Guide to running multiple GPU work units

[t_ski]

The 7770's can handle six WU's if you split the CPU righ (and you have a strong enough CPU).

[HammerON]

Well the GTX 580's did not seem to like 4 WU's each (thanks KieX for the app_info). I kept getting a "Computation Error" on two of the 8. I reduced it to 3 WU's and it appears to be working just fine
Taking about 10-12 minutes to complete a WU....

[Norton]

Originally Posted by HammerON

Well the GTX 580's did not seem to like 4 WU's each (thanks KieX for the app_info). I kept getting a "Computation Error" on two of the 8. I reduced it to 3 WU's and it appears to be working just fine
Taking about 10-12 minutes to complete a WU....

That's the time it takes them to complete on my 7870... but I'm doing 8 of them on the one card. The difference between the 7 series AMD cards and anything else is amazing!!!

[mstenholm]

I do 2x4 (7770s, one at 1000 MHz, one at 1090) and get 83-84K when all is working and 6 on my 7950 (at 990 MHz I think it was) and got the same 83-85K when I was using the PC 8 hours a day. Lets see if it pick up in my absence.

[KieX]

Originally Posted by HammerON

Well the GTX 580's did not seem to like 4 WU's each (thanks KieX for the app_info). I kept getting a "Computation Error" on two of the 8. I reduced it to 3 WU's and it appears to be working just fine
Taking about 10-12 minutes to complete a WU....

Good stuff. Glad to see you get going again. I'll post your version on the OP. If it resolved your problem then might help others too.

Originally Posted by mstenholm

I do 2x4 (7770s, one at 1000 MHz, one at 1090) and get 83-84K when all is working and 6 on my 7950 (at 990 MHz I think it was) and got the same 83-85K when I was using the PC 8 hours a day. Lets see if it pick up in my absence.

The speed of your CPU and GPU play a big difference in what return you get with more WU.

-My old 7770's @ 900MHz with 2.1GHz CPU = 3WU optimal for 33K
-Your 7770's @ 1000MHz+ with (3.xGHz? maybe) CPU = 4 WU optimal for 40K?

But tell you what, the 79xx series are a different animal:
-7950 @ 1025MHz with a full 3GHz CPU thread = 11 WU give 95-100K
-7970 @ 1100MHz with 0.44% of a 4GHz CPU thread = 9 WU give 100K-115K (not run a full day yet, these are preliminary)

[F150_Raptor]

Originally Posted by HammerON

Well the GTX 580's did not seem to like 4 WU's each (thanks KieX for the app_info). I kept getting a "Computation Error" on two of the 8. I reduced it to 3 WU's and it appears to be working just fine
Taking about 10-12 minutes to complete a WU....

What was the time that your 580's would get when you were only running 1 wu per card? You could be doing less work with 3 wu's per card.

The reason I ask is when I ran them on my 580's with only 1 wu per card, the time was 2 minutes 40 seconds per wu. When I ran 2 wu's per card it went up to 4 minutes and 30 seconds, which is an average of 2 minutes 15 seconds per wu.

I never tried 3 wu's, but your average per wu is 3 minutes 20 seconds to 4 minutes. That average just seems kind of high per wu, that's why I asked what they'd run with only 1 wu per card. You might get more points only running 2 wu's.

[HammerON]

Originally Posted by F150_Raptor

What was the time that your 580's would get when you were only running 1 wu per card? You could be doing less work with 3 wu's per card.

The reason I ask is when I ran them on my 580's with only 1 wu per card, the time was 2 minutes 40 seconds per wu. When I ran 2 wu's per card it went up to 4 minutes and 30 seconds, which is an average of 2 minutes 15 seconds per wu.

I never tried 3 wu's, but your average per wu is 3 minutes 20 seconds to 4 minutes. That average just seems kind of high per wu, that's why I asked what they'd run with only 1 wu per card. You might get more points only running 2 wu's.

I think it was about 3 to 4 minutes per WU...
I am not going to worry about it now as they are getting replaced hopefully tomorrow by 2 7970's

[mstenholm]

Originally Posted by KieX

The speed of your CPU and GPU play a big difference in what return you get with more WU.

-My old 7770's @ 900MHz with 2.1GHz CPU = 3WU optimal for 33K
-Your 7770's @ 1000MHz+ with (3.xGHz? maybe) CPU = 4 WU optimal for 40K?

But tell you what, the 79xx series are a different animal:
-7950 @ 1025MHz with a full 3GHz CPU thread = 11 WU give 95-100K
-7970 @ 1100MHz with 0.44% of a 4GHz CPU thread = 9 WU give 100K-115K (not run a full day yet, these are preliminary)

I can see that I'm not running my 7950 to its full potential. I ran 6 becasue of CPU temperature but it is now -15 C outside so I asume that my office is a bit colder too. I will try with 8 (one CPU to one GPU) when I get home. I don't think my wife want to poke around in the app_info

Originally Posted by HammerON

I think it was about 3 to 4 minutes per WU...
I am not going to worry about it now as they are getting replaced hopefully tomorrow by 2 7970's

Wow, knocked down a place in the pie.

[t_ski]

Originally Posted by HammerON

I think it was about 3 to 4 minutes per WU...
I am not going to worry about it now as they are getting replaced hopefully tomorrow by 2 7970's

Just as I was making a run for the #10 spot

[[Ion]]

Originally Posted by HammerON

I think it was about 3 to 4 minutes per WU...
I am not going to worry about it now as they are getting replaced hopefully tomorrow by 2 7970's

That's a hell of an upgrade--particularly from the WCG standpoint. Between those and the HD7770 system, I think you can count on 300k PPD

[t_ski]

manofthem is running dual 7970's, so you'd get close to his PPD on this system only. If you're putting these on the 970 you can do 6 WU's each and still give an entire thread to each WU.

[RejZoR]

Any idea how i'd run multiple GPU tasks at once for Help Conquer Cancer (HCC) project and leave CPU for FightAIDS and Human Proteome? I'd like to run these tasks simultaneously because GPU churns out massive number of completed WU's (today when iw as testing it it did like 10 WU's in 5 minutes or so). But only for HCC as other two only run on CPU.

I'd like to run all this under a single device and single WCG account. I'll try to use only GPU in my AMD E-450 based laptop and run only HCC there.

[KieX]

Originally Posted by RejZoR

Any idea how i'd run multiple GPU tasks at once for Help Conquer Cancer (HCC) project and leave CPU for FightAIDS and Human Proteome? I'd like to run these tasks simultaneously because GPU churns out massive number of completed WU's (today when iw as testing it it did like 10 WU's in 5 minutes or so). But only for HCC as other two only run on CPU.

I'd like to run all this under a single device and single WCG account. I'll try to use only GPU in my AMD E-450 based laptop and run only HCC there.

I replied to the first part in your other thread, hopefully that gets you going the way you wanted.

[manofthem]

Originally Posted by t_ski

manofthem is running dual 7970's, so you'd get close to his PPD on this system only. If you're putting these on the 970 you can do 6 WU's each and still give an entire thread to each WU.

This is truth. 2 7970s running averaging ~200k/day
I wish I had more cores/threads

[KieX]

Right. So seems I'm really late to the ball with this one, but the days for the current app_info.xml method are numbered. WCG and BOINC have been working together on the issues with this multi-gpu method and we'll need to switch to a different method once the new software is stable (currenlty in BETA).

This is because a bug was fixed in the BOINC client. World Community Grid has never supported the anonymous platform mechanism. However, because of two bugs, one in the client and one in the server, volunteers were able to fetch work using the anonymous platform at WCG.

The bug that is fixed in the BOINC client is that before 7.0.40, if you were using the app_info.xml file & your client supported an 'alt-platform', then the client would send 'anonymous' as the primary platform but also report an 'alt-platform' (which it should not have been sent).

The issue on the server side is as follows. When the server receives a request, it adds the primary platform and any alt-platforms that the project supports to an 'allowed platform' list. If the first platform on the allowed platform list is the anonymous platform, then the server code handles the request using the anonymous platform code path. If the list size is 0 then a message is reported back to the client that the platforms reported are not supported.

The server bug is that if the primary platform is the anonymous platform, then any alt-platforms should not be considered. Instead, the server code added the alt-platforms to the list anyway.

This is an issue for us because since we do not support the anonymous platform, that is discarded during processing and not added to the allowed platform list. However, the alt-platforms are added to the list. Since the list size is > 0, processing continues down the normal server side logic. It is this that caused the generation of the 'ignored app version' message. It also meant that if I user only specified hcc1 in their app_info.xml file, but they had opted into receiving work for more than hcc1, then they would be assigned work from those other projects and the client would immediately discard that work (and not report that fact to the server).

After discussing with BOINC, the decision was that volunteers should be able to have fine grain control over how many instances of a job are run on their GPU and that the anonymous platform mechanism is not the way that should be done. The decision was made to create the app_config.xml file so that volunteers could have control over this by making changes to the file which has been added in the 7.0.40 client. Volunteers at World Community Grid should be looking as this is stabilized as the way to run multiple copies of HCC1 GPU on your graphics card at one time.

Once the changes for the app_config.xml file are in a stable BOINC client release, we will disable the anonymous platform on the server side. We will provide warning in this forum before we do that and we will allow time for volunteers to upgrade to the stable client build before we disable the feature.

Source: https://secure.worldcommunitygrid.or...ad?post=404466

Seems like a few people have tested the new method successfully so I'll read up and do some testing of my own. Will start preparing a guide for that method. Looks simpler, so it may even be possible to code a small program to let users automatically generate and create the file according to their preferences without all the current hassle.

We will be given notice of the change when it becomes official with a small grace period to make changes, so hopefully we can have our team transition smoothly and keep rocking that #4 spot

[Norton]

I loaded the 7.0.42(x64) Boinc Manager in the Xeon rig for this reason (think that one will do the new system).

This link is in the options:

http://boinc.berkeley.edu/wiki/Client_configuration

Decent info but I don't have enough experience to be able to do anything with it yet?

[agent00skid]

FOR 7.0.40 and later:

Place this in app_config.xml in the worldcommunitygrid folder, to make each hcc1 WU use ,25 gpu and 1 cpu each.

<app_config>
   <app>
      <name>hcc1</name>
      <max_concurrent>4</max_concurrent>
      <gpu_versions>
          <gpu_usage>.25</gpu_usage>
          <cpu_usage>1</cpu_usage>
      </gpu_versions>
   </app>
</app_config>

Concurrency is not really needed in this instance, but can be used to limit the number of WU from one project running at the same time.

Example:

<app_config>
   <app>
      <name>cep2</name>
      <max_concurrent>1</max_concurrent>
   </app>
</app_config>

For multiple things, multiple <app></app>'s are just placed inside the <app_config></app_config>

<app_config>
   <app>
      <name>hcc1</name>
      <max_concurrent>4</max_concurrent>
      <gpu_versions>
          <gpu_usage>.25</gpu_usage>
          <cpu_usage>1</cpu_usage>
      </gpu_versions>
   </app>
   <app>
      <name>cep2</name>
      <max_concurrent>1</max_concurrent>
   </app>
</app_config>

[Mindweaver]

Hey guys I've got my 1055T w/ HD7770 Up and running. What would be the best config to run? I only see on the OP has ATI CPU/GPU = intel/ati. I need to run AMD/ATI. Thanks!

[t_ski]

The 1055T us six cores, right? Probably the best config would be running six WU's on the card and no GPU, unless you want to run more projects that just HCC. However, Norton is the expert when it comes to AMD CPU's.

[m&m's]

Set .33 or .25 for your GPU and use the other cores to crunch HCC or GFAM.

[[Ion]]

I would encourage four WUs. I'm running that many on my HD7770s, and the benefit over three is already very small. Six seems to me like an awful lot for such a small card.

[johnspack]

Hmmm, there's a 64 bit version? Darn, dling it now.

[Norton]

Originally Posted by Mindweaver

Hey guys I've got my 1055T w/ HD7770 Up and running. What would be the best config to run? I only see on the OP has ATI CPU/GPU = intel/ati. I need to run AMD/ATI. Thanks!

Try this one:
*4 HCC gpu wu's (1.0 CPU + 0.25 GPU) and 2 HCC cpu wu's

click me

<app_info> 
	<app> 
		<name>hcc1</name> 
		<user_friendly_name>Help Conquer Cancer</user_friendly_name> 
	</app> 
	<file_info> 
		<name>wcg_hcc1_img_7.05_windows_intelx86__ati_hcc1</name> 
		<executable/> 
	</file_info> 
	<file_info> 
		<name>hcckernel.cl.7.05</name> 
		<executable/> 
	</file_info> 
	<file_info>
		<name>wcg_hcc1_img_7.05_windows_intelx86</name>
		<executable/>
   </file_info>
	<app_version> 
		<app_name>hcc1</app_name> 
		<version_num>705</version_num> 
		<platform>windows_intelx86</platform> 
		<plan_class>ati_hcc1</plan_class> 
		<avg_ncpus>1.0</avg_ncpus> 
		<max_ncpus>4.0</max_ncpus> 
			<coproc> 
				<type>ATI</type> 
				<count>.25</count> 
			</coproc> 
		<file_ref> 
			<file_name>wcg_hcc1_img_7.05_windows_intelx86__ati_hcc1</file_name> 
			<main_program/> 
		</file_ref> 
		<file_ref> 
			<file_name>hcckernel.cl.7.05</file_name> 
			<open_name>hcckernel.cl</open_name> 
		</file_ref> 
	</app_version>
	<app_version>
		<app_name>hcc1</app_name>
		<version_num>705</version_num>
		<platform>windows_intelx86</platform>
		<avg_ncpus>1.000000</avg_ncpus>
		<max_ncpus>1.000000</max_ncpus>
		<api_version>6.13.0</api_version>
		<file_ref>
			<file_name>wcg_hcc1_img_7.05_windows_intelx86</file_name>
			<main_program/>
		</file_ref>
	</app_version>
</app_info>

I've run into some minor scheduling glitches but it gives me 60-65k ppd on the FX 8350/7870 rig (6 gpu, 2 cpu)

[johnspack]

Originally Posted by HammerON

Well the GTX 580's did not seem to like 4 WU's each (thanks KieX for the app_info). I kept getting a "Computation Error" on two of the 8. I reduced it to 3 WU's and it appears to be working just fine
Taking about 10-12 minutes to complete a WU....

Yeah, I stick to 2 wus on my 480, about 6mins apiece.

[Norton]

Originally Posted by agent00skid

FOR 7.0.40 and later:

Place this in app_config.xml in the worldcommunitygrid folder, to make each hcc1 WU use ,25 gpu and 1 cpu each.

<app_config>
   <app>
      <name>hcc1</name>
      <max_concurrent>8</max_concurrent>
      <gpu_versions>
          <gpu_usage>.25</gpu_usage>
          <cpu_usage>1</cpu_usage>
      </gpu_versions>
   </app>
</app_config>

I switched my FX-8150/2x 7770 rig over to BOINC Manager 7.0.42(x64) and added the app_config to it. It started right up with 8 wu's (4 on each card)

I used the file above that agent00skid posted (All I changed was the concurrent # of wu's running from 4 to 8).

Thanks to KieX for the heads up and agent00skid for posting a draft app_config to try out!