WU take longer after update from 6.4.7 to 6.6.38
10 Oct 2009 15:59:03 UTC
Topic 194570
(moderation:
Hello,
I had BOINC 6.4.7 running for a long time, on several machines still do, but I got 33 CUDA apps. in one time and that was not funny. So I got the advice to upgrade to BOINC 6.6.38. I followed that advice and for almost two weeks everything runs fine without errors, the beta-app. as well.
However I found out that previous CUDA-WU’s took around 16000 seconds to proceed and S5R5 WU’s for around 20000 seconds. Now with BOINC 6.6.38 the CUDA WU’s takes around 24000 seconds and the S5R5 around 27000.
This is a huge difference, to me though. However is it just impatience me, or have other people the same experience?
Greetings from
TJ
WU take longer after update from 6.4.7 to 6.6.38
)
Please realise that BOINC itself doesn't crunch so that CPU time reported for each task should only vary with app version but not BOINC version. Even if a newer BOINC was itself consuming more CPU cycles than previously, that would not show up as a variation in crunch time. As with any other use of the CPU by whatever else might be running, increased BOINC use would show up as a drop in CPU efficiency. Unless you monitor these things very closely and keep a log of numbers over time, it's unwise to rely just on perceptions :-).
I have no nVidia cards, so no experience with CUDA. With S5R5, there is quite a cyclic variability between tasks depending on (amongst other things) where in the sequence each task lies. Variations of the magnitude you mention for run time are quite normal and should be accompanied by a similar variation in the granted credit. The ability to predict run time variations is not that easy so you will still see examples of run time variations that are not able to be properly compensated for. The system that is in place (developed, at least partly by Bikeman, I believe) does a pretty good job and is much better than the 'no correction' scenario we used to have.
Hopefully others using CUDA will respond but in my recent experience, with both S5R5 and CPU only ABP1, there is nothing unusual of the magnitude you quote happening due to changes in BOINC version. During this year I have gone from 5.10.45 to 6.4.7/6.5.0 to 6.10.13. I did skip the 6.6.x series completely, not from personal experience, but from the experiences reported by others. 6.6.38 was only quite recently made the official recommended Win version and I think it is mostly OK. I'm using 6.10.13 on some machines because I have some ATI cards and so far I'm not seeing any problems with EAH running perfectly normally on that version. There should soon be an 'official' 6.10.x release and 6.10.13 is the 'release candidate'.
Cheers,
Gary.
RE: However I found out
)
Are those times the CPU run times (as shown on your computers/tasks list), or are you talking about seeing the time difference in BOINC Manager?
I ask, as BOINC Manager changed between 6.4.7 and 6.6 from showing CPU time to showing the actual Wall time. You'd be surprised to see the difference between the two, to see that for a long time when the task is already running, it's doing nothing with the CPU.
Of course, CUDA didn't show accurate time on 6.4.7, as the CPU is only used for the small bit of time that it transfers data from memory to the GPU and back. That's also a major reason why the CPU time is no longer shown, but the more accurate wall time.
RE: Please realise that
)
I understand that Gary, BOINC is just a "distributor". I have several machines running for Einstein. Einstein is why I started with it and the other projects were started during outages of Einstein, they are all not working at my farm.
I look at progression closely, between different pc, different OS, and different versions of BOINC. I also run the Beta-apps. from Einstein on several.
But I saw a huge change in time crunching after the update and that got me wandering, off course I know that BOINC is not doing the crunching.
I do not bother with credits because I run this as being an “Einstein fanâ€, though I like them. But I see pc with more “potential†i.e. more RAM, faster processor end so on, but taking more time to crunch. Some WU’s take longer, that I have seen in the years and is thus normal for the project. But I saw that after the upgrade of BOINC, WU’s and the CUDA WU, take longer to finish on that pc (i7, 7 cores to use for crunching, and doing else during my sleep period). My older Q6600 with the 6.4.7 version and no CUDA is still finished within 14000-17000 seconds. There was no change in crunching time, and I monitor that daily. I check the “tasks for user†page for that.
Greetings from
TJ
RE: Are those times the CPU
)
Yes, I compare the times of finished results. I keep track of the results for all my pc’s, and have the experience that they are in the same range, varying only a few 1000 seconds.
After the update I found an increasing crunching time of 4000+ seconds, for the S5R5 WU’s and even more for the CUDA WU, on my fastest PC, the i7.
The estimated running time for the CUDA WU in BOINC manager is approx. 4 times longer, than the actual calculation time, on the i7 with 6.4.7. But after the upgrade to 6.6.x actual running time is still much shorter then estimated.
Greetings from
TJ
There was an issue with tasks
)
There was an issue with tasks taking longer, up to 4 times longer, on some versions of BOINC. 6.6.20 was the worst of the bunch and it affected both the CPU and GPU side though it seemed more common with the GPU tasks.
I could not find a specific "smoking gun" to point to though you could monitor BOINC and stop and restart it to "clear" the issue. Some of the later versions in the 6.6.3x area seemed relatively immune ... YMMV
In that I was never able to pin this one down to specific messages / anomalies in the messages log the photo proofs I submitted were blown off as inconclusive. You may be seeing this issue, or having something else ...
RE: There was an issue with
)
Hi Paul,
Thanks for proofing that I am not the only one who is seeing a difference.
Greetings from
TJ
RE: RE: There was an
)
The issue we suspected was that BOINC was starting and stopping the tasks... I could not prove that so ... :(
The thing is, things are not always simple and cut-and-dried ... BOINC is complex enough that many things interact and unexpected results abound ...
Anyway, you can prove that this may be the issue if you see tasks that are running "slow", stop BOINC and the science applications and restart BOINC... if it speeds up then this is likely what I had been seeing. You can also try one of the later BOINC versions to see if that helps also ...
RE: RE: RE: There was
)
Great info thanks!
Overclock with the MSI G31M3-L and Intel E8600 3.33Ghz
Intel D865GLC Socket 478 Motherboard Review
Overclock your ASUS 1005HA netbook and crunch more
RE: Great info
)
You are most welcome.
As always, YMMV ... :)
I have it closely watched
)
I have it closely watched over the past week and the Q6600 with Vista x86 took between 16.000 and 18.000 seconds to complete a S5R5 WU with BOINC 6.4.7.
The i7 with Vista x64 took between 21.000 and 24.000 seconds to complete S5R5 WU’s with BOINC 6.6.38.
Both systems ran 24/7 without any throttling, at manufactures (dell) settings.
To me this means that the Q6600 is more efficient in calculating and/or BOINC 6.6.38 has a slowing influence however I do know that it is not actually doing the crunching, it is only a manager.
Greetings from
TJ