GG

PappaLitto
Send message
Joined: 21 Mar 16
Posts: 511
Credit: 4,617,042,755
RAC: 0
Level
Arg
Scientific publications
watwatwatwatwatwatwatwat
Message 51155 - Posted: 30 Dec 2018 | 22:28:30 UTC

I'm also curious to see if this is possible as I tried to do this a long time ago.

Aurum
Avatar
Send message
Joined: 12 Jul 17
Posts: 399
Credit: 13,024,100,382
RAC: 766,238
Level
Trp
Scientific publications
watwatwat
Message 52303 - Posted: 18 Jul 2019 | 16:43:36 UTC

I still wish BOINC allowed this:

<app d=0>
<name>acemdlong</name>
<gpu_versions>
<cpu_usage>1.0</cpu_usage>
<gpu_usage>0.5</gpu_usage>
</gpu_versions>
</app>
<app d=1>
<name>acemdlong</name>
<gpu_versions>
<cpu_usage>1.0</cpu_usage>
<gpu_usage>0.5</gpu_usage>
</gpu_versions>
</app>

____________

Aurum
Avatar
Send message
Joined: 12 Jul 17
Posts: 399
Credit: 13,024,100,382
RAC: 766,238
Level
Trp
Scientific publications
watwatwat
Message 52304 - Posted: 18 Jul 2019 | 16:50:52 UTC - in response to Message 51107.

I took a look at your computers to find one with a similar setup to one of mine. You have a host with 2 1080s and Linux Mint, similar to mine with Mint and a 1080, both with SWAN_SYNC (SPIN) enabled. I compared a recent set of two PABLO...MOR18 WUs against my most recent of the same type.

http://www.gpugrid.net//results.php?hostid=487152

Your times were about 37000 and 38000 seconds each. Mine was just under 17000. There is no way this is definitive with such a small sample and also not knowing what else was running on your GPUs at the same time. However, it points out that it might be worth your effort to check out only running one GPUGRID WU with SWAN_SYNC enabled and compare overall points/time.

For what-it-is-worth, I do run two MilkyWay WUs at at time -- still better results. I think it has to do with the fact that during a Milkyway run, there are 'down times' on the GPU (watching on a Windows 10 setup with Afterburner) which get filled up when running two with offset start times.
Most likely I was running two GG WUs at the same time. If so, then it implies that 2 x 17000 = 34000 and running one at a time is faster overall. I still think being able to run a GG WU on the same GPU as a MW WU is the best way to go. Works great with just one GPU in a computer.

____________

mmonnin
Send message
Joined: 2 Jul 16
Posts: 332
Credit: 3,772,896,065
RAC: 4,765,302
Level
Arg
Scientific publications
watwatwatwatwat
Message 52305 - Posted: 18 Jul 2019 | 16:53:27 UTC
Last modified: 18 Jul 2019 | 16:53:38 UTC

You're far from the 1st user to want some type of slot system for processors within BOINC kind of like FAH. I don't know where the suggestions are going but I made it here:
https://www.cpdn.org/cpdnboinc/forum_thread.php?id=8799

My guess it is really just the BOINC Mgr even though the client was mentioned.

Aurum
Avatar
Send message
Joined: 12 Jul 17
Posts: 399
Credit: 13,024,100,382
RAC: 766,238
Level
Trp
Scientific publications
watwatwat
Message 52306 - Posted: 18 Jul 2019 | 17:08:39 UTC - in response to Message 52305.

You're far from the 1st user to want some type of slot system for processors within BOINC kind of like FAH. I don't know where the suggestions are going but I made it here:
https://www.cpdn.org/cpdnboinc/forum_thread.php?id=8799

My guess it is really just the BOINC Mgr even though the client was mentioned.

I read some where that there's no dedicated BOINC developers any more. I don't know. BOINC probably has a forum that may be the most appropriate place to post my Wish.
I posted it here hoping that Richard Haselgrove will see it and pass it along to his BOINC buddies. He's the best BOINC-connected person I've come across in the various forums.
http://www.gpugrid.net/show_user.php?userid=30277
____________

Aurum
Avatar
Send message
Joined: 12 Jul 17
Posts: 399
Credit: 13,024,100,382
RAC: 766,238
Level
Trp
Scientific publications
watwatwat
Message 52307 - Posted: 18 Jul 2019 | 17:29:43 UTC

I also posted my Wish here:
https://boinc.berkeley.edu/forum_thread.php?id=13036#92218
____________

mmonnin
Send message
Joined: 2 Jul 16
Posts: 332
Credit: 3,772,896,065
RAC: 4,765,302
Level
Arg
Scientific publications
watwatwatwatwat
Message 52308 - Posted: 18 Jul 2019 | 21:10:37 UTC - in response to Message 52306.

You're far from the 1st user to want some type of slot system for processors within BOINC kind of like FAH. I don't know where the suggestions are going but I made it here:
https://www.cpdn.org/cpdnboinc/forum_thread.php?id=8799

My guess it is really just the BOINC Mgr even though the client was mentioned.

I read some where that there's no dedicated BOINC developers any more. I don't know. BOINC probably has a forum that may be the most appropriate place to post my Wish.
I posted it here hoping that Richard Haselgrove will see it and pass it along to his BOINC buddies. He's the best BOINC-connected person I've come across in the various forums.
http://www.gpugrid.net/show_user.php?userid=30277


I think David Anderson if anyone. But his funding is towards Science United side which is going the opposite direction of a more basic setup (pick a type of science and go) vs our more advanced client functions.

[VENETO] boboviz
Send message
Joined: 10 Sep 10
Posts: 158
Credit: 388,132
RAC: 0
Level

Scientific publications
wat
Message 52323 - Posted: 22 Jul 2019 | 10:15:07 UTC - in response to Message 52306.

I read some where that there's no dedicated BOINC developers any more. I don't know. BOINC probably has a forum that may be the most appropriate place to post my Wish.


Boinc has developers, but has a "community model".
Lates Boinc Workshop is clear on this transition (started 3ys ago).

david
Send message
Joined: 5 Jun 09
Posts: 5
Credit: 13,138,998
RAC: 0
Level
Pro
Scientific publications
wat
Message 54052 - Posted: 25 Mar 2020 | 10:51:10 UTC

Re multiple projects on one GPU. In linux, we can run nvtop and see how much memory is being used. So is there a method setting GPU memory usage per task? Even if true, are there multiple engines, like multiple cores on the CPU, where both WU's run literally simultaneously?

Post to thread

Message boards : Wish list : Make GPU do different projects

//