Skywise <into@oblivion.nothing.com> wrote:
DaveT <svirtftcyt@yahoo.com> wrote in
news:aircu4t2pc02k1salenm7qq2tjes76epsj@4ax.com:
BTW, my machine is chewing through these CUDA units in about
2:15 - 2:30.
That's hours and minutes. I left the seconds off. I've lain here
and watched them tick by. Maybe they've changed this with version
6.6.20. It appears that the time in the "elapsed" collum is roughly
equal to wall clock time, although I haven't timed it exactly. But
I've watched it from beginning to end so I can approximately
vouch for those times.
CUDA workunits do not report time properly. The 'time' shown
is CPU time, not wall-clock time. The CPU time shown is just
the small amount of CPU that has been used to keep feeding
data to the GPU.
This I actually can monitor with Process Explorer (www.sysinternals.com).
It is so miniscule it is hard to see. It goes between 0% - 0.75% - 1.5% - 3%
in sort of a rhythmic pattern.
CUDA does not run 100% on the GPU. A tiny piece also runs
on one of the CPU cores, which I understand is the overhead
to monitor the GPU and send it data.
Per above.
You can test this by waiting for a CUDA work unit to start
and noting the time on a clock, then noting the same clock
when the work unit is done.
I've been doing this since it started. They run about 2 hours
15 minutes to about 2 hours 30 minutes.
A side effect of this is - at least with my versions, it
may have been fixed? - is that when BOINC requests more
work units for CUDA, it keeps asking for more and more and
more. It thinks a work unit only takes 3 minutes (CPU time),
and if you're set to download 1 days worth, it tries to
download 480 work units (broad generalization).
They must have changed this. I changed the setting from 0.25 days
to 1 day. The most I am able to obtain is one unit runnung, one on deck,
and, occasionally, one completed until it goes away. I get regular
groups of messages that appear as follows:
4/15/2009 11:03:30 PM SETI@home Sending scheduler request: To fetch work.
4/15/2009 11:03:30 PM SETI@home Requesting new tasks
4/15/2009 11:03:40 PM SETI@home Scheduler request completed: got 0 new tasks
4/15/2009 11:03:40 PM SETI@home Message from server: No work sent
4/15/2009 11:03:40 PM SETI@home Message from server: No work available for the applications
you have selected. Please check your settings on the web site.
4/15/2009 11:03:40 PM SETI@home Message from server: CPU jobs are available, but your
preferences are set to not accept them
Then occasionally I will get a group that looks like this:
4/15/2009 11:23:56 PM SETI@home Sending scheduler request: To fetch work.
4/15/2009 11:23:56 PM SETI@home Requesting new tasks
4/15/2009 11:24:06 PM SETI@home Scheduler request completed: got 0 new tasks
4/15/2009 11:24:06 PM SETI@home Message from server: No work sent
4/15/2009 11:24:06 PM SETI@home Message from server: No work is available for SETI@home
Enhanced
4/15/2009 11:24:06 PM SETI@home Message from server: No work available for the applications
you have selected. Please check your settings on the web site.
But somehow it always seems to get a unit by the time it is ready for one
(and it only downloads one).
Eventually, I will let it download some CPU work. Right now, I am content to run it as is.
I was irritated by the fact that I wanted to try CUDA and it immediately downloaded
two Astropulse units that stuck in the computer for a week and it wouldn't download
any GPU work. That was with Boinc 6.6.4 or whatever, the 'recommended" one.
So you'll have to watch that you don't download too many work
units. I have to watch this because I like to keep several
days of work on the drive for those inevitable server outages.
I'd like to get a few but it won't do it.
Brian