Advanced search

Message boards : Graphics cards (GPUs) : Does Gpugrid support running multiple jobs on one GPU? How to config?

Author Message
ltgobis
Send message
Joined: 19 Dec 18
Posts: 5
Credit: 51,912,073
RAC: 0
Level
Thr
Scientific publications
watwat
Message 51262 - Posted: 11 Jan 2019 | 1:05:58 UTC

I would like to load and run more than one job on my 1080. I've read that some BOINC projects support this operation and can be configured in the app_config.xml file. Is this possible with Gpugrid?

PappaLitto
Send message
Joined: 21 Mar 16
Posts: 508
Credit: 4,607,207,341
RAC: 107,803
Level
Arg
Scientific publications
watwatwatwatwatwatwatwat
Message 51281 - Posted: 11 Jan 2019 | 16:29:42 UTC

If the goal is to maximize the utilization of your GPU, try SWAN_SYNC

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2265
Credit: 15,986,076,810
RAC: 2,649
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 51290 - Posted: 12 Jan 2019 | 10:50:59 UTC - in response to Message 51281.

If the goal is to maximize the utilization of your GPU, try SWAN_SYNC
Under Linux - I would like to add.

STFC9F22
Send message
Joined: 10 Nov 17
Posts: 7
Credit: 154,876,594
RAC: 0
Level
Ile
Scientific publications
watwat
Message 51330 - Posted: 15 Jan 2019 | 22:35:02 UTC

Hi Itgobis,

I'm not an expert and do not know whether GPUGrid supports this, but I suspect that running more than a single GPUGrid job on your graphics card might not increase the utilisation and might even cause the jobs to run less efficiently than had they been run sequentially.

From your profile I see that, like me, one of your machines is running Windows 10 Professional, albeit mine is Win 10 Home and I am running a GTX 1070. I was surprised to see that my Windows Task Manager typically reports the GPU load at less than 5% when running GPUGrid Tasks, however, on viewing greater detail ie Task Manager (More Detail) -> Performance -> GPU X (where GPU X is the relevant card) if you then click on the title of one of the GPU X charts in the panel to the right and select the Compute_0 option, in my case, a continous Compute_0 loading normally between 85% and 95% is reported when running GPUGrid tasks. At this level, I presume there is little headroom to run a second task effectively and, I am speculating here but, doing so might introduce an overhead in managing the two tasks?

Perhaps someone with expertise in graphics card loading could comment.

Aurum
Avatar
Send message
Joined: 12 Jul 17
Posts: 295
Credit: 10,242,918,856
RAC: 2,068
Level
Trp
Scientific publications
watwatwat
Message 51338 - Posted: 17 Jan 2019 | 2:56:28 UTC - in response to Message 51330.

Try using GPU-Z and click on the Sensors tab. Make its window as wide as your screen. Watch the GPU Load under different scenarios. I wish GPU-Z worked for Linux.

Aurum
Avatar
Send message
Joined: 12 Jul 17
Posts: 295
Credit: 10,242,918,856
RAC: 2,068
Level
Trp
Scientific publications
watwatwat
Message 51339 - Posted: 17 Jan 2019 | 3:00:33 UTC - in response to Message 51262.

This is an example of an app_config.xml file to put in your Project folder. Each GPU WU requires a logical CPU thread of its own:
<app_config>
<app>
<name>acemdlong</name>
<gpu_versions>
<cpu_usage>1.0</cpu_usage>
<gpu_usage>0.5</gpu_usage>
</gpu_versions>
<fraction_done_exact>1</fraction_done_exact>
</app>
<app>
<name>acemdshort</name>
<gpu_versions>
<cpu_usage>1.0</cpu_usage>
<gpu_usage>0.5</gpu_usage>
</gpu_versions>
<fraction_done_exact>1</fraction_done_exact>
</app>
<app>
<name>QC</name>
<max_concurrent>8</max_concurrent>
<fraction_done_exact>1</fraction_done_exact>
</app>
<app_version>
<app_name>QC</app_name>
<plan_class>mt</plan_class>
<avg_ncpus>4</avg_ncpus>
</app_version>
<app>
<name>QC_beta</name>
<max_concurrent>1</max_concurrent>
<fraction_done_exact>1</fraction_done_exact>
</app>
<app_version>
<app_name>QC_beta</app_name>
<plan_class>mt</plan_class>
<avg_ncpus>4</avg_ncpus>
</app_version>
</app_config>[/quote]

Aurum
Avatar
Send message
Joined: 12 Jul 17
Posts: 295
Credit: 10,242,918,856
RAC: 2,068
Level
Trp
Scientific publications
watwatwat
Message 51340 - Posted: 17 Jan 2019 | 3:01:49 UTC - in response to Message 51262.

This is an example of an app_config.xml file to put in your Project folder. Each GPU WU requires a logical CPU thread of its own:
<app_config>
<app>
<name>acemdlong</name>
<gpu_versions>
<cpu_usage>1.0</cpu_usage>
<gpu_usage>0.5</gpu_usage>
</gpu_versions>
<fraction_done_exact>1</fraction_done_exact>
</app>
<app>
<name>acemdshort</name>
<gpu_versions>
<cpu_usage>1.0</cpu_usage>
<gpu_usage>0.5</gpu_usage>
</gpu_versions>
<fraction_done_exact>1</fraction_done_exact>
</app>
<app>
<name>QC</name>
<max_concurrent>8</max_concurrent>
<fraction_done_exact>1</fraction_done_exact>
</app>
<app_version>
<app_name>QC</app_name>
<plan_class>mt</plan_class>
<avg_ncpus>4</avg_ncpus>
</app_version>
<app>
<name>QC_beta</name>
<max_concurrent>1</max_concurrent>
<fraction_done_exact>1</fraction_done_exact>
</app>
<app_version>
<app_name>QC_beta</app_name>
<plan_class>mt</plan_class>
<avg_ncpus>4</avg_ncpus>
</app_version>
</app_config>

Aurum
Avatar
Send message
Joined: 12 Jul 17
Posts: 295
Credit: 10,242,918,856
RAC: 2,068
Level
Trp
Scientific publications
watwatwat
Message 51341 - Posted: 17 Jan 2019 | 3:02:04 UTC - in response to Message 51262.

This is an example of an app_config.xml file to put in your Project folder. Each GPU WU requires a logical CPU thread of its own:
<app_config>
<app>
<name>acemdlong</name>
<gpu_versions>
<cpu_usage>1.0</cpu_usage>
<gpu_usage>0.5</gpu_usage>
</gpu_versions>
<fraction_done_exact>1</fraction_done_exact>
</app>
<app>
<name>acemdshort</name>
<gpu_versions>
<cpu_usage>1.0</cpu_usage>
<gpu_usage>0.5</gpu_usage>
</gpu_versions>
<fraction_done_exact>1</fraction_done_exact>
</app>
<app>
<name>QC</name>
<max_concurrent>8</max_concurrent>
<fraction_done_exact>1</fraction_done_exact>
</app>
<app_version>
<app_name>QC</app_name>
<plan_class>mt</plan_class>
<avg_ncpus>4</avg_ncpus>
</app_version>
<app>
<name>QC_beta</name>
<max_concurrent>1</max_concurrent>
<fraction_done_exact>1</fraction_done_exact>
</app>
<app_version>
<app_name>QC_beta</app_name>
<plan_class>mt</plan_class>
<avg_ncpus>4</avg_ncpus>
</app_version>
</app_config>

KAMasud
Send message
Joined: 27 Jul 11
Posts: 24
Credit: 161,881,111
RAC: 50,309
Level
Ile
Scientific publications
watwat
Message 51343 - Posted: 17 Jan 2019 | 7:10:29 UTC

What is the use of multiple tasks? My Afterburner is reporting 85% to 92% GPU utilization. Except in the case for short runs, my GPU utilization is still 65%.

Aurum
Avatar
Send message
Joined: 12 Jul 17
Posts: 295
Credit: 10,242,918,856
RAC: 2,068
Level
Trp
Scientific publications
watwatwat
Message 51346 - Posted: 17 Jan 2019 | 15:03:59 UTC - in response to Message 51343.

The idea is do you get more work done in the long run. For you is???

Time{WU1 in parallel with WU2} < Time{WU1 in series with WUs}

ltgobis
Send message
Joined: 19 Dec 18
Posts: 5
Credit: 51,912,073
RAC: 0
Level
Thr
Scientific publications
watwat
Message 51599 - Posted: 6 Mar 2019 | 23:32:39 UTC - in response to Message 51340.

I don't see a app_config.xml in my C:\ProgramData\BOINC directory. Is this file hidden?

Zalster
Avatar
Send message
Joined: 26 Feb 14
Posts: 211
Credit: 4,496,324,562
RAC: 61
Level
Arg
Scientific publications
watwatwatwatwatwatwatwat
Message 51600 - Posted: 6 Mar 2019 | 23:47:42 UTC - in response to Message 51599.

I don't see a app_config.xml in my C:\ProgramData\BOINC directory. Is this file hidden?


There isn't one. You need to construct one yourself if you want to do that. However, GPUGrid provides bonus for work units returned before 24 hours after being sent. By running more than 1 work unit at a time on your GPU, it's highly doubtful you will finish it in order to get the bonus. Some work units work your GPU hard enough with only 1 on them (example High Temps and throttling) All these need to be considered if you are planning on running more than 1 at a time.
____________

mmonnin
Send message
Joined: 2 Jul 16
Posts: 289
Credit: 1,154,516,551
RAC: 141,192
Level
Met
Scientific publications
watwatwatwatwat
Message 51612 - Posted: 8 Mar 2019 | 1:57:40 UTC

A GTX 970 can complete all tasks easily within 24 hours at 2x. A GTX 1080 is over double the performance. It'll be fine.

Azmodes
Send message
Joined: 7 Jan 17
Posts: 23
Credit: 1,226,212,102
RAC: 28,019
Level
Met
Scientific publications
watwatwat
Message 51615 - Posted: 9 Mar 2019 | 11:01:31 UTC

I used to play around with multiple tasks on a single card, since projects like MilkyWay or some of PG's sieving apps benefit nicely from it, but I never really conclusively figured out if it actually makes a worthwhile difference. Combined with the 24h-bonus and the fact that each task hogs an entire CPU thread due to SWAN_SYNC I decided to forego any potential minute gains and just concentrate on one task per card.

Post to thread

Message boards : Graphics cards (GPUs) : Does Gpugrid support running multiple jobs on one GPU? How to config?