| Subject: Re: How do I tell if BOINC is using CUDA? |
| From: Skywise |
| Date: 15/04/2009, 05:43 |
DaveT <svirtftcyt@yahoo.com> wrote in
news:4m98u45del8auufi04558tgb4tp28i53n8@4ax.com:
Thanks. This is excellent and useful information.
You're welcome. I usually just lurk, but my experiences
appeared useful, so I shared. Isn't that what it's all
about? :)
Skywise <into@oblivion.nothing.com> wrote:
After reading the existing posts I thought I'd chime in with
my experience. First, I am not a S@H expert even though next
month is my 10 year anniversary with the program.
I think I started a month later. Was 1999 the year it started?
Seti@home went live on 5-17-99. I signed two days later on 5-19.
Anyway, I have a working CUDA system.
What a struggle it has been to make this happen.
The two Asstropulse units that downloaded will finally finish tomorrow
so I can take another stab at CUDA.
Hardware:
Core2Quad 3Ghz Q9650 CPU on Supermicro X38 based mobo w/4 gigs
EVGA Nvidia GeForce 9800 GTX+ with 512meg
What OS are you running? My understanding is that XP and Vista 32 bit
can only utilize 3GB of RAM. I have a dual-core XP64 machine in addition
to the XP32but it is AGP 8x and I'm not looking for a CUDA capable AGP
8x card.
The machine in question is running XP-Pro SP2.
The question of how much memory it can use is a complicated one.
Even if it can access only 3 gigs, if it's a 32 bit system it can
only access 4 gigs max anyway. And since things like video cards,
audio, USB, ports, pci busses, and other motherboard services
require memory, that will be taken from the 4 gig total anyway.
Simple example - start with 4 gigs. Get a video card with 512 megs.
That 512 megs has to occupy address space, so it's deducted from
the 4 gigs. That alone leaves you with 3.5 gigs left.
(If you only have 2 gigs, that 512 megs of video memory will be
outside of main memory, so you don't lose any)
This is part of the advantage of 64 bit systems. More addressable
memory space.
Now, I do have to babysit the system a bit to keep it fed with
the right work units.
You will have to go into your seti settings in your BOINC
account online where you will find check boxes to permit or
deny which type of work units you want done.
I'm a bit confused as to whether this change is made locally in Boinc
or whether the change is implemented on the Seti end.
You go to the seti web site and log into your account settings.
http://setiathome.berkeley.edu/home.php
About the middle you will find a link "SETI@home preferences".
In there you will find the check boxes for selecting what
type of work you want your system to do.
I currently do not D/L astropulse even though my system chugs
through them pretty quick, in about 36 hours. (x4 as I have
four cores)
I'm puzzled by this. My 2.5 GHz dual-core AMD slogs through 2 workunits
in about 7 days. That would be 3.5 days per unit, but it still takes 7
days to get them done. Do you mean that your 3 GHz machine processes 4
Ap workunits in 36 x 4 = 144 hours, or will it actually chew through a
workunit in 36 hours? I know Intel CPU's have large caches, but that's a
five-fold increase.
OK. Each core is it's own independent CPU. One process runs on
one core.
Say it takes a core 24 hours to process one work unit. But if you
have 4 cores, you can process four work units simultaneously.
Each work unit still takes 24 hours, but you are doing 4 at once.
One work unit will not be spread across four cores to be processed
in 6 hours.
I find my NVidia card with one GPU can burn through as many
CUDA workunits as all four cores of the CPU (or faster). So
I have to keep the number of CUDA vs CPU workunits balanced.
What a hassle. You'd think they'd CUDA-ize Astropulse since it drags
along so slowly.
I think it's being worked on.
Brian