Big disparity in times
log in

Advanced search

Message boards : Number crunching : Big disparity in times

Author Message
Bok
Send message
Joined: 17 Jun 09
Posts: 23
Credit: 7,404,841
RAC: 0
Message 9 - Posted: 3 Jul 2009, 14:22:31 UTC

Is this normal ?

This is the first wu I've completed. It's a CPU wu on a Phenom X2 940. The difference in time it took for mine and the wingman is a bit much :)

Rather generous credit for other one..

Or is this a GPU vs CPU smackdown ? If so, it might not be worth running this on cpu..

Bok

Profile Crunch3r
Volunteer moderator
Project developer
Project tester
Avatar
Send message
Joined: 30 Jun 09
Posts: 219
Credit: 7,515,663,568
RAC: 10,847
Message 10 - Posted: 3 Jul 2009, 14:52:10 UTC - in response to Message 9.

Is this normal ?

This is the first wu I've completed. It's a CPU wu on a Phenom X2 940. The difference in time it took for mine and the wingman is a bit much :)

Rather generous credit for other one..

Or is this a GPU vs CPU smackdown ? If so, it might not be worth running this on cpu..

Bok


Yes, it's CPU vs GPU ;)


____________

Team BOINC United.Join Science that matters.

Bok
Send message
Joined: 17 Jun 09
Posts: 23
Credit: 7,404,841
RAC: 0
Message 11 - Posted: 3 Jul 2009, 15:05:06 UTC
Last modified: 3 Jul 2009, 15:05:31 UTC

I guess the problem being that if it had been GPU vs GPU then the granted credit would be small, but seeing as it's CPU vs GPU and quorum of 2, then the GPU gets a massive increase in granted credits 783 for ~ 8mins, compared to 783 for 18hours on the cpu.

I know it's early stages in the project, so this is just information right now :)

Maybe only have CPU vs CPU and GPU vs GPU ?

Bok

Profile Crunch3r
Volunteer moderator
Project developer
Project tester
Avatar
Send message
Joined: 30 Jun 09
Posts: 219
Credit: 7,515,663,568
RAC: 10,847
Message 12 - Posted: 3 Jul 2009, 15:11:28 UTC - in response to Message 11.
Last modified: 3 Jul 2009, 15:12:36 UTC

I guess the problem being that if it had been GPU vs GPU then the granted credit would be small, but seeing as it's CPU vs GPU and quorum of 2, then the GPU gets a massive increase in granted credits 783 for ~ 8mins, compared to 783 for 18hours on the cpu.


No. Credit is fixed, there's all the info you need -> Credits are fixed


I know it's early stages in the project, so this is just information right now :)

Maybe only have CPU vs CPU and GPU vs GPU ?

Bok


I see no reason for that since CPU and GPU are doing the same work. No need to split the project.
____________

Team BOINC United.Join Science that matters.

Bok
Send message
Joined: 17 Jun 09
Posts: 23
Credit: 7,404,841
RAC: 0
Message 13 - Posted: 3 Jul 2009, 15:25:35 UTC

ah, didn't see that. Thanks for pointing it out :)

I didn't mean to split the project at all, you can limit within BOINC server to only allow wu's to go to similar architectures for validation, but if it's not needed then I agree, no need at all.

Looking forward to my CUDA wu's running once I have a few mammoth Aqua units finish up tomorrow or so.

Hopefully a linux CUDA app will be along at some point so I can get some other machines in on it too...

Bok

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 14 - Posted: 3 Jul 2009, 15:25:40 UTC - in response to Message 9.

Yes, it is CUDA vs CPU result and yes, it might not pay to run this project on a CPU. The CPU apps were created to make sure people who wanted credit in all projects would have a chance even if they don't have a CUDA or ATI card. In addition, it gives me something with which to get baseline credit readings from a PIII 800Mhz box (which is STILL working on its first v1.07 WU by the way).

Due to the nature of the 3x+1 algorithm, none of the results affect each other. That allows the GPUs to do very well due to their parallel/stream processing capabilities and the very tight loop and repetative nature of the algorithm. About the only CPU time used for the GPU apps is to update the percent done and to iterate through the output to find the number with the largest steps.

The ATI app also does much better than the CPU version but not quite as well as the CUDA cards. I'm working on a new version that uses the latest ATI Stream SDK which has better async processing and memory management which might bring the ATI cards up to the same level as the CUDA cards.

Credit is currently being awarded using the results from v1.04 app which was using the MPIR library. The current version, v1.07, uses the GMP library for handling large numbers and it is considerably faster than MPIR. Once the baseline machine returns some results, I'll be adjusting the credit.

At some point, I will have to make the credit variable and adjust it according to the actual work done in each WU since there is 5-10% difference in IOPS from one WU to another and that could increase as the average steps increases as the numbers get larger.

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 15 - Posted: 3 Jul 2009, 15:33:36 UTC - in response to Message 13.

Hopefully a linux CUDA app will be along at some point so I can get some other machines in on it too...


That is planned although I can't say exactly when. There are a couple other items with a higher priority right now (validator that works 100% of the time, ATI apps that don't crash quads and V8's., etc.)

Profile Gipsel
Volunteer moderator
Project developer
Project tester
Send message
Joined: 2 Jul 09
Posts: 279
Credit: 77,436,650
RAC: 76,989
Message 17 - Posted: 3 Jul 2009, 16:14:09 UTC - in response to Message 14.

The current version, v1.07, uses the GMP library for handling large numbers and it is considerably faster than MPIR. Once the baseline machine returns some results, I'll be adjusting the credit.

At some point, I will have to make the credit variable and adjust it according to the actual work done in each WU since there is 5-10% difference in IOPS from one WU to another and that could increase as the average steps increases as the numbers get larger.

Maybe these arbitrary precision libraries are a bit of an overkill? Have you thought about using the exactly same algorithm as for the GPU applications? As you only have a division by 2 (a one bit shift), a multiplication by three and an addition of 1 as necessary operations, these libraries may have some overhead (for the added functionality) you don't need here.

On the credit issue I think it would be best to grant it simply according to the number of steps for each individual WU. It would be conveniently self adjusting as you mention in the credit thread.

Profile STE\/E
Avatar
Send message
Joined: 12 Jul 09
Posts: 581
Credit: 761,710,729
RAC: 0
Message 78 - Posted: 12 Jul 2009, 12:26:36 UTC
Last modified: 12 Jul 2009, 12:27:27 UTC

Being new here I'd like to know more about the big differences in run time's, I have some Wu's that are @ only 10% after 2 Hours of run time. I have CPU work Un-Checked in my Preferences so why are these Wu's taking so long compared to a lot of other ones.

Also I take it the Credits are Fixed so if a Wu takes 20 hours which some of these seem like they will you'll only get the 408 Credits for them like the ones that take only 4-6 Minutes ... ???

Profile Crunch3r
Volunteer moderator
Project developer
Project tester
Avatar
Send message
Joined: 30 Jun 09
Posts: 219
Credit: 7,515,663,568
RAC: 10,847
Message 80 - Posted: 12 Jul 2009, 13:23:05 UTC - in response to Message 78.

Being new here I'd like to know more about the big differences in run time's, I have some Wu's that are @ only 10% after 2 Hours of run time. I have CPU work Un-Checked in my Preferences so why are these Wu's taking so long compared to a lot of other ones.

Only 10% after 2hours is unusual. Some more info about the machine would help solving the mystery.


Also I take it the Credits are Fixed so if a Wu takes 20 hours which some of these seem like they will you'll only get the 408 Credits for them like the ones that take only 4-6 Minutes ... ???


Yes.
____________

Team BOINC United.Join Science that matters.

Profile STE\/E
Avatar
Send message
Joined: 12 Jul 09
Posts: 581
Credit: 761,710,729
RAC: 0
Message 82 - Posted: 12 Jul 2009, 13:51:13 UTC - in response to Message 80.
Last modified: 12 Jul 2009, 13:53:12 UTC

Being new here I'd like to know more about the big differences in run time's, I have some Wu's that are @ only 10% after 2 Hours of run time. I have CPU work Un-Checked in my Preferences so why are these Wu's taking so long compared to a lot of other ones.

Only 10% after 2hours is unusual. Some more info about the machine would help solving the mystery.


Also I take it the Credits are Fixed so if a Wu takes 20 hours which some of these seem like they will you'll only get the 408 Credits for them like the ones that take only 4-6 Minutes ... ???


Yes.


It was like that on most of my Box's but I was running other CPU Wu's & GPUGrid Wu's & even some AQUA CUDA & MT Wu's so I think the Collatz & the other Wu's just weren't getting along very good & fighting over CPU Share. Several Collatz Wu's would be only running in the 4-6 Minutes Range but each Box had 2-4 Collatz Wu's in the 2 hour range that were only @ 10%.

I Suspended all other CPU Wu's & the GPUGrid WU's I ran out & I haven't had any long running ones since ... Running GTX 200 Series NVIDIA Cards on Windows XP Pro SP2 64-Bit OS's on Intel Quad Cores Q6600 & 6700 & 1 i7 920 Box ...


Post to thread

Message boards : Number crunching : Big disparity in times


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.