Forum > Discussion Forum
Seti is down again
Richard Haselgrove:
--- Quote from: Claggy on 23 Oct 2010, 04:29:51 am ---
I'm running 6.12.4 on my T8100 Laptop, but since it can only run one Cuda project, can't tell if it's fixed.
--- End quote ---
I've loaded v6.12.4 on my Fermi, and zeroed debts on the way in so everything starts level.
The good news: yes, task startup by debt is working. Starting from zero, it finished off the SETI task in progress, building up a small amount of debt. When that finished, it pulled a GPUGrid task from well down the page, and started that next.
The bad news: it's overly agressive on task switching. 8 minutes later (triggered by a SETI work allocation), it preempted the GPUGrid task, and started running SETI again. Later on, it preempted SETI, and went back to GPUGrid.
And I'm still getting 'leakage' into CPU STD for projects which should be ineligible.
So, test with caution: this isn't the end of the story. Credit hounds may wish to avoid this version - all that task switchinbg, especially given the heavy SETI start-up penalty, isn't going to be efficient.
msattler:
--- Quote from: Richard Haselgrove on 23 Oct 2010, 01:29:19 pm ---
--- Quote from: Claggy on 23 Oct 2010, 04:29:51 am ---
I'm running 6.12.4 on my T8100 Laptop, but since it can only run one Cuda project, can't tell if it's fixed.
--- End quote ---
I've loaded v6.12.4 on my Fermi, and zeroed debts on the way in so everything starts level.
The good news: yes, task startup by debt is working. Starting from zero, it finished off the SETI task in progress, building up a small amount of debt. When that finished, it pulled a GPUGrid task from well down the page, and started that next.
The bad news: it's overly agressive on task switching. 8 minutes later (triggered by a SETI work allocation), it preempted the GPUGrid task, and started running SETI again. Later on, it preempted SETI, and went back to GPUGrid.
And I'm still getting 'leakage' into CPU STD for projects which should be ineligible.
So, test with caution: this isn't the end of the story. Credit hounds may wish to avoid this version - all that task switchinbg, especially given the heavy SETI start-up penalty, isn't going to be efficient.
--- End quote ---
I thought the task switching rate was a user configurable option.......switch between tasks every......
Does that not work here?
Raistmer:
--- Quote from: Richard Haselgrove on 23 Oct 2010, 01:29:19 pm ---The bad news: it's overly agressive on task switching. 8 minutes later (triggered by a SETI work allocation), it preempted the GPUGrid task, and started running SETI again. Later on, it preempted SETI, and went back to GPUGrid.
And I'm still getting 'leakage' into CPU STD for projects which should be ineligible.
So, test with caution: this isn't the end of the story. Credit hounds may wish to avoid this version - all that task switchinbg, especially given the heavy SETI start-up penalty, isn't going to be efficient.
--- End quote ---
IMO not only "credit hounds" but all who take care about their host performance. GPU task switching should be at minimum even stronger than CPU task switching (any task switching is bad enough, but GPU is just worst).
Interesting, will this version do preemting of MW GPU tasks that don't checkpoint at all ?
I've seen foolish pre-emting of 3min-only ATI MW tasks in prev builds (for example, when another GPU task from another project un-suspended and BOINC decides to continue it immediately (and no, no high priority mode and far-far from deadline) ).
Very nature of GPU apps implies much bigger startup overhead, for different projects this overhead will be different of course, but in all cases it will be much bigger than corresponding CPU overhead. So switching GPU task in progress w/o absolute need is counterproductive.
Raistmer:
--- Quote from: msattler on 23 Oct 2010, 02:17:19 pm ---
I thought the task switching rate was a user configurable option.......switch between tasks every......
Does that not work here?
--- End quote ---
LoL, it works ~ in same extent as user-configurable project shares in all prev builds before "backup project" was implemented properly. In short, it almost doesn't work.
[
BTW, if BOINC would respect user settings more and switch to another task only after user-specified number of minutes (if no high-priority mode involved) it could complete GPU task and do switch on task boundary cause most GPU tasks (maybe except GPU grid with their long GPU tasks) will finish earlier than even default of 60 min.
]
Richard Haselgrove:
Agree totally with you both - but this is an early, TEST build. I'll be filing a bug report, of course, with logs.
But I've no experience with MilkyWay - I'm not attached. Someone else had better do that one.
Let's see if we can get the bugs out of this one before they move on to the v6.13.xx line :P
Navigation
[0] Message Index
[#] Next page
[*] Previous page
Go to full version