Intel GPUs Supported
log in

Advanced search

Message boards : News : Intel GPUs Supported

1 · 2 · 3 · Next
Author Message
Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 16574 - Posted: 24 May 2013, 20:01:27 UTC

The Collatz server was upgraded with the latest BOINC server software to resolve the work request issues for Intel GPUs and I am happy to report that Intel GPU work requests are now working properly. If you are experiencing any side effects from this deployment, please let me know ASAP.

zombie67 [MM]
Volunteer tester
Avatar
Send message
Joined: 3 Jul 09
Posts: 156
Credit: 612,749,810
RAC: 196
Message 16577 - Posted: 24 May 2013, 22:29:14 UTC

Perfect!

http://boinc.thesonntags.com/collatz/result.php?resultid=140825238

Looks like the run time is about 5.5 hours for the solo_collatz. Not bad for "free" credits!


____________
Dublin, California
Team: SETI.USA

Dylan
Send message
Joined: 16 Jul 12
Posts: 8
Credit: 12,262,824
RAC: 0
Message 16638 - Posted: 28 May 2013, 22:32:49 UTC
Last modified: 28 May 2013, 22:41:25 UTC

This is great. However, I have a question. How can I set up one of my computers to run Collatz tasks on the HD 4000, but also allow the GTX 670 to continue crunching another project (GPUGRID)?

Is it not easily done, due to the limitations of Boinc?

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 16640 - Posted: 28 May 2013, 23:25:18 UTC - in response to Message 16638.

This is great. However, I have a question. How can I set up one of my computers to run Collatz tasks on the HD 4000, but also allow the GTX 670 to continue crunching another project (GPUGRID)?

Is it not easily done, due to the limitations of Boinc?



See the "exclude_gpu" section of the cc_config.xml file at
http://boinc.berkeley.edu/wiki/Client_configuration

e.g.
<exclude_gpu>
<url>http://boinc.thesonntags.com/collatz/</url>
<device_num>0</device_num>
<type>intel_gpu</type>
<app>mini_collatz</app>
</exclude_gpu>
<exclude_gpu>
<url>http://boinc.thesonntags.com/collatz/</url>
<device_num>0</device_num>
<type>intel_gpu</type>
<app>solo_collatz</app>
</exclude_gpu>
<exclude_gpu>
<url>http://boinc.thesonntags.com/collatz/</url>
<device_num>0</device_num>
<type>intel_gpu</type>
<app>collatz</app>
</exclude_gpu>

It isn't efficient, but since BOINC doesn't use a DTD, I don't think you can repeat the app element within the exclude_gpu element. In other words, I'm not sure the following would work, but you could give it a try instead:

<exclude_gpu>
<url>http://boinc.thesonntags.com/collatz/</url>
<device_num>0</device_num>
<type>intel_gpu</type>
<app>collatz</app>
<app>mini_collatz</app>
<app>solo_collatz</app>
</exclude_gpu>

Also, it isn't documented well, but I believe the device number starts at 0 for the first device. Note that the device number is the device number for that type of GPU. With one Intel and one NVidia, they would each be device 0 but the type would be different.

Profile Beyond
Avatar
Send message
Joined: 30 Jul 09
Posts: 213
Credit: 939,653,252
RAC: 3,969
Message 16652 - Posted: 29 May 2013, 14:36:17 UTC - in response to Message 16638.

This is great. However, I have a question. How can I set up one of my computers to run Collatz tasks on the HD 4000, but also allow the GTX 670 to continue crunching another project (GPUGRID)?

I believe that you can just uncheck "Use NVIDIA GPU" in your project preferences from your account page. That's the way I do it to run my ATI/AMD GPUs here (or elsewhere) and the NVidia GPUs at GPUGrid.

Dylan
Send message
Joined: 16 Jul 12
Posts: 8
Credit: 12,262,824
RAC: 0
Message 16654 - Posted: 29 May 2013, 19:50:46 UTC - in response to Message 16652.
Last modified: 29 May 2013, 19:51:07 UTC

I just now tried this.

The good part: I wasn't sent any CPU or Nvidia/AMD GPU tasks.

The bad part: I wasn't sent any Intel GPU tasks.

This is what I get:

5/29/2013 3:38:38 PM | Collatz Conjecture | Scheduler request completed: got 0 new tasks

5/29/2013 3:38:38 PM | Collatz Conjecture | No tasks sent

5/29/2013 3:38:38 PM | Collatz Conjecture | Tasks for CPU are available, but your preferences are set to not accept them

5/29/2013 3:38:38 PM | Collatz Conjecture | Tasks for NVIDIA GPU are available, but your preferences are set to not accept them

5/29/2013 3:38:38 PM | Collatz Conjecture | Tasks for AMD/ATI GPU are available, but your preferences are set to not accept them


Also, when manually updating:


5/29/2013 3:38:54 PM | Collatz Conjecture | update requested by user

5/29/2013 3:38:59 PM | Collatz Conjecture | Sending scheduler request: Requested by user.

5/29/2013 3:38:59 PM | Collatz Conjecture | Not requesting tasks: project is not highest priority

5/29/2013 3:39:01 PM | Collatz Conjecture | Scheduler request completed

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 16655 - Posted: 29 May 2013, 23:01:09 UTC - in response to Message 16654.

I just now tried this.

The good part: I wasn't sent any CPU or Nvidia/AMD GPU tasks.

The bad part: I wasn't sent any Intel GPU tasks.

This is what I get:

5/29/2013 3:38:38 PM | Collatz Conjecture | Scheduler request completed: got 0 new tasks

5/29/2013 3:38:38 PM | Collatz Conjecture | No tasks sent

5/29/2013 3:38:38 PM | Collatz Conjecture | Tasks for CPU are available, but your preferences are set to not accept them

5/29/2013 3:38:38 PM | Collatz Conjecture | Tasks for NVIDIA GPU are available, but your preferences are set to not accept them

5/29/2013 3:38:38 PM | Collatz Conjecture | Tasks for AMD/ATI GPU are available, but your preferences are set to not accept them


Also, when manually updating:


5/29/2013 3:38:54 PM | Collatz Conjecture | update requested by user

5/29/2013 3:38:59 PM | Collatz Conjecture | Sending scheduler request: Requested by user.

5/29/2013 3:38:59 PM | Collatz Conjecture | Not requesting tasks: project is not highest priority

5/29/2013 3:39:01 PM | Collatz Conjecture | Scheduler request completed



Don't you love how they have improved the scheduler? LOL

First, enable sched_op_debug in the cc_config.xml file (http://boinc.berkeley.edu/wiki/Client_configuration) located in the c:\program data\boinc folder (if you didn't change the location during the install). That will give a lot more info.

Do you have solo_collatz selected as a project? That's the only one with the Intel GPU app.

Lastly, this is all somewhat experimental. I think Collatz is the first to implement Intel GPU apps so we are the lucky ones that get to work out all the kinks with the BOINC software. I know there are known issues when using GPUs from different manufactures and the fix isn't a simple one and would likely break several existing projects applications. (You can't always fix a poor design with patches. That's like putting a sports car into a school bus and expecting it to be able to drag race in under 8 seconds. Not going to happen. Of course, hind sight is 20/20 when evaluating designs.)

Profile Beyond
Avatar
Send message
Joined: 30 Jul 09
Posts: 213
Credit: 939,653,252
RAC: 3,969
Message 16663 - Posted: 30 May 2013, 17:01:16 UTC - in response to Message 16654.

The good part: I wasn't sent any CPU or Nvidia/AMD GPU tasks.
The bad part: I wasn't sent any Intel GPU tasks.

This is what I get:

5/29/2013 3:38:38 PM | Collatz Conjecture | Scheduler request completed: got 0 new tasks
5/29/2013 3:38:38 PM | Collatz Conjecture | No tasks sent
5/29/2013 3:38:38 PM | Collatz Conjecture | Tasks for CPU are available, but your preferences are set to not accept them
5/29/2013 3:38:38 PM | Collatz Conjecture | Tasks for NVIDIA GPU are available, but your preferences are set to not accept them
5/29/2013 3:38:38 PM | Collatz Conjecture | Tasks for AMD/ATI GPU are available, but your preferences are set to not accept them

Is "Use Intel GPU" checked? Also you might try checking the AMD/ATI box and see what happens.

Also, when manually updating:

5/29/2013 3:38:59 PM | Collatz Conjecture | Not requesting tasks: project is not highest priority

Try putting: <fetch_on_update>1</fetch_on_update> in your cc_config.xml and restart BOINC.

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 16673 - Posted: 3 Jun 2013, 14:46:06 UTC - in response to Message 16663.
Last modified: 3 Jun 2013, 14:47:04 UTC

Try putting: <fetch_on_update>1</fetch_on_update> in your cc_config.xml and restart BOINC.


One person suggests a change and the developers change the default BOINC behavior with little or no discussion. Then, the volunteers think something is wrong with the project because of it. I find that very frustrating. I'm also sure it won't be the last time it happens. Thanks for providing the work around so that the needs of the one no longer outweigh the needs of the many.

Profile Beyond
Avatar
Send message
Joined: 30 Jul 09
Posts: 213
Credit: 939,653,252
RAC: 3,969
Message 16677 - Posted: 3 Jun 2013, 15:48:31 UTC - in response to Message 16673.

Try putting: <fetch_on_update>1</fetch_on_update> in your cc_config.xml and restart BOINC.

One person suggests a change and the developers change the default BOINC behavior with little or no discussion. Then, the volunteers think something is wrong with the project because of it. I find that very frustrating. I'm also sure it won't be the last time it happens. Thanks for providing the work around so that the needs of the one no longer outweigh the needs of the many.

You're very welcome. This behavior (Not requesting tasks: project is not highest priority) was quite irritating and it took a reasonable amount of prodding to get the workaround. But it was provided and it helps. Hopefully the default situation is being improved in BOINC 7.1.x. The round robin work requests would get stuck on some misc project in 7.0.x and it would take forever for the situation to become resolved. So far work fetch in 7.1.3 seems to be working better. The above was not the only work fetch bug in the current "stable" 7.0.64 (there were many) and the bugs are still being rooted out.

ExtraTerrestrial Apes
Avatar
Send message
Joined: 22 Aug 09
Posts: 56
Credit: 262,397,889
RAC: 208,227
Message 16683 - Posted: 3 Jun 2013, 20:32:53 UTC

I see solo_collatz 4.05 Intel GPU claims 0.02 cores, yet it's using a full core of my i7 3770K with driver 9.18.10.3071 (should be the current one).

MrS

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 16684 - Posted: 3 Jun 2013, 21:02:50 UTC - in response to Message 16683.

I see solo_collatz 4.05 Intel GPU claims 0.02 cores, yet it's using a full core of my i7 3770K with driver 9.18.10.3071 (should be the current one).

MrS


That has been a problem lately, and not just with Intel. nVidia has the same screwed up code where the asynchronous calls aren't always non-blocking. About all you can do is experiment by using fewer threads, fewer items_per_kernel and fewer kernels_per_reduction. However, even with the minimum values, NVidia still uses 100% of a CPU with certain drivers.

As I do not have an Intel GPU, I need feedback from the volunteers in order to know what settings to use as defaults.

ExtraTerrestrial Apes
Avatar
Send message
Joined: 22 Aug 09
Posts: 56
Credit: 262,397,889
RAC: 208,227
Message 16687 - Posted: 3 Jun 2013, 21:59:50 UTC - in response to Message 16684.

For now I'm using an app_config to tell BOINC the app is actually using 1 core. It hence launches one CPU thread less, so the CPU is not as over-commited. Not sure if it affects performance or throughput. You could set "1 CPU + 1 GPU" on the server side.

Regarding actually eleminating that CPU usage: I'd be happy to help you testing, but would need further instructions.

MrS
____________
Scanning for our furry friends since Jan 2002

ExtraTerrestrial Apes
Avatar
Send message
Joined: 22 Aug 09
Posts: 56
Credit: 262,397,889
RAC: 208,227
Message 16702 - Posted: 4 Jun 2013, 18:03:32 UTC - in response to Message 16663.

Try putting: <fetch_on_update>1</fetch_on_update> in your cc_config.xml and restart BOINC.

Wow, thank you!!! This "not requesting work, project not highest priority" madness was driving me nuts!

MrS
____________
Scanning for our furry friends since Jan 2002

Profile mikey
Avatar
Send message
Joined: 11 Aug 09
Posts: 3242
Credit: 1,688,469,767
RAC: 6,020,339
Message 16713 - Posted: 5 Jun 2013, 11:00:49 UTC - in response to Message 16702.

Try putting: <fetch_on_update>1</fetch_on_update> in your cc_config.xml and restart BOINC.

Wow, thank you!!! This "not requesting work, project not highest priority" madness was driving me nuts!

MrS


ME TOO!!!

Profile Beyond
Avatar
Send message
Joined: 30 Jul 09
Posts: 213
Credit: 939,653,252
RAC: 3,969
Message 16718 - Posted: 5 Jun 2013, 13:22:28 UTC
Last modified: 5 Jun 2013, 13:33:04 UTC

You're very welcome. Sometimes it's beneficial to hang out on the mailing lists. I was there trying to get a fix for the long standing "work fetch with GPU exclusions and app_info" bug. Saw this "not requesting work, project not highest priority" problem and requested a fix from DA. He didn't want to change the default behavior so I asked if a cc_config option was possible. He came up with the "fetch on update" flag. BTW, he fixed the "work fetch with GPU exclusions and app_info" too. That was a tough one to track down and required major changes. So thanks David for both fixes!

BTW, someone else to thank is Jacob Klein, who's lately been doing more testing on BOINC bugs than anyone. I'd say he's been instrumental in helping to track down many of the problems in the 7.0.x series and one of the reasons why 7.1.x is looking so good at such an early stage. So thanks Jacob, and let's not forget Rom too!

ExtraTerrestrial Apes
Avatar
Send message
Joined: 22 Aug 09
Posts: 56
Credit: 262,397,889
RAC: 208,227
Message 16724 - Posted: 5 Jun 2013, 20:22:27 UTC - in response to Message 16684.
Last modified: 5 Jun 2013, 20:22:53 UTC

Slicker wrote:
That has been a problem lately, and not just with Intel. nVidia has the same screwed up code where the asynchronous calls aren't always non-blocking. About all you can do is experiment by using fewer threads, fewer items_per_kernel and fewer kernels_per_reduction. However, even with the minimum values, NVidia still uses 100% of a CPU with certain drivers.

As I do not have an Intel GPU, I need feedback from the volunteers in order to know what settings to use as defaults.

Could you build an executable for testing purposes which works locally on some small set of numbers and maybe takes the parameters mentioned above from an input file? This way we could quickly run various configurations and see if we can free up that core.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 16725 - Posted: 5 Jun 2013, 21:03:48 UTC - in response to Message 16724.

Slicker wrote:
That has been a problem lately, and not just with Intel. nVidia has the same screwed up code where the asynchronous calls aren't always non-blocking. About all you can do is experiment by using fewer threads, fewer items_per_kernel and fewer kernels_per_reduction. However, even with the minimum values, NVidia still uses 100% of a CPU with certain drivers.

As I do not have an Intel GPU, I need feedback from the volunteers in order to know what settings to use as defaults.

Could you build an executable for testing purposes which works locally on some small set of numbers and maybe takes the parameters mentioned above from an input file? This way we could quickly run various configurations and see if we can free up that core.

MrS


It already does, so here's "Collatz App Testing 101" :-)

The 4.xx Collatz applications use a configuration file named "collatz.config" that must be located in the same folder as the executable. (See the thread about optimizing Collatz 4.xx apps for the details on the config file.)


Create an input file named "in.txt" in the same folder and place the following in it:
2361183357323965081960
2361183357326112565608
2361183357325482218569
1568
1208877788665


The first two lines are the true input and are the starting and ending numbers to be checked. That's all the app actually uses. The 3rd-5th lines are the results you should see in stderr.txt when it finishes, mainly the best number, number of steps, and total steps.

I use the following commands in a batch file in order to run it so that it deletes all the temp files BOINC creates prior to running and the displays the output:

@echo off
del stderr.txt
del out.txt
del boinc_finish_called
del boinc_lockfile
del checkpoint.txt
del init_data.xml
solo_collatz_4.07_windows_intelx86__opencl_intel_gpu_100.exe --device 0
type stderr.txt
pause


or in Linux:
#!/bin/sh
rm stderr.txt
rm out.txt
rm boinc_finish_called
rm boinc_lockfile
rm checkpoint.txt
rm init_data.xml
./solo_collatz_4.07_windows_intelx86__opencl_intel_gpu_100.exe --device 0
cat stderr.txt | more



If you get a driver reset while testing, you will need to kill the application via task manager in Windows or via the kill command in Linux as it will remain hung.

Profile mikey
Avatar
Send message
Joined: 11 Aug 09
Posts: 3242
Credit: 1,688,469,767
RAC: 6,020,339
Message 16727 - Posted: 6 Jun 2013, 11:07:03 UTC - in response to Message 16718.

You're very welcome. Sometimes it's beneficial to hang out on the mailing lists. I was there trying to get a fix for the long standing "work fetch with GPU exclusions and app_info" bug. Saw this "not requesting work, project not highest priority" problem and requested a fix from DA. He didn't want to change the default behavior so I asked if a cc_config option was possible. He came up with the "fetch on update" flag. BTW, he fixed the "work fetch with GPU exclusions and app_info" too. That was a tough one to track down and required major changes. So thanks David for both fixes!

BTW, someone else to thank is Jacob Klein, who's lately been doing more testing on BOINC bugs than anyone. I'd say he's been instrumental in helping to track down many of the problems in the 7.0.x series and one of the reasons why 7.1.x is looking so good at such an early stage. So thanks Jacob, and let's not forget Rom too!


Years ago I used to talk to Rom, online, and always found him helpful!!

ExtraTerrestrial Apes
Avatar
Send message
Joined: 22 Aug 09
Posts: 56
Credit: 262,397,889
RAC: 208,227
Message 16737 - Posted: 6 Jun 2013, 19:01:42 UTC - in response to Message 16725.

Wow, that looks nice! However, I could not yet run it successfully:

- copied everything from the BOINC project directory into a temporary directory
- created the "in.txt" and put the contents in
- created "collatz.config" (still empty, just like the file with the executable name)
- started the executable

Here's what I get:

20:52:33 (2828): Can't open init data file - running in standalone mode
Collatz Conjecture v4.05 x86_64 for OpenCL 1.0 or later
Based on the AMD Brook+ kernels by Gipsel
GPU Type INTEL
WARNING: Requested OpenCL platform not found. Aborting.
20:52:33 (2828): called boinc_finish

OpenCL is working under BOINC, so I have no idea what's missing. I could surely edit the regular .config, and maybe even in the BOINC slot, but a stand-alone test seems much nicer.

MrS
____________
Scanning for our furry friends since Jan 2002

1 · 2 · 3 · Next
Post to thread

Message boards : News : Intel GPUs Supported


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.