Help needed for doublecheck sieving!

[henryzz]

why isnt there a program that dishes out sieving assignments like the lrrserver dishes out llr

[mdettweiler]

Originally Posted by gd_barnes

Anon,

What is your P-rate per second at P=300G on this sieve?


Thanks,
Gary

114247 p/sec.

[mdettweiler]

Originally Posted by henryzz

why isnt there a program that dishes out sieving assignments like the lrrserver dishes out llr

I tried writing something like that a few weeks ago in Perl, but I have nowhere near the programming skills or knowledge to be able to do it.

Maybe Geoff could try to put something together, since he wrote all the sr(x)sieve programs?

[henryzz]

maybe the ecmserver could be modified
that might be easier

[mdettweiler]

Originally Posted by henryzz

maybe the ecmserver could be modified
that might be easier

Some things probably could be borrowed either from ecmserver or LLRnet, but most of it would likely have to be written from scratch, because sieving is handled in a fundamentally different way than LLR, ECM, etc. Sieving deals in ranges of p working on files with lots of candidates, whereas LLR, ECM, etc. deal with candidates individually.

[gd_barnes]

Originally Posted by Flatlander

For information, my rates are (per core):
160,000 p/sec
9 sec per factor
C2D 3000MHz
(About 4% better when one core is idle.)

OK, then jasong's resevation should take 184000/160000 - 1 = 15% longer than what I estimated above. I'll send him a PM. With 3 people running sieving, the only reason that one person should reserve such a large range is if he can finish it in < 1 week. Otherwise it will end up taking the collective group longer to finish in calendar days even if the total CPU days is the same.

Chris, can you run an LLR test on a candidate somewhere around 500*2^212000-1 on the same machine that you're sieving? Let me know how long it takes. Since core duo's are about equally good at LLRing and sieving, we should be able to get relatively close on an optimal sieve depth.

It's possible that jasong may have even reserved past the optimal sieve depth here. Although I don't think so.

To get the most accurate test at this point, though, we need to remove all the factors up to P=300G and then see what the P-rate/sec. is but we can slightly adjust your P-rate upwards for the estimate.


Gary

[Flatlander]

Originally Posted by gd_barnes

...

Chris, can you run an LLR test on a candidate somewhere around 500*2^212000-1 on the same machine that you're sieving? Let me know how long it takes. Since core duo's are about equally good at LLRing and sieving, we should be able to get relatively close on an optimal sieve depth.
...
Gary

For completeness (and my curiosity):

500*2^100000-1 is not prime. Res64: 34C2F9B0B7467EE7 Time : 7.713 sec.
500*2^100000-1 is not prime. Res64: 34C2F9B0B7467EE7 Time : 7.697 sec.
500*2^212000-1 is not prime. Res64: 41965284C2C61EB0 Time : 33.148 sec.
500*2^212000-1 is not prime. Res64: 41965284C2C61EB0 Time : 33.152 sec.
500*2^260000-1 is not prime. Res64: C54B16F3B30FEAC0 Time : 53.565 sec.
500*2^260000-1 is not prime. Res64: C54B16F3B30FEAC0 Time : 53.579 sec.

I suppose it will also help you decide the size of LLR work units.

[gd_barnes]

Updated optimal sieve depth analysis:

Using Chris's machine:

P-rate @ P=300G: 160K P/sec
LLR 500*2^212000-1: 33.15 secs

Sieve test P=300G-310G:
expectation: 4357.58 factors

P=10G sieve would take: 10G/160K = 62500 secs

1 factor would take 62500/4357.58 = 14.34 secs

Optimal sieve depth:
300G * 33.15 / 14.34 = ~693G

**************************

Using my machine (non over-clocked 1.66 Ghz Dell CD):

P-rate @ P=300G: 56.3K P/sec
LLR 500*2^212000-1: 65.8 secs

(same sieve test and expectation as above)

P=10G sieve would take: 10G/56.3K = 177620 secs

1 factor would take 177620/4357.58 = 40.76 secs

Optimal sieve depth:
300G * 65.8 / 40.76 = ~484G

**************************

Clearly the over-clocking on Chris's machine has much greater impact on the sieving than it does LLR. He's getting 3X the sieve rate and 2X the LLR rate of my 1.66 Ghz Dell CD, resulting in a higher optimal sieve depth.

So my original estimate of P=550G was very close; in between the two. Adding the 150 k's had much smaller impact on the optimal sieve depth than I had speculated. I was way off base in guessing P=1T.

My opinion at this point, let's somewhat average the two depths and round up. Let's make the optimal sieve depth P=600G.

We could make this more exact by sieving to P=500G, removing all factors, and then doing the analysis again but the impact would likely be minimal.

I will contact Jasong via PM, find out if he knows the amount of work he has, and let him know that he can stop at P=600G.

Edit: I have sent a PM to Jasong. I changed the first post here to show an optimal depth of P=600G and to reduce his range.


Gary

[mdettweiler]

300G-310G complete.

[mdettweiler]

Originally Posted by Anonymous

300G-310G complete.

Oh, drats--I just accidentally deleted my factors file for 300G-310G!

I guess that means 300G-310G will have to be done over again. Thus, I'm making that range available again for someone else to take.

[Flatlander]

Originally Posted by Anonymous

Oh, drats--I just accidentally deleted my factors file for 300G-310G!

I guess that means 300G-310G will have to be done over again. Thus, I'm making that range available again for someone else to take.

I guess you've checked your recycle bin?