Giant performance hit to SMP in Linux when running GPU?

[Chibo]

I have an i7 at 4GHz on Ubuntu with an 9800GT, videocard is doing a solid 5k PPD at stock clock and I thought the SMP client doing 10k PPD was fine until I mentioned it in IRC and was told it's extremely low for my clock. I turned off my GPU client and immediately jumped to 18k PPD based on TPF, so I tried switching the GPU client to GPU2 only, still no change.

Here is an example of TPF while messing with the machine:

[06:05:20] Project: 6025 (Run 1, Clone 148, Gen 248)
[06:05:20] 
[06:05:20] Assembly optimizations on if available.
[06:05:20] Entering M.D.
[06:05:26] Completed 0 out of 500000 steps  (0%) - Ignore
[06:08:51] Completed 5000 out of 500000 steps  (1%) - Ignore
[06:13:29] Completed 10000 out of 500000 steps  (2%) - GPU Client Running TPF ~4:30 (9.8k PPD)
[06:18:23] Completed 15000 out of 500000 steps  (3%) - GPU Client Running TPF ~5:00 (8.4k PPD)
[06:22:25] Completed 20000 out of 500000 steps  (4%) - Ignore
[06:25:27] Completed 25000 out of 500000 steps  (5%) - GPU Client Off TPF ~3:00 (18k PPD)
[06:28:30] Completed 30000 out of 500000 steps  (6%) - GPU Client Off TPF ~3:00 (18k PPD)

I know it's not very scientific but is this normal? I assume I should just run SMP only and try moving to -bigadv? I never really paid attention because I never started getting bonus points until today so the GPU way outpaced the SMP.

 

[sirmonkey1985]

on linux most likely.. im guessing the gpu clients being run through wine? since that would have a fair bit of overhead and dedicating a specific amount of cpu cycles to wine..

 

[Chibo]

Yeah, WINE. I'm worried about going bigadv because that needs to be run through wine too. Wish I had a computer to stuff this GPU in because 5k ppd isn't bad.

 

[Tobit]

on linux most likely.. im guessing the gpu clients being run through wine? since that would have a fair bit of overhead and dedicating a specific amount of cpu cycles to wine..

Unless A3 changed things, this was never the case. Back in Jan/Feb, I used to run all 7 of my GPUs via wine on Linux and never had a problem. I'd see a drop of maybe 500 to 1000 PPD per GPU depending on WU but had no problems folding A2 -bigadvs along side two GPUs. For the OP to lose that much PPD is unacceptable.

Standford, please tell me you haven't totally dropped the ball on Linux A3.

 

[APOLLO]

Standford, please tell me you haven't totally dropped the ball on Linux A3.

I don't think that's it, I think something else is up. Maybe the system needs optimizing or the GPU WUs are just CPU intensive lately. Haven't touched Linux in months...

To the OP, yes the PPD seems low compared to Win A3 SMP but that's to be expected up to a point. I don't know what production to expect in Linux with an i7. Did you try isolating the GPU client to a specific CPU core? Sometimes this helps on certain systems that are experiencing SMP performance issues with GPU. It's worth a shot if you know how to do this.

 

[Chibo]

What should I be expecting out of a pure SMP setup from an i7 at 3.8-4ghz in Windows? I'd almost be willing to migrate the box to Windows 7 since I have a key for it... I just have no real use for a Windows box.

I tried setting the smp client to 7 cores and leaving one to the GPU, that was the worst PPD option.

 

[musky]

My I7 920 at 3.85 gets 18K-20K on regular smp units with no GPU, so yours sounds about right with the GPU off. I also got the same ppd in Linux w/ Wine and the WinSMP client, so the Wine layer is not your problem. I never had much luck GPU folding in Linux with I7s for this same reason. If I were you, I'd scrap the GPU client and run WinSMP + Wine + bigadv. I'd expect you would see 30K+ ppd on that machine, and less power draw to boot.