Finally got a POEM@Home OpenCL unit! :) Was barely using GPU :(

[blckgrffn]

After upgrading my BOINC Client to 7.025, I re-attached to POEM for GPU only work yesterday.

This morning, checking on the cruncher in question, I see it has a .095 + 1 ATI GPU WU running. Yay, right?

Except that it had been running for 2+ hours and was 5% completed. Boo. I noticed my 4770 was running at "low" speed (250mhz) and was being hit for 5-40% utlization, which is underwhelming.

http://boinc.fzk.de/poem/forum_thread.php?id=708#5850

It looks like I'd need to do some configuration to reserve 2x CPUs and hit this guy (my 4770) with 2x GPU WUs in order for this to be worthwhile?

POEM++ OpenCL version - (Unsent) 1 - (In Progress) 12,517 - (AVG time in hours) 0.54 (0.21 - 1.74) - (Users) 211

http://boinc.fzk.de/poem/server_status.php

It looks like my experience may be a typical? And it looks like it would be hard to get the multiple WU's needed in order to fill up the GPU and take advantage of the CPUs. Until this gets a bit better sorted out, I think I'll remove Poem as an active project again.

Any one out there having better luck than I? Maybe it's my 4770?

 

[biodoc]

on linux with my 4850's I just freed up one core or one thread for the gpu on Poem. My cc_config.xml file was as follows for a Q6600:

<cc_config>
<options>
<ncpus>3</ncpus>
</options>
</cc_config>

 

[wayliff]

My experience was not with Poem but with MilkyWay@Home.

When all my CPU cores were assigned to something else, my GPU would only get used around 50% - 80%. Once I freed one CPU then the GPU was always 98%+. That made a ~30k points difference in one day.

This was combining MW@H for GPUs and WCG for CPUs.

 

[blckgrffn]

on linux with my 4850's I just freed up one core or one thread for the gpu on Poem. My cc_config.xml file was as follows for a Q6600:

<cc_config>
<options>
<ncpus>3</ncpus>
</options>
</cc_config>

Did that allow Poem to stretch its legs more efficiently? Or were you just on the ball and doing that preemptively?

My experience was not with Poem but with MilkyWay@Home.

When all my CPU cores were assigned to something else, my GPU would only get used around 50% - 80%. Once I freed one CPU then the GPU was always 98%+. That made a ~30k points difference in one day.

This was combining MW@H for GPUs and WCG for CPUs.

Luckily (?) enough, it seems like Collatz and MW are both able to use the GPU effectively without dedicating a core on this machine... I guess I've been doing mostly Collatz so I'll have to watch MW@H next time I have it running.

Thanks for the input, guys

 

[biodoc]

Did that allow Poem to stretch its legs more efficiently? Or were you just on the ball and doing that preemptively?

I believe I read that in the Poem forums but being stubborn, I tried w/ and w/o freeing up a core and saw a dramatic difference. I suspect the more powerful cards vs my 4850 would require freeing more than one core or using one of the more complicated cc_config.xml variants where you would need to load 2+ WUs for "co-crunching" on a single gpu.

Even freeing up a core on my Q6600 vs freeing up a thread on my 2700K with identical 4850s in both machine, the 2700K/4850 was 10-20% faster on the GPU WUs.

It's a good project and I will go back hopefully in the near future but I stopped because it didn't play well with sim1 as I recall.

 

[blckgrffn]

I believe I read that in the Poem forums but being stubborn, I tried w/ and w/o freeing up a core and saw a dramatic difference. I suspect the more powerful cards vs my 4850 would require freeing more than one core or using one of the more complicated cc_config.xml variants where you would need to load 2+ WUs for "co-crunching" on a single gpu.

Even freeing up a core on my Q6600 vs freeing up a thread on my 2700K with identical 4850s in both machine, the 2700K/4850 was 10-20% faster on the GPU WUs.

It's a good project and I will go back hopefully in the near future but I stopped because it didn't play well with sim1 as I recall.

Great, thanks for the info

I am running six core thubans for a reason, right?

I'll free up a core like you did and see what happens. There are only so many projects the old 4770 can participate in... I hope that Poem is one of them. If the 4850 can do it effectively, so should this card...

 

[salvorhardin]

Here is my thread where I put up the settings that worked for me. I haven't upgraded to anything higher than 12.1 so I don't know if there has been anymore performance increases. It was working for me as of 3 weeks ago. I switched from my 5850 to my 4830 but I couldn't get more than 2 wus at a time and only when I would update the project.

 

[blckgrffn]

Here is my thread where I put up the settings that worked for me. I haven't upgraded to anything higher than 12.1 so I don't know if there has been anymore performance increases. It was working for me as of 3 weeks ago. I switched from my 5850 to my 4830 but I couldn't get more than 2 wus at a time and only when I would update the project.

Thanks for that! I intend to give that a shot if I have sometime tomorrow.