Join PrimeGrid
Returning Participants
Community
Leader Boards
Results
Other
drummers-lowrise
|
Message boards :
Cullen/Woodall prime search :
CW sieve now running faster?
| Author |
Message |
|---|
|
|
|
Hi all,
Just a quick note - all the CW sieves i've downloaded since yesterday evening have been completing around 4-5% faster than before. I've changed nothing on my PC in that time (no reboot or graphics driver updates) and the sieves always finish within a fairly narrow window of time so the increase in speed is quite noticeable. I've run this subproject for many weeks now without seeing any change to runtimes in this period, so was wondering if anyone else noticed a similar thing or is it just me? Has there been any code changes at all? Just mildly curious as to how to explain this...
Cheers,
Darryl
____________
 | |
|
Ken_g6 Volunteer developer
Send message
Joined: 4 Jul 06
Posts: 940
ID: 3110
Credit: 261,910,272
RAC: 18,715
                            
|
|
CW Sieve on nVIDIA uses the same basic algorithm as PPS Sieve, so it scans a large number of N's == a large number of bits at a time. But due to the size of the K's used here, it can't jump as far: about 22-24 bits last I recall. As the P's we test against get bigger, so too does the jump size, one bit/N at a time. I'm guessing that's what happened here.
These are rare and getting rarer - getting to each sequential jump size takes about twice as long as the previous one did, minus the small speed increase. PPS Sieve is long past a 32-bit upper limit, so speeds there don't jump like this anymore; but they used to.
____________
| |
|
|
|
|
Was just wondering if the sieves have changed in size or something. Once again the run times have dropped. I was running them between 171 - 172 seconds and now I can run them between 163 - 164 seconds. That's a pretty big drop in time.
____________
@AggieThePew
 | |
|
 rroonnaallddVolunteer developer
Volunteer tester
Send message
Joined: 3 Jul 09
Posts: 1213
ID: 42893
Credit: 34,634,263
RAC: 0
               
|
|
The app was not changed. Maybe we have reached a higher search range...
____________
Best wishes. Knowledge is power. by jjwhalen
 | |
|
|
|
|
Are we still finding factors by sieving? I haven't found a factor in a really long time sieving these 24/7 | |
|
|
|
Are we still finding factors by sieving? I haven't found a factor in a really long time sieving these 24/7
We are, but not very frequently. In the first week of 2012 i completed 4200 sieves and found 3 factors. I'm guessing the optimal depth is somewhere around 1 factor per week or so - each LLR test for the numbers being seived now would probably take at least 5 or 6 days and would still need double checking (so double that time).
Still, i can't imagine this subproject has many more months left...
____________
| |
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 14011
ID: 53948
Credit: 430,128,819
RAC: 1,087,180
                         
|
Are we still finding factors by sieving? I haven't found a factor in a really long time sieving these 24/7
We are, but not very frequently. In the first week of 2012 i completed 4200 sieves and found 3 factors. I'm guessing the optimal depth is somewhere around 1 factor per week or so - each LLR test for the numbers being seived now would probably take at least 5 or 6 days and would still need double checking (so double that time).
Still, i can't imagine this subproject has many more months left...
Except that "the end" is a target that keeps moving...
Don't forget that if you make the sieve faster -- ironically, that extends the lifetime of the sieve rather than shortening it.
Imagine that the next gen GTX 6xx cards come out in a few months and are 25% faster than the 5xx cards. That means that sieving just got 25% more efficient as compared to doing LLR tests on the CPU. Thus, new life gets breathed into the sieving life expectancy.
Conversely, the new gwnum libraries 16 months ago sped up the CPU LLRs by about 10%, and this hastened the end of the PSP/SoB sieving project.
Things seem to work backwards here. Make sieving faster, and rather than getting to the end quicker, you get there slower because the end moves further away.
____________
My lucky number is 75898524288+1 | |
|
JohnHonorary cruncher
Send message
Joined: 21 Feb 06
Posts: 2875
ID: 2449
Credit: 2,681,934
RAC: 0
               
|
Conversely, the new gwnum libraries 16 months ago sped up the CPU LLRs by about 10%, and this hastened the end of the PSP/SoB sieving project.
Anyone want an extra 20%-50% increase in LLR speed. ;) If so, stay tuned...
p.s. You might want to get your CW Sieve badge sometime this year...or sooner.
____________
| |
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 14011
ID: 53948
Credit: 430,128,819
RAC: 1,087,180
|
Conversely, the new gwnum libraries 16 months ago sped up the CPU LLRs by about 10%, and this hastened the end of the PSP/SoB sieving project.
Anyone want an extra 20%-50% increase in LLR speed. ;) If so, stay tuned...
p.s. You might want to get your GCW Sieve badge sometime this year...or sooner.
AVX?
____________
My lucky number is 75898524288+1 | |
|
JohnHonorary cruncher
Send message
Joined: 21 Feb 06
Posts: 2875
ID: 2449
Credit: 2,681,934
RAC: 0
|
Conversely, the new gwnum libraries 16 months ago sped up the CPU LLRs by about 10%, and this hastened the end of the PSP/SoB sieving project.
Anyone want an extra 20%-50% increase in LLR speed. ;) If so, stay tuned...
p.s. You might want to get your GCW Sieve badge sometime this year...or sooner.
AVX?
That is correct. Please see: AVX build of llr (20%-50% faster)
____________
| |
|
|
|
Imagine that the next gen GTX 6xx cards come out in a few months and are 25% faster than the 5xx cards. That means that sieving just got 25% more efficient as compared to doing LLR tests on the CPU. Thus, new life gets breathed into the sieving life expectancy.
Michael, I wonder how do you compare a GPU to a CPU? Is it like you compare i7 to GTX580? What about more threads in a CPU?
And how is LLRing compared to sieving? Is there a test sieve that one can understand how deep (say "T" or "P") sieving progresses in an hour, day, week at a specific "n"?
Again, comparing a CPU to a GPU we need to have something in common. Maybe power consumption? How much progress is done on one watt consumed?
Looks tricky to me )))
P.S. Also, if we stop sieving on PPS and Woodall then we should start other sieve projects or cruchers will go to other projects and everyone looses. Even if it's not very effective to sieve it's better sieve than not, because many crunchers don't care about the science behind but care more about the credits and fun. | |
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 14011
ID: 53948
Credit: 430,128,819
RAC: 1,087,180
|
Imagine that the next gen GTX 6xx cards come out in a few months and are 25% faster than the 5xx cards. That means that sieving just got 25% more efficient as compared to doing LLR tests on the CPU. Thus, new life gets breathed into the sieving life expectancy.
Michael, I wonder how do you compare a GPU to a CPU? Is it like you compare i7 to GTX580? What about more threads in a CPU?
And how is LLRing compared to sieving? Is there a test sieve that one can understand how deep (say "T" or "P") sieving progresses in an hour, day, week at a specific "n"?
Again, comparing a CPU to a GPU we need to have something in common. Maybe power consumption? How much progress is done on one watt consumed?
Looks tricky to me )))
P.S. Also, if we stop sieving on PPS and Woodall then we should start other sieve projects or cruchers will go to other projects and everyone looses. Even if it's not very effective to sieve it's better sieve than not, because many crunchers don't care about the science behind but care more about the credits and fun.
If "you" is meant to mean me, it's very simple -- I just compare one core of my fastest CPU to my fastest GPU -- neither of which are particularly fast. It's simplistic, but usually what I'm concerned about personally is "how fast can I finish this?" Note that this is different than "how much can I get done?", which is why I only look at a single core.
As for power consumption, I generally only crunch in the winter, so the power heats the house. Efficiency (say, credits vs. watts) is irrelevant when waste heat is useful.
But, if "you" means PrimeGrid, John will have to answer. There's lots of factors that go into the decision. For example, Sieving for GFN at N=4194304 will likely stop just above 9000P. You could do comparisons of CPU Sieving speed vs. GPU PRP speed,, or power consumption analysis, capital depreciation, or whatever. But in that case, it's simply the point where the sieve program will stop working.
____________
My lucky number is 75898524288+1 | |
|
|
|
|
So, we need a quick new calculations for CW Sieve project. ;) | |
|
Message boards :
Cullen/Woodall prime search :
CW sieve now running faster? |
