POEM

Gilthanis

Gilthanis

[H]ard|DCer of the Year - 2014
Joined
Jan 29, 2006
Messages
8,723
I am adding this info under a POEM thread so that it is easier to find.

This is an app_config file info to run more than one work unit per GPU. This setting is for 2 work units and reserving 1 CPU thread to run them. It can be changed to run more work units but it is up to you to find what is the best setting for you.

<app_config>
<app>
<name>poemcl</name>
<gpu_versions>
<gpu_usage>0.50</gpu_usage>
<cpu_usage>1.0</cpu_usage>
</gpu_versions>
</app>
</app_config>


Edit: I created an account called HardOCPtest so that anyone whom wanted to test or contribute anonymously could do so.


&#8226;to attach a computer to your account without using the BOINC Manager. To do so, install BOINC, create a file named account_boinc.fzk.de_poem.xml in the BOINC data directory, and set its contents to: <account>
<master_url>http://boinc.fzk.de/poem/</master_url>
<authenticator>65158_90e69af3238e26ef858cf76d349447fd</authenticator>
</account>

teamdaily.php
 
Last edited:
Hi,

you're right, we also figured out that memory bandwidth is the bottleneck of our current GPU simulations.
For users which are benchmarking and tweaking configuration files, the current practice with reserving almost one complete core, may become counterproductive depending on the system they use. But the point is to keep the application running with default BOINC client configurations.
So we made this choice to prevent hanging GPU jobs on most clients.

However, this problem should be gone with our next release.

Also a Happy new Year to you!
Thomas
Click to expand...

http://boinc.fzk.de/poem/forum_thread.php?id=1103#10047
 
POEM++ 2.17 Released
The new POEM@HOME version is now live. Both CPU and OpenCL versions have been updated.
In combination with the release, we also started three new types of GPU work units. Special thanks go to Vladimir Tankovich, whose GPU code made this possible.
Important: If you configured your client to run multiple GPU jobs on a single ATI card, jobs may fail. We are still investigating in this case. If you get this error, please adjust your settings to only run one POEM job per GPU in the meantime.
Click to expand...

I may add a few GPU's on here since it is rare to have a large number of GPU work units available...
 
I have been getting lots of GPU work from them for a few days now. Nice change from before, when I was lucky to get GPU work more than once a week, or so.
 
So they finally fixed their OpenCL apps now? Excellent!
 
Is anyone using the app_config.xml file from the first post? I'm wondering if I'm limiting my GPU throughput by using a stock config where I have 32 cores running BOINC CPU tasks plus GPUs trying to run these OpenCL tasks.

When I was running GPUGrid, CPU usage per task was minimal. Looking at top with these OpenCL tasks, they appear to consume 100% CPU besides using the GPU.

I'm using NVidia driver 340.24. Is there another specific version I should use to help these OpenCL tasks?
 
I will have to look again later today, but OpenCL tasks tend to take more CPU than CUDA tasks in general. It would not surprise me if you were starving your GPU's. I typically set aside a full core/thread per GPU to prevent that.
 
I am using an app_config, brilong, for exactly the reason you describe. The Poem OpenCL app requires some CPU power when crunching on the GPU, with the amount of CPU needed dependent on the GPU architecture.

On my AMD GPU's, which run OpenCL more efficiently than Nvidia GPU's, I have app_config set up to use .5 GPU's and .5 CPU's per WU. This allows me to run two GPU tasks at once on my GPU's, while giving each task 1/2 a CPU core (using BOINCTasks allows me to see CPU usage on all tasks in real time, so I know that Poem tasks on my AMD GPU's use about 10-15% of a CPU core while running). On my Nvidia GPU's, Poem needs almost 100% of a CPU core per task. On those machines, I have app_config set up to use 1 GPU and 1 CPU per task/WU.

Try it out and see if you notice a difference in processing times. I have a feeling you will see Poem WU's complete much faster than they do currently for you.

Note that you do not need to shut down or restart the BOINC Client to make the new app_config active. Just use the "Read Config Files" option in the client once you create the file and save it in the Poem directory. It will then be read and take effect immediately.

Here is the exact app_config I am using on my AMD machines. For Nvidia, just change the 0.5's to 1's.

<app_config>
<app>
<name>poemcl</name>
<max_concurrent>4</max_concurrent>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.5</cpu_usage>
</gpu_versions>
</app>
</app_config>

*Note that I am also using the max_concurrent flag to limit my machines to 4 running tasks at once. This is optional and this line can be removed if you do not wish to use this flag.

Paste that into a plain text editor (Notepad) and save it in your Poem directory as app_config.xml. That's all there is to it. Happy crunching!

Update 1: I apologize if I tend to go into too much detail in my posts, or break things down too simply. I am an engineer, so am of the mindset that I must explain everything in detail to ensure accuracy. However, this can sometimes be interpreted as condescension from me, when that is the last thing on my mind. Please just know I am being thorough as much for the person I am responding to, as I am for other people who may read this thread and learn something from one of my posts. Much of the knowledge I've gained in both of my hobbies (Boinc and Home Audio/Theater) has come from good posts in various forums, so I am simply giving back a little.
 
Last edited:
I appreciate the detailed explanation and I don't think it's condescending at all. I've put the Nvidia tweak into place (using 1 in both spots) and using BOINCTasks I told it to re-read the config files. I don't believe this worked, so I also ran the following:
Code: 
boinccmd --read_cc_config

After running the above command, I found the following in stdoutdae.txt (which was not there after I tried with BOINCTasks):
Code: 
28-Jan-2015 06:59:27 [Poem@Home] Found app_config.xml

Just for clarity, I put the app_config.xml file in /var/lib/boinc-client/projects/boinc.fzk.de_poem.
 
I have found that sometimes the client doesn't take the config files until the client had been restarted under Windows. (even when doing as RFGuy suggested) It may also come down to which version of the client you are running too as they work bugs out of features.

As far as location of the file under Linux, you would have to check with someone else. All I know is that the file should be in the BOINC data directory under the project it was designed for (POEM in this case).
 
I have an issue with the app_config.xml I'm using for POEM. I have specified 1 in gpu_usage and cpu_usage, but BOINC is still running 63 CPU-intensive processes (Malaria) as well as two GPU tasks (POEM). Each GPU tasks wants 100% of a CPU. When I enable the app_config.xml and told BOINC to re-read configs, it paused one Malaria process. I assumed it would pause two Malaria processes, one for each GPU. Any ideas on tweaks I should make? Thanks.
 
Also, are the Malaria WU's running in High Priority mode by chance? They often do on my machines. If so, Boinc seems to ignore everything when WU's are in High Priority mode and won't stick to the settings in your app_config. This is the only time I've found when app_config won't work as expected.
 
RFGuy_KCCO said:
Also, are the Malaria WU's running in High Priority mode by chance? They often do on my machines. If so, Boinc seems to ignore everything when WU's are in High Priority mode and won't stick to the settings in your app_config. This is the only time I've found when app_config won't work as expected.
Click to expand...

And this is by design and not a bug...just to be clear.
 
Gilthanis said:
What was in your app_config.xml file?
Click to expand...

Code: 
<app_config>
<app>
<name>poemcl</name>
<max_concurrent>4</max_concurrent>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>
</app_config>

I confirmed my 32-core system is running 30 Malaria tasks, none of which are high priority. I also confirmed my 64-core systems are running 62 Malaria tasks, but I just patched and rebooted them (went to nvidia 3.40-29 driver). Before they were running 63 Malaria tasks. Load is still 67+ on one of them.

Code: 
 14:56:46 up  2:27,  1 user,  load average: 67.73, 66.36, 65.77

Another thing I've noticed is that since I implemented these changes, I've not seen CPU-only POEM tasks. I wonder if I changed something incorrectly at boincstats.com or on my clients directly.
 
It may be BOINC load balancing between Malariacontrol and POEM.

The load you speak of... is that the GPU load? And are you wanting to load it up more or is that too high or what is the question? If you want to load it up more to utilize more of the GPU, then change your settings to run more tasks and assign it more CPU cores accordingly...

For example: To get 2 work units on the GPU, change that setting to .5 and then change the CPU setting to 2. To get 4 work units on the GPU, change that setting to .25 and then change the CPU setting to 4. That will also reduce the number of CPU work units running accordingly.
 
If everything works as expected, the release will happen tomorrow.

I used to stop task creation only on the Test Server, but queued tasks will then still be sent out. As they are of no further use, I have decided to switch it completely off now.
Please cancel any left over work units in your client GUI.
Click to expand...

http://boinc.fzk.de/poem/forum_thread.php?id=1097&sort_style=6&start=150

I wanted to point that out in case anyone still had test work units from the test server...
 
Just saw this on their page:
Intel GPU Support
We are currently running a version of POEM++ for Intel GPUs on our Test Server. We only have a very limited amount of machines for testing this, so it would be great to find some volunteers with Intel GPU hardware to connect to the Test Project.

To do so, use the &#8220;Add project&#8221; wizard of your BOINC client and paste the Project URL &#8220;http://int-boinctest.int.kit.edu/poem/&#8221; into the corresponding text field, as our test server is not listed as an official BOINC project. We appreciate every help for testing, but please keep the following in mind:
- Test server results will not be used for scientific publications.
- Credits will not be listed in the BOINC cross-project statistics.
- Test server applications may crash your system, so be sure to save any valuable data beforehand.
Best practice is to return a few tasks, and immediately detach from the POEM@TEST project.



Thanks in advance,
Thomas 16 Jun 2015, 12:03:45 UTC · Comment
Click to expand...

Edit: Wow. It would help if I would read things first... I see MrrVlad had already posted this info a couple of weeks ago. Sorry for the spam... I hang my head in shame...
 
I just added it to two of the machines I'm burning in, but the project appears to be out of work.

I went ahead and started a team as well.
 
Last edited:
Last edited:
Yes, I'm aware of the team for POEM, I started that one as well many moons ago. :cool:

I'm talking about the POEM@TEST. I still like to see the right team for all the projects I've attached regardless of stats.
 
Looks like POEM@home has not being sending any more work and will be shut down. Server says zero task ready to send. It's a nice project, alternative to FAH that utilizes GPU power. Apparently the GPU speed advances so rapidly that there is no enough WUs, according to their website. This is kind of strange as FAH has pretty much unlimited WUs.

I didn't reach my 10M goal. Stopped at 8.66M point but happy to contribute to the research.

Our team finished at 18th place. Hopefully we will remain at that position when the cleaning up is finished.
 
Yes this was a nice project. Something to keep in mind though is that different research depts. have different goals and needs. FAH is looking into a hand full of things where POEM may be looking a specifics. For those that may come after the project closes:

The Future of POEM@HOME
Dear POEM@HOME community,

first of all we want to express our deepest gratitude for every contributor to our project. Last year we developed a GPU optimized version of SIMONA, which is used as part of POEM@HOME, which speeds up the calculations by various magnitudes. Due to this speedup it has been very hard for us to supply sufficient work to saturate our BOINC project.
Therefore after eight years running, we have so much computational power that we do not require your large and generous contributions to conduct our science anymore. At the same time other, smaller and bigger projects require more computational power than ever.

The GPU accelerated algorithm will be open-sourced so that other groups can use the efficient forcefield freely to conduct research on GPU clusters on their own (the timeframe for this will be until Christmas. The code is ready to be shipped and the paper is written, but needs to be peer-reviewed before).

We are sorry for the hiccups in the last month; we would have liked to prepare a cleaner and better run of the final tasks.

Thank you again for your amazing support and we hope that your contributions will help other BOINC projects from now on. We want to especially thank Vladimir Tankovich for his contributions to the GPU code.

Best regards,
The POEM@HOME admin team

P.S.: We will keep the forums and website open for some time and will gradually empty the queue and shutdown the server daemons after approximately 2 months. 4 Oct 2016, 12:48:08
Click to expand...
 
Beside FAH, gpugrid may be the one that I'll try to contribute next in the protein folding segment. Did I miss out any other folding projects? So many projects and not enough hardware (and money)
 
Do you mean currently or in the past? there have been a few others in the past that completed or moved all in house.
 
I meant current active GPU biomedical folding projects. Might give a shot on gpugrid if there is nothing else.
 
GPUGrid and Folding at Home are the only two I'm aware of at the moment.
 
That's a shame about this project ending. I was just thinking about hitting it hard again to try for 100 million points. Looks like I will have to be satisfied staying at my current ~50 million points.
 
Me too. I'm running gpugrid right now. I saw Gilthanis just passed 100M mark. Going for 1B?:D That's too far for me to see.
 
Since I don't have any high end GPU's...it will be quite a while before I hit that achievement.
 
You must log in or register to reply here.
Back
Top