GPU's gone wild

[Bender10]

The PS3Grid project has picked up a lot of speed since it branched out to include Nvidia GPU's. There are teams posting +200k a day, and this is still in beta!?

Because of this, we have dropped from 11th down to 14th, with more sliding to follow.

The clients are pretty stable. Most of the errors are user induced. Users are pushing the envelope a bit to help work out the kinks in the clients. I have have a stable Linux quad running for about 3 months. And Vista seems to be working ok also (XP has some kind of memory issue with boinc 6.3.10 and below).

http://www.ps3grid.net/

Stop by the site and look around. Don't get overwhelmed by the user/testers (most everyone). And most people forget to post what hardware they are running. But, there is some good information to be found (or ask me, I'll help).

[mitchellds]

I put my two ps3's back on grid yesterday,(gotta wait for 36 hours for results). One I have running the penos, and the other I upgraded to YD 6.0. I wanted to compare the results.

I think I'll add some GPU's and run under linux once the new app is out that wont chew up my cpu power when running the gpu code. When is that supposed to be? Any idea?

[liuqyn]

I'm going to be switching my ps3 over here soon as well. however, I don't have any good nvidia cards to run it on, so thats not an option.

[Bender10]

The new Linux client (6.3.10??, I'm on 6.3.8) is supposed to run a PS3grid wu without any (or very few) cpu cycles. Client 6.3.8 does a wu with 1/10 of the normal cpu cycles.

I just found that today. Gotta run down my wu's before I do the upgrade.

My win64 client does a wu using ~15 - 20 cpu seconds.

The BIG issue is getting the client to utilize the un-used core for normal Boinc projects. There is a back-door way for the user to do this using 'ncpu' in the Boinc cc_config (?) file.

[mitchellds]

Bender,

Are you saying with 6.3.10, on a typical quad core with a gpu, that it would still only use one core of the GPU, or there would still be one core of the quad cpu left unavailable under this setup, for a total of 5 cores instead of 6

(3 cpu + 2 GPU) ?
(4 cpu + 1 GPU) ?

how about:
(4 cpu + 2 non sli GPU cards installed)
or
(8 cpu + 2 non sli GPU cards installed)

Anyone done that yet?

[Bender10]

I typed up a great answer, and could not post due to getting timed out....

Shorter answer:

Yes

there would still be one core of the quad cpu left unavailable under this setup

1 'core' is still needed to feed the GPU. So, 1 gpu wu + 3 Boinc wu's is all you can run with a quad on a stock client, and the same with a dual (1 gpu wu + 1 Boinc wu). The cpu usage in support of the GPU is going down with each client update.

My Linux box uses ~4,000 seconds of cpu time per GPU wu. (8800GS = 1 wu in ~20 hours)
My XP box uses ~15-20 seconds of cpu time per GPU wu. (8800GT = 1 wu in ~16 hours)
My PS3 box uses ~81,000 sec of cpu time per wu. (Thats ~22 hours)

They run different clients.....But a new Linux client came out today that should change those numbers.


Funny note:

The Devs made a 'mistake' a few days ago and a bunch of users (me too) crunched for a day or so on all cores+1 (quad = 1 gpu + 4 cpu). They fixed their 'mistake', But that caused quite a stir.



EDIT: added run times for different platforms.

[Bender10]

I just checked the Free-DC stats, and we are at 13th, for a few more days. Thats not too bad for the amount of GPU's other teams are throwing at this project. The teams making the big points have a bunch of people running just 1 GPU each.

Looking at the stats, If we can maintain an ~35k average output, we will stay around 13-14th place here.

I'm running 2 PS3's and 2 GPU's to output around 12-15k per day.

[Bender10]

My WinXP64 box (and other users on XP64) is having problems crunching. It runs for a couple of days fine, then burps, and errors out some wu's. Rebooting the box every couple of days keeps this from happening (and that is a pain to remember). The project admins have not come up with a reason for this yet.

I have to keep a sticky on the monitor to remind myself...

Win32 and Linux64 seem to behave fine.

I may have to build another Linux box.....

[PoorBoy]

My WinXP64 box (and other users on XP64) is having problems crunching. It runs for a couple of days fine, then burps, and errors out some wu's. Rebooting the box every couple of days keeps this from happening (and that is a pain to remember). The project admins have not come up with a reason for this yet.

I brought this up in the PS3Grid Forum but apparently as usual nobody took me seriously. Whats happening is the BOINC Manager in Win 64 is slowly taking up all of your memory until the PC won't run or acts very sluggish & will start to Error out the PS3 Wu's until you Re-Boot & that refreshes the memory again.

A simple solution for me on the Win 64 Box's is to not leave the BOINC Manager Window on any Manager Window that has activity. By activity I mean like Time Counting Down until a Project Updates it's self again. I noticed by watching the Task Manager & BOINC Manager @ the same time that for every Tick of a Second in the Manager the Managers Memory usage increased. I've seen it's usage get as high as 5 Million MB. So now I always before Minimizing the Manager put the Manager Window to either Statistics or Disk, this keeps the Manager on Win 64 Box's from slowly sucking up all the Systems Memory Resources & I haven't had a PS3 Wu Error out since I've been doing that.

Try it & see if it works for you Bender or anybody else too, this works on the v6.3.14 that I'm using but would probably work on other Versions too, hopefully or maybe the next Version Release won't need this simple work around ...

[Brucifer]

Well I'll ask a question or two here in friendlier turf before I go to the website and run into the hard cores being as I'm totally ignorant on the GPU thing...

For starters, what GPU cards are you talking about here? Like what is the best bang for the buck on the GPU cards, assuming running a linux system. What is the power consumption on the card(s)?

What's the best path to follow here, ie get a really low power consumption cpu to handle the GPU maintenance requirements, and then put all the heavy duty processing into the GPU? Can you run multiple GPU cards?