LMH goals...

[Xyzzy]

I have been wondering about what we are working towards...

My guess is to have exponents factored completely before PrimeNet begins giving them out, but to do that wouldn't we all work in very narrow ranges right above where PrimeNet is now?

Or is the goal just to find as many factors as possible? If so, what meaning does that have in the overall scheme?

I also wonder about some of those huge ranges that are out... Wouldn't it make sense to hand out ranges of at most 500,000 exponents? Right now there is almost nothing to give out...

As it is, someone factoring to 59 or 60 bits shaves just a minute or two off of a factoring job... Factoring to 68 or whatever would be much more useful... But if we all work spread out over 50 million exponents getting to 68 bits is never going to happen in time... I have a box here that is probably one of the fastest factoring boxes you can get and it is going to take me 9 days to take a 500,000 range to 2^59, and when it is all over that work will save the average cruncher 1-2 minutes?

I'm just wondering... Please don't jump on me too hard... :)

[nitro]

Originally Posted by Xyzzy

I have been wondering about what we are working towards...

My guess is to have exponents factored completely before PrimeNet begins giving them out, but to do that wouldn't we all work in very narrow ranges right above where PrimeNet is now?

When I first got started in this sub project - about 18 months to 2 years ago, it was just because we wanted to be a little different.

Originally Posted by Xyzzy

Or is the goal just to find as many factors as possible? If so, what meaning does that have in the overall scheme?

That sums it up pretty nicely, I have found around 150,000 or so factors.

Originally Posted by Xyzzy

I also wonder about some of those huge ranges that are out... Wouldn't it make sense to hand out ranges of at most 500,000 exponents? Right now there is almost nothing to give out...

I guess it is a case of get here early and stake a claim. Bit like the gold rush really. As someone who has one of those "huge" ranges, I have put in an awful lot of work on that range, basically from 32 bits up to now complete to 58 bits. 60-64m is complete to 59 bits and 60-61 is at 60 bits.

I can't stop anyone trampling all over the rest, but it would really tick me off. I am stopping at 60 bits, after that do what you like.

Originally Posted by Xyzzy

As it is, someone factoring to 59 or 60 bits shaves just a minute or two off of a factoring job...

That depends on the speed of the machine, but the real gain comes when you do find a factor. Keep on reading...

Originally Posted by Xyzzy

Factoring to 68 or whatever would be much more useful... But if we all work spread out over 50 million exponents getting to 68 bits is never going to happen in time...

It will all need to be done *eventually*, but given that first time checks are only just past 20m it is not going to be a pressing problem for a *long* time.

Originally Posted by Xyzzy

I have a box here that is probably one of the fastest factoring boxes you can get and it is going to take me 9 days to take a 500,000 range to 2^59, and when it is all over that work will save the average cruncher 1-2 minutes?

Right, so you have the fastest Athlon chip available. Now lets do some quick and dirty calculations.

In a range of 500,000 let's say there are 12500 candidates (in 64-64.5m there are 12989). So total time taken to test is roughly 17 days at 2 mins per test.

Factors should be found roughly one in every 59 numbers tested, so that means about 212 factors found.

Now let's nip over to the gimps home and look at the status page. In this range there are currently 251,447 exponents that are "unknown", with an expected cpu test time of 6,707,276 years. Dividing one by the other gives an average of 26.67 cpu years per exponent.

So, you found on average 212 exponents with a factor, you just "saved" the project 212 * 26.67 = 5654 cpu years of effort.

Originally Posted by Xyzzy

I'm just wondering... Please don't jump on me too hard... :)

The above maths tells you why I got into doing this. My 150,000 factors have saved the project around 4 MILLION cpu years of testing.

It has long been a gripe of mine that we don't have a third top producers chart, this time ranked by who has contributed most to the project by *shortening* it.

It was and still is just about the only way that us small guys (half a dozen home boxes) can compete against and trounce the big guys with hundreds of university computers to lay with.....

[trif]

I suspect the goals are highly individual. The collective goal of LMH is to not have work duplicated inadvertently.

For the huge ranges, those were reserved back when LMH had fewer participants. If we run short on space for new people coming in, I suspect those having large assignments would be quite willing to break them up, but until we run out of other space, it hurts nothing to have larger assignments. If someone has a particular interest in working space that someone else has "reserved" I'm sure something could be worked out.

As for my personal interest, there is the fact that Primenet only gives credit for the last four bit levels, so I work to a standard that aims for an ending level of four less than the final bit level. Also some people have machines that are just dogs for doing anything other than stuff under 62 bits. There is also the fascination of watching a large range, and seeing the tantilizing patterns of where factors are found.

[cheesehead]

Originally Posted by Xyzzy

As it is, someone factoring to 59 or 60 bits shaves just a minute or two off of a factoring job...

... just a minute or two off each of several tens of thousands or hundreds of thousands factoring jobs, that is.

500,000 minutes = ~ 1 year.

Factoring to 68 or whatever would be much more useful...

Well, the total amount of trial factoring work to be done to, say, 68 bits for the remaining Mnumbers in a given range is the same no matter how you break it up, so it's not a matter of usefulness in that sense (disregarding variations in relative TF efficiencies by various systems).

(By total amount of TF work, I mean the work needed to reach the upper limit or find a factor, whichever occurrs first.)

But if we all work spread out over 50 million exponents getting to 68 bits is never going to happen in time...

It's going to take the same amount of time either way (disregarding variations ...).

OTOH, your perception of the progress may differ between horizontal progress and vertical progress. Try looking at the chart while reclining or lying on your side. ;)

I have a box here that is probably one of the fastest factoring boxes you can get and it is going to take me 9 days to take a 500,000 range to 2^59, and when it is all over that work will save the average cruncher 1-2 minutes?

It'll save 500,000 average crunchers 1-2 minutes each, for a grand total of a year or two.


Clearly, you're happier taking a few exponents deep than taking a lot of exponents shallow ... so why not just do that? Instead of reserving, say, 41M-42M to 2^61, reserve 41M-41.004M to 2^68.

Of course, whoever keeps track of LMH reservations may consider ranges of 0.004M to be a hassle.

(hmmm ... wonder if that's why LMH has been doing broad and shallow rather than narrow and deep ... no automated reservation system?)

[Xyzzy]

Okay, it makes more sense now...

I'm not sure how the math works that says that 4 million years of work was saved, since those factors would be found anyways, but I'll take your word for it since this is all way over my head...

No matter what, it is obvious that this project is popular and fun and that is what is most important...

(Perhaps someone could add a more in-depth mission statement to the sticky thread?)

Thanks!

[apocalypse]

I'm running LMH on my work machine so I can help the project with ZERO chance of winning money in the process. That would only confuse and annoy my employers. Even with permission to run the software, I'm not allowed to earn anything on personal projects with company equipment.

I do like chalking up the factors, too. I'm relatively new so I only have a couple hundred.

Maybe we aren't really saving work, but at least we're tightening the work remaning estimate. :(

[antiroach]

it would be pretty cool if we organized something to take a small range (maybe 1M) upto the "required" bit depth. Instead of reserving a 1 million range per person each person could take .1 or .05 million and take it to a certain depth. like non p4 users could take the range upto 64 bits, then we could have the p4 users take that range past the 64 bit mark. Its cool to find factors for like exponents of 70+million but these factors wont really save any work in the near future (on my box it would take 330days to LL a 79.3M exponent :) ) By finishing some smaller exponents to the required bit depth would save the people running LL tests on these exponents some time. I have an athlon and I would hate having to factor a # to 68bits, when a p4 could do so much faster (well above 64 bits). It would be also a cool idea to maybe start a p-1 thread so that when these exponents get to 1st time LL's they would have had p-1 done also and that would save people some more time as well. also dedicated p-1'ers probably have a decent amount of ram (256+ megs plus allotted to prime95). Some people (newbies) may not know how much ram to allot to prime95 or they may not have enough to do p-1 (im talking mainly about stage2). I think I may be repeating myself in this post but It took me several minutes to write this thing and i'm too tired to re-read / re-write this thing. It's pretty late. If I dont make any sense please let me know :)

[lycorn]

What you say makes actually a lot of sense.
I have been participating in LMH for fun, and as a way to make use of some old boxes, too slow for relevant LL testing.
As for the point of whether it is useful or not to test large exponents in a broad and shallow manner, I think we should look at the project *as a whole*, a pool of 4,630,913 prime numbers to test, with exponents ranging from 2 to 79.3M. So it is as useful to find a 5-bit as a 73-bit factor. It will save twice the time spent on the LL test, be it in the near or far future.
Coming to your point, I agree it would be interesting to organize the search so as to avoid having machines performing work they are not really fit to. Particularly, using P3/Athlons to TF to 64 bits would be beneficial in terms of overall throughput.
But LMH is not automated. It is a "handcrafted" project, so detailed coordination isn´t that easy. Also, the mix between Primenet and LMH would be very hard, if not at all impossible to manage. I mean situations where for example exponents in the 33,7M range were handed over to LMH users to be TFed and/or P-1ed before getting into the Primenet pool, but then the user is not catching up with Primenet´s leading edge. In that case, either the work would be lost because Primenet assigned the exponents, or we would start having "holes" in Primenet assignments, which I´m afraid would be pretty messy. So I think LMH shall keep a "safe distance" from Primenet´s leading edge.
In a future release of the server software, these forms of assigning exponents should be considered. E.g. p-1 as a separate assignment, TFing to levels lower than the default one.
For now, it would not be practical to manage hundreds of small ranges (garo knows what I´m talking about... ). And, even the way we are doing it now, *every* bit of LMH work is useful to GIMPS, which is our common goal.

[jocelynl]

Xyzzy


LMH is generally the place for people with very slow machine. It is now the only way they can contribute to the search. A small contribution to a very long project. I don't think anybody would try LL at 20M on a 100Mhz.
It would take forever.

[garo]

As being the coordinator of the reservation thread I for one certainly do not relish the prospect of managing 0.05 million sized ranges. I think LMH works well as it is and lycorn has raised many valid points that rule out a major change in the nature of the effort. In any case, Primenet always sends out assignments in a way that all factoring and P-1 is done before the LL test (P-1 only if the machine is capable). So if LMH misses out stuff it's no sweat! Duplicating work due to overlap is much worse.

Futhermore, LMH was really designed to allow people with slow machines like 100 or 200MHz computers to contribute. And perhaps for faster machines that have no net connection as with LMH you do not even need to connect once in 60 days like on Primenet. If you have spanking new P4, gosh even a 1.6MHz P4 why join LMH? Get your work from Primenet!! Less work for you and me!

In the long run, I'd really just like to see all exponents done to 60 bits and then 62 bits before they get to Primenet so that P4s do not waste time. I'd also like people to take 1-3 months worth of work at time. Run a trial exponent or two if you are a first timer to get a sense of how long your range will take.

[cheesehead]

garo, have you been coordinating the LMH reservations all along, or was there a predecessor? (I didn't pay much attention when the Yahoo forum was going, though I browsed occasionally.)