Yeah unfortunately they only have GPU clients for Windows. There's no Linux version of the GPU/CUDA cores. Maybe you could set f...@h's priority way down with nice and see how that does? I'm not sure if it runs at defaut niceness level, but if not maybe setting it to 19 scheduling priority would increase system responsiveness? Might want to look into it a bit, if you're not swapping active programs out of RAM that might help you without having to stop/start f...@h when you need the computer.

Do you have a f...@h team? If not we could start up an MHVLUG one.

I really need to get my 9600GSO chugging for Folding, but I've been too lazy to get multilib setup on Slack64. I'm really considering just putting the 9600GSO in the other Windows machine I have and getting a GeForce 210 for the Linux box - which is more than enough for VDPAU accelerated decoding (which I actually use on that machine). Probably kind of sad that I'd rather spend 2 minutes swapping cards than 10 minutes getting multilib Slack going. I really wish they just released it multi-lib out of the box!

I was considering using the CPU clients, but they'd take my CPU off idle and I'm not sure if it's worth the power drain. The video cards have massive folding performance and great folding/power ratios.

-Frank

--------------------------------------------------
From: "John D. Mort" <[email protected]>
Sent: Friday, June 04, 2010 3:25 PM
To: <[email protected]>
Subject: Re: [mhvlug] Folding/CUDA/Multilib x86-64 distro

I use Ubuntu 64 bit with the standard f...@h install, it takes advantage
of my dual core CPU.  I haven't messed around with having it touch my
GPU.  My personal experience is that even though the processes are
given the lowest priority, I have response issues with processor or
RAM intensive programs.  (it's probably only one of those, but I
haven't bothered to determine which)

I am now in the habit of stopping f...@h when I'm going to do something
with large files, and then turn it back on later when I'm just going
to be webbing around or leave the computer idle for a bit.

--
John D. Mort
http://john.mort.net




On Fri, Jun 4, 2010 at 2:48 PM, Frank D. <[email protected]> wrote:
Does anyone here run fold...@home? Is there an MHVLUG f...@h team?

If anyone knows if what it takes to run multiple copies at f...@h under WINE
with multiple nVidia cards, I'd appreciate some input.

As far as rig options, here's what I'm looking at:
AMD 635 Propus Athlon II X4 @ 2.9GHz
ASRock 785G motherboard w/ 3x physical PCI-E x16 slots
3x GTS250, standard clocks @ ~705GFLOPS/card
Rosewill 850W PSU
4GB RAM
2x32GB A-DATA SSDs in RAID-0
My standard preferred CoolerMaster 590 case
Headless terminal setup, using remote X

The GTS250s would give 2.226GFLOPS/USD for total system cost of $950
Or use overclocked GTX260s @ ~900GFLOPS/card and an upgraded power supply
for around $1400 - this gives just shy of 2GFLOPS/USD but an impressive
2600-2700GFLOPS total output.
Or I could use "Green GTS250s" with a lower clock rate + voltage. This would actually cost $50 more and reduces the cards to around 585GFLOPS for a total
of 1755 for $1000 and giving 1.755GFLOPS/$$.  It does reduce peak rated
power consumption by 90-150 watts, though. At $0.18/kwh, that's a savings of
up to 65 cents a day..  or using around 600w vs 750w.

The actual power reduction depends on the actual clock rate change and
voltage changes, but for a 9800GT green (same G92 GPU) it's about a 30w
reduction through the voltage drop and 100MHz lower core/250MHz lower shader
clock. That's 1625MHz shaders down to 1375 on the 9800GT. The 9800GT uses
112 cores down from the 9800GTX/GTX+ @ 128.

The clock rate for the 9800GTX+/GTS250 is 1836MHz, being reduced to 1500 on
the GTS250 green. That equates to a 336MHz reduction, or 18% - up from
225/15.5% on the 9800GT. It also affects 16 more shaders on the GTS250 than
on the 9800GT. If the voltage drop is equivalent, I'd expect the power
reduction to then be greater than 30w considering the clock reduction and
the fact that its affecting 14% more shaders.

I had some initial indecision about the GTS250 vs "Green" GTS250, so I did a
little math to help me decide. I'm STILL undecided.

Normal GTS250:
$950 total system cost, 2115GFLOPS
~2.82 GFLOPS/Watt assuming 750w total system draw from the wall
$3.24/day ($0.18/kwh)
$1182/year ($0.18/kwh)
3 year cost (power+system): $4496
3 year cost($)/GFLOPS: 2.125 (lower is better)

"Green" GTS250s - max estimated power reduction.
$1000 total system cost, 1755GFLOPS
~2.925 GFLOPS/Watt assuming 600w total system draw from the wall
$2.59/day ($0.18/kwh)
$946/year ($0.18/kwh)
3 year cost (power+system): $3838
3 year cost ($)/GFLOPS:  2.187 (lower is better)

"Green" GTS250s - min estimated power reduction
$1000 total system cost, 1755GFLOPS
~2.65 GFLOPS/Watt assuming 660w total system draw from the wall
$2.85/day ($0.18/kwh)
$1041/year ($0.18/kwh)
3 year cost (power+system): $4123
3 year cost ($)/GFLOPS: 2.349 (lower is better)

As you can see, even the minimum estimated power reduction would save me
$373 over 3 years, although at reduced efficiency (energy:work - only about 90% the efficiency of the normal GTS250) and the ability to do only 83% of the work of the cheaper more power hungry system. If the power reduction is
toward the higher side, it could save as much as $658 with almost equal
efficiency.. albeit only doing 83% of the faster system's work.

I did ignore power supply efficiency, I assume efficiency will drop off at around the 500w mark, which is where the max estimated Green GTS250 number
is. If efficiency does drop, these numbers are fairly correct. But
efficiency may remain fairly constant from 300w draw on up, which would
actually put the Green GTS250s in further favor for power efficiency. I have
no data on the PSU's efficiency at different levels.

A) Is WINE+CUDA stable for f...@h?
B) Are multiple instances of f...@h on WINE+CUDA stable, and is it a huge
hassle to get it to run like that?
C) Is there a performance hit using WINE? (I'm inclined to think no, but if
someone has a definitive answer I'd like to hear it)
D) Is 4GB enough for 3 f...@h GPU2/3 clients?
E) Is a Propus 635 enough? I only run 5-20% CPU on my 940 with the f...@h
viewer disabled so I'm inclined to think the answer to this is "yes,
marginally."


MediaCoder has support for CUDA-accelerated H264 encoding under Windows,
which is PHENOMENAL. 2x-5x speed up for my transcodes, as a test I just did a 720x304 video to 640x264 @ 2300Kbps. It encoded @ 231FPS w/ CUDA vs 57FPS on CPU with an AMD 940/192c GTX260. Is there anything like that for Linux?
MediaCoder site says "due to licensing agreement with nVidia we are not
allowed to release the source to the CUDA encoder." I wouldn't even be
opposed to using a binary for that kind of encoding speed on my Linux
machines.. but they don't even have that.

Thoughts on video card selection? Power usage? f...@h GPU client under Linux?
Someone please tell me there's a CUDA encoder coming for Linux!

Can someone recommend a multi-lib 64 bit distro? I can get it going on
Slack64, but for this machine it might be less of a PITA to have something where I can just install the packages.. need multilib for WINE if I want a
64 bit OS.

-Frank
_______________________________________________
Mid-Hudson Valley Linux Users Group                  http://mhvlug.org
http://mhvlug.org/cgi-bin/mailman/listinfo/mhvlug

Upcoming Meetings (6pm - 8pm)                         MHVLS Auditorium
 Jul 7 - Patent Absurdity - The Movie
 Aug 4 - Samba
 Sep 1 - BOINC

_______________________________________________
Mid-Hudson Valley Linux Users Group                  http://mhvlug.org
http://mhvlug.org/cgi-bin/mailman/listinfo/mhvlug

Upcoming Meetings (6pm - 8pm)                         MHVLS Auditorium
 Jul 7 - Patent Absurdity - The Movie
 Aug 4 - Samba
Sep 1 - BOINC

_______________________________________________
Mid-Hudson Valley Linux Users Group                  http://mhvlug.org
http://mhvlug.org/cgi-bin/mailman/listinfo/mhvlug

Upcoming Meetings (6pm - 8pm)                         MHVLS Auditorium
 Jul 7 - Patent Absurdity - The Movie
 Aug 4 - Samba
 Sep 1 - BOINC

Reply via email to