MW WUs not using all of GPU power?

[DrPop]

It's been a while since I crunched any MW. Kat's rig is back online now and she wants to crunch MilkyWay, of course. So...I fired it up last night, and realized this morning that the WUs are taking way longer than I remember them taking. She's got the 2GB version of Sapphire's HD4870 clocked at 790MHz...but I can't figure something out. It's only loading the GPU at like 70%. It is crunching Collatz, FreeHal and WuProp on the CPU if that matters. Also, if I suspend MW and let Collatz crunch on the GPU, Collatz will load the GPU to 98 or 99% like it should.

Any ideas of what's going on with MW here and why it won't load up the GPU higher?

[trigggl]

In Linux with a 4870, most of my tasks are completing in just under 3 minutes. What are you seeing?

[DAD]

Open cl sucks and is inefficient ?

Seriously, I dunno I don't use ati. But it could be the mw open cl code isn't optimized well

[shralper]

I just tried Milky Way on my machine. It seems Milky Way is only using around 80% of my 560 and 570 while prime grid uses 99% of both of them. So something MW is doing...?

[DrPop]

This is weird. Did MW re-code and cave in to OpenCL then? It used to be "native" ATI code, and was very efficient. I could get loads of 98 - 99% all the time back when I crunched on both a 4870 and 5870. Hmmm...I will have to look into this. I wonder if I suspend the CPU work if that will make a difference? I'll try some different scenarios when I have time here between patients and report back.

[Al]

Not seeing any changes on my xp box, my 6950 runs 98 - 99%, even while running Climate Prediction, Boincsimap and numberfields. Nvida cards are another story, mw still isn't optimized for them so they always run at 70 to 75% useage max.

[Fire$torm]

This is weird. Did MW re-code and cave in to OpenCL then? It used to be "native" ATI code, and was very efficient. I could get loads of 98 - 99% all the time back when I crunched on both a 4870 and 5870. Hmmm...I will have to look into this. I wonder if I suspend the CPU work if that will make a difference? I'll try some different scenarios when I have time here between patients and report back.

I just checked MW's app page and the ATI app is still listed as ATI14 which uses Stream code. Take a look at the MW wu's, are they tagged as ATI14?
When I first started crunching MW 2+ years ago, one could run two wu's simultaneously but that was using CCC 10.10. When ATI/AMD started adding openCL to the driver's it fubared the ability to do this, at least for my 4850s. You may wish to try an app_info file like this one (if ATI14)

<app_info>
 <app>
 <name>milkyway</name>
 </app>
 <file_info>
  <name>milkyway_separation_0.82_windows_x86_64__ati14.exe</name>
  <executable/>
 </file_info>
 <app_version>
  <app_name>milkyway</app_name>
  <version_num>82</version_num>
    <flops>1.0e11</flops>
    <avg_ncpus>0.05</avg_ncpus>
    <max_ncpus>1</max_ncpus>
    <plan_class>ati14ati</plan_class>
    <coproc>
      <type>ATI</type>
      <count>0.5</count>
    </coproc>
    <cmdline></cmdline>
  <file_ref>
   <file_name>milkyway_separation_0.82_windows_x86_64__ati14.exe</file_name>
   <main_program/>
  </file_ref>
 </app_version>
</app_info>

You might also have to change the exe reference in the config if it doesn't match the app in the project folder.

You can also try the opti_apps written by Arkayn and Crunch3r (Link) although MW tends to roll the latest opti_app into the server for client DL'ing.

Hope this helps.

F$

[DAD]

I just grabbed some nv gpu MW wu... all of them are opencl.. lame... looks like they caved to it...

I mean it's EASIER for them to code this way so they don't have to code cuda and ati separate, but it makes it very inefficient on either brand of GPU

They are also putting a LOT lighter load on my GPUs vs PG... with PG I have to do everything I can to keep them from overheating... with these MW ones, the cards are staying nice and cool

[DAD]

Those MW GPU apps, at least for NV annoy the hell out of me.. they keep running even after I close boinc down... and refuse to suspend when boinc tells them to.

They also eat up an enormous about of CPU use. ONE GPU WU is taking 25% of my CPU time.. that's 2 full cores... IMO, their GPU WUs blow... I stick to their multi CPU WUs which are actually very nice imo

[DrPop]

Yeah, MW has never been that great for CUDA cards. ATI / AMD cards though - it was always good credits (when it was up).

OK, finally got a chance to look at this rig here. Busy day. Thanks for all the replies and suggestions.
I found the problem. Running Collatz, FreeHal, and WuProp on the 6 core CPU was hogging all the CPU and starving the GPU. Strange, since this never happened before, and especially with a "small" GPU like the 4870.
When I suspended Collatz on the CPU, the GPU % immediately jumps up to 97 or 98%. At 790MHz core clock, this is spitting out WUs in 2:23 now. Very nice, much like what I remember. For the record, they are the native ATI14 code WUs.

However, what did they change? So on Kat's rig, I will have to disable one CPU core from BOINC crunching just to feed a lowly 4870 GPU MW WUs?
That is lame. Anyone notice anything similar, and is there perhaps a setting somewhere that I can change to allow the GPU feeding to take precedence over the CPU WUs, or is this a new "feature" in MW that it uses higher % of the CPU now?