Too much WU at a time for Arecibo Pulsar search
1 May 2010 13:59:39 UTC
Topic 194901
(moderation:
Hi,
I noticed that some of my computers, especially those which aren't up 24/24, are receiving sometimes almost 100 WUs to compute in one shot!!!
In addition with a very short deadline (15 days), which isn't adequate for computers not up 24/24... The delay is correct when you have only ONE WU for Einstein, not when you're on multiple projects and than Einstein gives you 100 WUs to compute!!!
Result is always the same: the Boinc's scheduler puts these WU in high priority mode, other project are put aside to the benefit of Einstein. I don't like this at all...
As I don't want to start the "guru tuning" of my Boinc preferences, I'm glad to receive in one bunch 100 WUs to compute, but at least give me a deadline of a month or more.
Regards
Too much WU at a time for Arecibo Pulsar search
)
That's not the fault of the Einstein project! It's a bug with the newest BOINC clients on machines with a CUDA device that has not enough memory.
You should disable the use of the nVidia (CUDA) device in your Einstein@Home preferences as long as this error persists.
Gruß,
Gundolf
Computer sind nicht alles im Leben. (Kleiner Scherz)
Thanx for that prompt
)
Thanx for that prompt response!
Arf! OK!
On the concerned computers, I'm using 6.10.43 x86_64 with GeForce 8800 GTX 512Mo!!! Not enough?!
OK. Should I give a try to the newest client: 6.10.50? I hate to modify my global prefs, coz I forget to put them back after a while!!!
RE: I'm using 6.10.43
)
What Gundolf meant to say is that sometimes tasks get stuck in video memory. We don't know why this happens, but as long as such a task is stuck, the GPU won't have all memory free. Although there is code in 6.10.43 that should prevent more work to be downloaded to a GPU whose memory is partially filled by a stuck task, this code has a bug in it. That bug is fixed in later code.
Don't try 6.10.50 as it is a development version with a whole range of bugs on its own, one of them pretty seriously, having to do with the Network availability preferences. If you want to, you can go back to what's now the recommended version, 6.10.18, which doesn't have any code preventing GPUs to download more work when something is stuck... but instead in the case of that version the tasks will err out. Get enough that err and your daily quota will go down.
Or do as Gundolf said and change your project preferences (not Global preferences !!).
RE: Don't try 6.10.50 as it
)
OK. Refering to previous exchange we had, 6.10.4x was interessant to bypass scheduling problem between cpu and gpu.
So, as you both said, until further debugging:
I'll do that!!!
Thx to both of you.
.51 is supposed to fix the
)
.51 is supposed to fix the snarfl in the preferences ...
RE: .51 is supposed to fix
)
Though it doesn't exist for either Windows or Linux yet. Waiting is.
RE: RE: .51 is supposed
)
Hmm, Zen perhaps? Waiting is ...
Or is it more like me tying to figure out what happened to my other C-Chip cookie? I only ate one last night ... can't find the other one today ... and only me and the dogs and they would have left crumbs ...
I run boinc 6.10.56 and I
)
I run boinc 6.10.56 and I also have the problem, I even put einstein to 50% ressource and reset debt to 0, but after a few days, more and more eintein WU stack up and at a moment it runs in high priority mode on all CPUs.
My GPU is CUDA with only 256Mb so I disabled it in einstein preference weeks ago when I saw this hread, but it changes nothing on the behaviour :-(
That has to be another
)
That has to be another problem, since you write that the tasks are running on the CPU. The problem of the original poster was that GPU tasks were downloaded but never started to run.
Since your GPU has only 256MB, you'd get no GPU tasks anyway, regardless of your preferences, since Einstein needs more than that.
What are your preferences for "Computer is connected to the Internet about every XX days" and "Maintain enough work for an additional YY days"?
Do you use local preferences?
Is it very important for you to hide your computers? We can get much information from those details without invading your privacy.
Gruß,
Gundolf
Computer sind nicht alles im Leben. (Kleiner Scherz)
I use connect about every 1
)
I use connect about every 1 day and additional work buffer is 0.5 days
It's a quad xeon so 4 CPUs, it runs seti and milyway at 100% and einstein at 50% (was 100 before).
What I see is that even if it should have work for about 1.5 days, einstein WU stack up days after days and represent dozens of hours pending so after a week for instance, einstein runs 4 WUs in high priority.
Let's say that if einstein download 38 hours of works every 36 hours, it drifts slowly... next time I will let it crunch in high priority and see what happen next, if he continues to d/l too much work or re-balance debt.