POEM: Upcoming client release

[Slicker]

I am running a 5770 and if I run only Poem I get between 150 to 200K, The more CPU you give the more you have of GPU on it. Now I am giving 2 cores to 1 GPU and I am a bit over 60%. (60/70%) of the GPU crunching. I would like to have more cores.

I increased the CPU for each and that helped a lot, but it is still at 84% utilization. So.... 4.0 CPU cores to utilize 0.84 GPUs. Calling it a GPU app is a lie. A better description would be a multi-threaded CPU app that also requires a GPU.

[Duke of Buckingham]

I say that for 5770 the GPU would require about 4 cores at 2.8 GHz to get between 90 and 100% and maybe 6 cores for the 6970. And yes slicker you are so right about your complains but Poem is now the best paying project once we understood this all "equation". I am as you all know very bad at GPU crunching but I have been studying the best I can to understand the relations between CPU and GPU.

The ones that have spare CPU power to crunch old or not very strong GPUs have much to gain on this kind of project. I will in a small while see the influence of OC both CPU and GPU. I hope I don't fry both...

Go slicker and give them a lesson about multi-threaded CPU apps. I am to bad for that. Making the app and make it work at more higher performance was quite a task for me.

Very Bad At GPUs Crazy Duke.

[Al]

I increased the CPU for each and that helped a lot, but it is still at 84% utilization. So.... 4.0 CPU cores to utilize 0.84 GPUs. Calling it a GPU app is a lie. A better description would be a multi-threaded CPU app that also requires a GPU.

I agree, it uses too much CPU. At the same time, the Gpu wus give a lot better credit than the CPU only wus. I went from next to nothing to 1.3 million in a few days. I usually run MWay on my ATI cards, especially during challenges, poem's ATI app gives me another option for non-challenge times.

Sent from my ADR6300

[zombie67]

Does anyone have an app_info they could share, for running more than one GPU task at a time?

[Duke of Buckingham]

I don't know very well how to do this but I will try to send the XML

<?xml version="1.0" encoding="utf-8"?>
<app_info>
<app>
<name>poemcl</name>
<user_friendly_name>POEM++ OpenCL version</user_friendly_name>
</app>
<file_info>
<name>poemcl_0.1_windows_intelx86__opencl_ati_10 0</name>
<executable />
</file_info>
<app_version>
<app_name>poemcl</app_name>
<version_num>1</version_num>
<avg_ncpus>0.666667</avg_ncpus>
<max_ncpus>0.666667</max_ncpus>
<flops>2.5e11</flops>
<plan_class>opencl_ati_100</plan_class>
<file_ref>
<file_name>poemcl_0.1_windows_intelx86__opencl_ati _100</file_name>
<main_program />
</file_ref>
<coproc>
<type>ATI</type>
<count>0.333333</count>
</coproc>
<gpu_ram>256.000000</gpu_ram>
</app_version>
</app_info>

I don't know better how to send this file, I am sorry. Better as a notepad file

[Slicker]

I don't know very well how to do this but I will try to send the XML

<?xml version="1.0" encoding="utf-8"?>
<app_info>
<app>
<name>poemcl</name>
<user_friendly_name>POEM++ OpenCL version</user_friendly_name>
</app>
<file_info>
<name>poemcl_0.1_windows_intelx86__opencl_ati_10 0</name>
<executable />
</file_info>
<app_version>
<app_name>poemcl</app_name>
<version_num>1</version_num>
<avg_ncpus>0.666667</avg_ncpus>
<max_ncpus>0.666667</max_ncpus>
<flops>2.5e11</flops>
<plan_class>opencl_ati_100</plan_class>
<file_ref>
<file_name>poemcl_0.1_windows_intelx86__opencl_ati _100</file_name>
<main_program />
</file_ref>
<coproc>
<type>ATI</type>
<count>0.333333</count>
</coproc>
<gpu_ram>256.000000</gpu_ram>
</app_version>
</app_info>

I don't know better how to send this file, I am sorry. Better as a notepad file

I'd use <max_ncpus>1.0</max_ncpus>. That's how much BOINC will allow, and there are times when it does want a full CPU for a split second.

Depending upon the speed of the CPU, 0.66667 works for <avg_ncpus> but you may be able to set it to 0.5 on a faster one.

I have <count> set to 0.166666667 so I can run 6 at a time which is only 88% utilization on the HD 6970. I may need to switch to 0.125 and go with 8 at a time using 4 cores to run them.

[Mr. Hankey]

So here is a question for you. I have two GPUs (5850) in a box with a dual core E5300 @ 2.6, what configuration would work best? I used the stock app and did 1 task / gpu, I then put in an app-info and did 2 wu / gpu, but it sounds like I don't have enough cpu to push WU to GPU. Here are the tasks I have completed so far:

http://boinc.fzk.de/poem/results.php...state=3&appid=

I just switched it to 3 wus / gpu to let that run for a while.

[Bikermatt]

Mr. H, check out this thread, good info on optimization.

http://www.setiusa.us/showthread.php...ng-POEM-OpenCL