"To Completion" inaccurate
log in

Advanced search

Message boards : Windows : "To Completion" inaccurate

Author Message
xyzzy
Send message
Joined: 9 Jan 10
Posts: 4
Credit: 28,072,032
RAC: 0
Message 5911 - Posted: 30 Jan 2010, 14:54:34 UTC

Collatz 2.0 (sse)tasks run like clockwork in 14+ hrs on my computer, so why is "to completion" set at 100+ hrs for new downloaded tasks??

Collatz 2.03 (cuda23) runs like clockwork in 45 min, but new downloaded tasks set "to completion" at over 3 hrs??

What's wrong with Boinc, or my setup?

xyzzy
Send message
Joined: 9 Jan 10
Posts: 4
Credit: 28,072,032
RAC: 0
Message 5916 - Posted: 30 Jan 2010, 23:27:20 UTC - in response to Message 5911.

Thanks for all the input?

KWSN-Gemjunkie[TeaM]
Avatar
Send message
Joined: 2 Aug 10
Posts: 1
Credit: 91,279,820
RAC: 0
Message 9072 - Posted: 2 Aug 2010, 11:09:08 UTC

My guess is the time references the benchmark machine they test WU's on.

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 9073 - Posted: 2 Aug 2010, 14:30:32 UTC

Here are a few reasons....
1. The flops estimate was set too high which causes the over-estimation. Fixing it too fast causes 20,000 computer to suddenly think that the 9 days cache is only 1 and download 72 days more work which they won't finish in 14 days and the majority not get returned before the due date.

2. The longer you crunch on a project, the more BOINC adjusts so that the estimates match the actual credit.

3. BOINC does a really bad job at estimating GPU run time. There are several reasons for this: a) not all GPU capable projects utilize the GPU to the same extent (e.g. compare Einstien or SETI to MW or Collatz) b) the run-time is based solely on flops and there is a lot of discussion on how to best estimate run time on GPUs when no one hits the theoretical limits (and those that don't uitilize GPUs well because they are trying to fit a square peg in a round hole do about 25% at best).

4. BOINC does its estimates from floating point benchmarks and Collatz does all integer math. Even if the estimate works for one machine, it will be way, way off on others.

5. The various GPU families vary quite a bit in run time. The HD 5xxx cards have new instructions sets which allow them to run 15-30% faster than the same speed HD 4xxx cards.

6. Running both CPU and GPU tasks really screws up BOINC. It has a single duration correction factor that it uses per project (not even per application) which means if your CPU is 6 times faster and you GPU is 20 times faster, it will split the difference making it 13 times faster which results in both being way off.

7. My guess is that your computers sense your urgency and respond by over estimating the task length so that if they don't finish at exactly the time they said they would, you won't sell them or something worse. How would they sense this? At the beginning of each WU, the Collatz app looks at your posts on the MB and attempts to determine whether you are more relaxed or stressed. Given that you get anxious when no one responds to your post within 9 hours, the app worries that you might go ballistic if it estimates too short and it takes longer than estimated, so it over compensates. The more impatient you are, the higher the estimate will be. (OK, I made this last one up.)

Tahg
Send message
Joined: 26 Sep 10
Posts: 1
Credit: 28,322
RAC: 0
Message 9787 - Posted: 27 Sep 2010, 3:36:32 UTC

Shouldn't this just be Elapsed/Progress - Elapsed = To Completion?

Profile mikey
Avatar
Send message
Joined: 11 Aug 09
Posts: 3244
Credit: 1,695,398,874
RAC: 5,303,870
Message 9790 - Posted: 27 Sep 2010, 11:22:04 UTC - in response to Message 9787.

Shouldn't this just be Elapsed/Progress - Elapsed = To Completion?


Nah Boinc is MUCH more complicated than that. Boinc is also a work in progress meaning it has been 'adjusted/tweaked' many, MANY times over its life span and although updated it still carries alot of code in it that works so is left alone.

Perhaps you can find more here http://www.boinc-wiki.info/Main_Page

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 1
Message 9798 - Posted: 27 Sep 2010, 21:35:54 UTC - in response to Message 9787.

Shouldn't this just be Elapsed/Progress - Elapsed = To Completion?


While that seems to make sense, it won't work unless you are only allowed to connect to one project at a time and must complete one workunit for the project before downloading more. Otherwise, how would the BOINC scheduler work? How would it know how many hours of work you have cached unless it has actually done work on each of the WUs in order to calculate the estimated completion time?

0 hours elapsed / 0 % progress = divide by zero error

Some projects take 20% of the time to do the first 1% of the work. Other projects take 20% of the time to do 20% of the work. So, even when partially done, you really can't tell how much time it will take. Others have no idea at all how long it will take. They make a WAG (wild ____ guess) which is why some projects will continue on to 150% complete before actually being done.

Maybe there should be several formulas available and the projects should be able to choose which method works best for them. Unfortunately, the BOINC client does all the estimating now and the only thing the project can do is adjust the estimated flops which, for Collatz, doesn't help much because it only uses IOPS and not FLOPS.

uBronan
Avatar
Send message
Joined: 30 Aug 09
Posts: 110
Credit: 29,134,968
RAC: 0
Message 9811 - Posted: 28 Sep 2010, 5:00:12 UTC
Last modified: 28 Sep 2010, 5:02:04 UTC

To be honest i fear that will be impossible, for instance if i look at world community grid te same type of units gets done between 5 minutes while others are done in 15 hours.
Or look at the scary times with CPDN where i get an estimate for 1600 hours to completion, while is finished in 280 hours.

So what i do is ignore those estimates and just use the time which it does for real, ofcourse with projects such as wcg won't help much but in time you get a fair idea what the real times are in general.

On collatz its alot easier those units allways ( if not get slowed down by other influences "games or pr0n movies ;)") they will finish in almost allways similar times

In my fastest case between 300 and 345 seconds, no matter what happens :)

In some cases you could fix those estimates by messing with the settings files of the project, but if you don't know what your doing you get more problems then you wish for.

So my advise ignore it ;)

elvis
Send message
Joined: 23 Oct 10
Posts: 6
Credit: 861,685
RAC: 0
Message 10136 - Posted: 2 Nov 2010, 19:36:26 UTC

Is there then a way to only do colatz gpu tasks seeing colatz loves gpu's and set say 2nd aqua@home to cpu only as it only like cpu's at this stage ?

rroonnaalldd
Send message
Joined: 19 Oct 10
Posts: 175
Credit: 404,110
RAC: 0
Message 10142 - Posted: 3 Nov 2010, 0:40:02 UTC - in response to Message 10136.
Last modified: 3 Nov 2010, 0:43:24 UTC

Aqua is CPU-only. For Collatz you will find optimized appz and enable GPU in your preferences.


[edit]
links
____________


Post to thread

Message boards : Windows : "To Completion" inaccurate


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.