Help needed for doublecheck sieving!

[mdettweiler]

Originally Posted by gd_barnes

In theory, since the rate of sieving varies by the square root of the # of k's, it should sieve sqrt (500/351)-1 = ~19.4% slower. But it will find more factors / P-range.

With the efficiency of more k's, sieve depth should increase slightly. After I said it, 1T seemed a little high but it might be real close with 500 vs. 351 k's. It's very unusual to sieve to 1T for a non-top 5000 range but then again, it's very unusual to sieve 500 k's at once too!


G

Oh, I see, yeah. I forgot for a moment there that sieve speed does not increase/decrease linearly, but instead by square root!

As for sieve depth: okay, cool. One thing's for sure: when we're done with this, we're going to have some REALLY well-sieved ranges!

BTW, speaking of those well-sieved ranges, when we are finally ready to LLR them, what should we call the resulting team drive? "Doublecheck Drive #1"? Or just plain old "Team Drive #3" (or #4, if #3 has already started by then)? I'm more inclined towards "Doublecheck Drive #1", but I'd like to get everyone else's opinion on this first. Anyway, regardless of the name, it will be similar to any other team drive, except for one important thing: it will be limited to known-stable, non-overclocked machines, since we don't have any first-pass residuals to compare our results with, and thus we have to rely on our doublecheck results completely. (If someone's got a machine that's overclocked, however, but not too heavily, and they've confirmed it to be fine with a Prime95/mprime stress test, we might be able to let it help out--any thoughts on this, please?)

[Flatlander]

Taking 130G - 270G

[gd_barnes]

Originally Posted by Anonymous

Oh, I see, yeah. I forgot for a moment there that sieve speed does not increase/decrease linearly, but instead by square root!

As for sieve depth: okay, cool. One thing's for sure: when we're done with this, we're going to have some REALLY well-sieved ranges!

BTW, speaking of those well-sieved ranges, when we are finally ready to LLR them, what should we call the resulting team drive? "Doublecheck Drive #1"? Or just plain old "Team Drive #3" (or #4, if #3 has already started by then)? I'm more inclined towards "Doublecheck Drive #1", but I'd like to get everyone else's opinion on this first. Anyway, regardless of the name, it will be similar to any other team drive, except for one important thing: it will be limited to known-stable, non-overclocked machines, since we don't have any first-pass residuals to compare our results with, and thus we have to rely on our doublecheck results completely. (If someone's got a machine that's overclocked, however, but not too heavily, and they've confirmed it to be fine with a Prime95/mprime stress test, we might be able to let it help out--any thoughts on this, please?)

I'm not sure we would need to call it a drive. Perhaps an 'effort'. I would only consider this to be a lower priority behind getting all 300<k<=1001 searched from n=260K to ~400K. As long as we're staying about n=50K above the drop-off point of top-5000, I'm good with a coordinated effort to put large resources on this for periods of time. The goal that I initially had for this was completion sometime in 2010. Since we're striving to complete to n=600K a year earlier than its original goal, then it makes since to set a goal for this for completion sometime in 2009.

Yes, only non-overclocked machines can be used for a double-check effort in which we have no residues to compare against. Or as you alluded to, if someone has a slightly overclocked machine and they can 100% say that it is virtually dust-free and running at a good core temperature, then I'm good with that.


Gary

[mdettweiler]

Originally Posted by gd_barnes

I'm not sure we would need to call it a drive. Perhaps an 'effort'. I would only consider this to be a lower priority behind getting all 300<k<=1001 searched from n=260K to ~400K. As long as we're staying about n=50K above the drop-off point of top-5000, I'm good with a coordinated effort to put large resources on this for periods of time. The goal that I initially had for this was completion sometime in 2010. Since we're striving to complete to n=600K a year earlier than its original goal, then it makes since to set a goal for this for completion sometime in 2009.

Yes, only non-overclocked machines can be used for a double-check effort in which we have no residues to compare against. Or as you alluded to, if someone has a slightly overclocked machine and they can 100% say that it is virtually dust-free and running at a good core temperature, then I'm good with that.


Gary

Oh, okay. I was thinking that "Doublecheck Drive #1" sounded a little more "catchy", especially considering that this is hardly the last doublecheck effort this project will be doing. However, in the first post of the thread, I could still make it clear that it is much lower priority than getting both 300<k<400 and 300<k<1001 searched past the "danger zone" at the bottom of the top-5000.

[gd_barnes]

Originally Posted by Anonymous

Oh, okay. I was thinking that "Doublecheck Drive #1" sounded a little more "catchy", especially considering that this is hardly the last doublecheck effort this project will be doing. However, in the first post of the thread, I could still make it clear that it is much lower priority than getting both 300<k<400 and 300<k<1001 searched past the "danger zone" at the bottom of the top-5000.

That works.

When we're done with this and if we're at n=600K on the main drives at that point, we will have tested all 351 k's from n=1 to 600K as well as 150 k's from n=1 to 260K!

Going forward from there, I would probably see a slowly progressing and ongoing double-check effort that stays around n=300K-400K behind the current testing limit. Once we get things caught up to where they should be in the scheme of things, then nice and steady works well.

The idea with double-checking is that it is done with machines that are far faster than the machines that were originally used to do the searches because it historically only corrects 0.5%-2% of the data. My hope is at least 2X as fast. For n=1 to 260K, I'm sure it's more like 4X as fast if people are using their quads since many of the ranges were completed to n=260K by 2005.


Gary

[mdettweiler]

Originally Posted by gd_barnes

That works.

When we're done with this and if we're at n=600K on the main drives at that point, we will have tested all 351 k's from n=1 to 600K as well as 150 k's from n=1 to 260K!

Going forward from there, I would probably see a slowly progressing and ongoing double-check effort that stays around n=300K-400K behind the current testing limit. Once we get things caught up to where they should be in the scheme of things, then nice and steady works well.

The idea with double-checking is that it is done with machines that are far faster than the machines that were originally used to do the searches because it historically only corrects 0.5%-2% of the data. My hope is at least 2X as fast. For n=1 to 260K, I'm sure it's more like 4X as fast if people are using their quads since many of the ranges were completed to n=260K by 2005.


Gary

As for doublechecking in the future--once we reach n=260k for the doublechecks, we'll have residuals to compare for 300<k<1001. Thus, even somewhat unstable machines can help out with doublechecking, since even if they produce bad residuals, we'll simply notice that when we compare first-pass and doublecheck residuals, and mark those particular numbers as needing to be done a third time.

Thus, once we start doublechecking in the range that we have first-pass residuals for, we might want to set up an LLRnet server for doublechecking. (No, Carlos, don't worry--we know you can't handle any more LLRnet servers. We'd find someone else to host it, in the case that you wouldn't be able to.) It would probably be a great way to encourage people to do doublechecking work, since otherwise it's decidedly less appealing than first-pass tests.

[Kman1293]

Taking 270G-300G

[gd_barnes]

Originally Posted by Kman1293

Taking 270G-300G

Welcome to the effort Kman! Thanks for helping.

[Flatlander]

Originally Posted by gd_barnes

Yes, only non-overclocked machines can be used for a double-check effort in which we have no residues to compare against. Or as you alluded to, if someone has a slightly overclocked machine and they can 100% say that it is virtually dust-free and running at a good core temperature, then I'm good with that.
Gary

That's me out then. My C2D is overclocked about 70% and running at 72C. It hadn't occurred to me that I could be missing primes/getting false residues. My prime finding rate looks about average though, and I've never had any errors or false positives. Not sure if I've got the patience to run it at stock.

[mdettweiler]

Originally Posted by Flatlander

That's me out then. My C2D is overclocked about 70% and running at 72C. It hadn't occurred to me that I could be missing primes/getting false residues. My prime finding rate looks about average though, and I've never had any errors or false positives. Not sure if I've got the patience to run it at stock.

Don't worry, stable overclocked machines are still welcome. The only requirement is that you run a Prime95/mprime stress test for a few hours to see whether your machine will produce good residuals (it probably will).

Please note, however, that even machines that produce bad LLR residuals will still do just fine with sieving.

[Flatlander]

Originally Posted by Anonymous

Don't worry, stable overclocked machines are still welcome. The only requirement is that you run a Prime95/mprime stress test for a few hours to see whether your machine will produce good residuals (it probably will).

Please note, however, that even machines that produce bad LLR residuals will still do just fine with sieving.

Excellent!