Everyone thinks their systems are the baseline for calculating ratio of
BRP4 to BRP5. I have seen as many ratio's of 10 to 1 as I have seen 7 to 1
I get either 10 to 1 or 10.5 to 1 runtime ratio's running two at time.
I think the credits awarded should be equal to or slightly greater than BRP4
to make up for the longer runtimes where one job has a greater chance of failing
Can the server calculate the average runtime of BRP4 vs BRP5?
I think there is a big variation of runtime of BRP5. I've looked through the finishe BRP5 wus and I can see runtime vary from 12,900sec up to 41,500sec and one with 29,000sec.
My finished BRP4 wu range from 1,300sec to 6,500sec.
am i right in understanding of points is a Ñurrency without gold standart?
you can set any amount of points for a WU without relation to calculating job done in floating point calculation?
am i right in understanding of points is a Ñurrency without gold standart?
you can set any amount of points for a WU without relation to calculating job done in floating point calculation?
Yes you are. There's a gold definition/recommendation, but no gold standard.
What's worse, is that the points (however granted) are converted back into floating point operations at the golden definition rate. Which results in some absurd claims of GFLOP contributions on pages like Top 100 volunteers.
I think there is a big variation of runtime of BRP5. I've looked through the finishe BRP5 wus and I can see runtime vary from 12,900sec up to 41,500sec and one with 29,000sec.
My finished BRP4 wu range from 1,300sec to 6,500sec.
I think it would be accurate to say there is a big variation of their time on your system, but not traceable to inherent computational differences in the WUs themselves. On other systems they are remarkably consistent in elapsed time (within the basic groupings of BRP4 and BRP5).
Everyone thinks their systems are the baseline for calculating ratio of
BRP4 to BRP5. I have seen as many ratio's of 10 to 1 as I have seen 7 to 1
I get either 10 to 1 or 10.5 to 1 runtime ratio's running two at time.
I think the credits awarded should be equal to or slightly greater than BRP4
to make up for the longer runtimes where one job has a greater chance of failing
Can the server calculate the average runtime of BRP4 vs BRP5?
Hi Tom.
I have to reluctantly agree with you, but there's a modifier.
I've looked at eleven of my machines and I'm surprised by what I am seeing.
My GT 240 is underpaid for the longer work - it doesn't do it very well.
Both of my AMD cards 6770 / 7770 seem to be overpaid.
Both dual 660Ti machines appear to be "just about right."
The GTX 470s are underpaid just a little.
The triple GTX 670 is underpaid.
In all of those we're talking about a 2-8% reduction or bonus.
I don't have enough GTX 460 numbers for me to feel good about a comparison.
But here's what surprised me:
All of the 500-series NVIDIA cards are the most under-paid (in my case, four machines representing 9 GPUs... I didn't count one... it's complicated...).
These are very close to 10:1, so they are being points-deprived by close to 20%. They just seem to be *really good* at doing the BRP4 Arecibo work.
But all of this is very preliminary because useful conclusions can only really be drawn after we're sure all of the other work has cleared from every machine.
Then my observation is that if 500-series cards are paid to parity, everything else is going to be overpaid by as much as the 500s seem to be underpaid.
I suspect all of this has to do with the efficiency of the application on any given card, which further complicates things. I'll bet the 500s would do much better if they could use CUDA 4.2, and I'm SURE the 600s would do better using CUDA 5.
There's apparently nothing we can do about that.
So, what do you suggest that would make everyone happy?
I'm sure we'll get to a good number and Bernd seems completely willing to work this out as best it can be worked-out.
Time keeps flowing like a river, to the sea.
I'm going to let this rest for a while until I have more reliable data. Since I'm a "sometimes" participant, I really don't even want to have any influence, but I have enough very different machines that I don't mind providing data if that will help the discussion.
All of the 500-series NVIDIA cards are the most under-paid (in my case, four machines representing 9 GPUs... I didn't count one... it's complicated...).
These are very close to 10:1, so they are being points-deprived by close to 20%. They just seem to be *really good* at doing the BRP4 Arecibo work.
During my small test on Albert I got a ratio of 10.8 : 1 for my GTX580
As I said when you asked some of us a similar question by email last month, "I think we ought to be careful about the distinction between projects which use the full-blown CreditNew package, and projects which use the runtime estimation and other server-side components that were introduced around that time, but use one of the alternative credit mechanisms - as you do with Albert."
What I meant here (and I'm almost sure that's what I wrote you at the time) was: Which projects do actually grant credit using CreditNew?
Quote:
Most readers here will have in-depth knowledge of maybe a dozen of those?
This is precisely the reason why I asked that question more publicly here. More readers should know about more projects, right?
I think it would be accurate to say there is a big variation of their time on your system, but not traceable to inherent computational differences in the WUs themselves. On other systems they are remarkably consistent in elapsed time (within the basic groupings of BRP4 and BRP5).
I agree that other programs like antivirus and so on may also influence runtime of WUs. Crunching different projects or 2 WUs on GPU in parallel as well.
After WUs were downloaded by boinc manager the estimated runtime is calculated and shown. I can observe that the estimated runtime for not started BRP4 WUs vary from 20 minutes up to 120 minutes. Sure thats not the real runtime for a WU but it is good indicator were the real runtime might end.
Do you have an explanation for this?
Edit: My computer is staffed with an i7-3930k and 64GB RAM and Radeon HD7850 with 2GB VRAM.
Everyone thinks their systems
)
Everyone thinks their systems are the baseline for calculating ratio of
BRP4 to BRP5. I have seen as many ratio's of 10 to 1 as I have seen 7 to 1
I get either 10 to 1 or 10.5 to 1 runtime ratio's running two at time.
I think the credits awarded should be equal to or slightly greater than BRP4
to make up for the longer runtimes where one job has a greater chance of failing
Can the server calculate the average runtime of BRP4 vs BRP5?
I think there is a big
)
I think there is a big variation of runtime of BRP5. I've looked through the finishe BRP5 wus and I can see runtime vary from 12,900sec up to 41,500sec and one with 29,000sec.
My finished BRP4 wu range from 1,300sec to 6,500sec.
am i right in understanding
)
am i right in understanding of points is a Ñurrency without gold standart?
you can set any amount of points for a WU without relation to calculating job done in floating point calculation?
RE: am i right in
)
Yes you are. There's a gold definition/recommendation, but no gold standard.
What's worse, is that the points (however granted) are converted back into floating point operations at the golden definition rate. Which results in some absurd claims of GFLOP contributions on pages like Top 100 volunteers.
Eric_Kaiser wrote:I think
)
I think it would be accurate to say there is a big variation of their time on your system, but not traceable to inherent computational differences in the WUs themselves. On other systems they are remarkably consistent in elapsed time (within the basic groupings of BRP4 and BRP5).
RE: Everyone thinks their
)
Hi Tom.
I have to reluctantly agree with you, but there's a modifier.
I've looked at eleven of my machines and I'm surprised by what I am seeing.
My GT 240 is underpaid for the longer work - it doesn't do it very well.
Both of my AMD cards 6770 / 7770 seem to be overpaid.
Both dual 660Ti machines appear to be "just about right."
The GTX 470s are underpaid just a little.
The triple GTX 670 is underpaid.
In all of those we're talking about a 2-8% reduction or bonus.
I don't have enough GTX 460 numbers for me to feel good about a comparison.
But here's what surprised me:
All of the 500-series NVIDIA cards are the most under-paid (in my case, four machines representing 9 GPUs... I didn't count one... it's complicated...).
These are very close to 10:1, so they are being points-deprived by close to 20%. They just seem to be *really good* at doing the BRP4 Arecibo work.
But all of this is very preliminary because useful conclusions can only really be drawn after we're sure all of the other work has cleared from every machine.
Then my observation is that if 500-series cards are paid to parity, everything else is going to be overpaid by as much as the 500s seem to be underpaid.
I suspect all of this has to do with the efficiency of the application on any given card, which further complicates things. I'll bet the 500s would do much better if they could use CUDA 4.2, and I'm SURE the 600s would do better using CUDA 5.
There's apparently nothing we can do about that.
So, what do you suggest that would make everyone happy?
I'm sure we'll get to a good number and Bernd seems completely willing to work this out as best it can be worked-out.
Time keeps flowing like a river, to the sea.
I'm going to let this rest for a while until I have more reliable data. Since I'm a "sometimes" participant, I really don't even want to have any influence, but I have enough very different machines that I don't mind providing data if that will help the discussion.
RE: [ But here's what
)
During my small test on Albert I got a ratio of 10.8 : 1 for my GTX580
RE: RE: Don't know about
)
Of GPU projects, only seti is using creditnew AFAIK. The WCG GPU app is now defunct. Several GPU projects tried creditnew but abandoned it quickly.
Hi Richard! RE: As I
)
Hi Richard!
What I meant here (and I'm almost sure that's what I wrote you at the time) was: Which projects do actually grant credit using CreditNew?
This is precisely the reason why I asked that question more publicly here. More readers should know about more projects, right?
BM
BM
archae86 wrote:I think it
)
I agree that other programs like antivirus and so on may also influence runtime of WUs. Crunching different projects or 2 WUs on GPU in parallel as well.
After WUs were downloaded by boinc manager the estimated runtime is calculated and shown. I can observe that the estimated runtime for not started BRP4 WUs vary from 20 minutes up to 120 minutes. Sure thats not the real runtime for a WU but it is good indicator were the real runtime might end.
Do you have an explanation for this?
Edit: My computer is staffed with an i7-3930k and 64GB RAM and Radeon HD7850 with 2GB VRAM.