Anyone factoring <5M?

[Prime95]

Originally Posted by cheesehead

Another incorrect recollection?

Will Edgington made an effort. We corresponded by email to nail down the bug's impact as well as the ranges thought to be affected. He then used his own program to cover the gaps. This was likely a fairly error-prone process.

The ranges were never sent out to be trial factored again by mainstream prime95 clients.

[cheesehead]

Poacher call-out:

"GrunwalderGIMP" poached my TF assignment on 1528291

I had 1528291 reserved through PrimeNet for TF from 61 to 64.

While I had it reserved, "GrunwalderGIMP" reported no factor found from 61 to 63.

When I reported no factor found from 61 to 64 just now, PrimeNet informed me that the first two levels were not needed, and gave me no credit for them.

Note: I don't care about the credit for myself! (Please don't bother to credit me for it, George.)

Poachers like "GrunwalderGIMP" not only

(a) steal credit from other folks, some of whom may care about their PrimeNet credits,

but also

(b) waste the time of folks who follow the rules. I do care about that.

I could've run some other PrimeNet assignment instead of doing those first two levels of TF, if "GrunwalderGIMP" had had the courtesy to register with PrimeNet.

[NBtarheel_33]

Originally Posted by cheesehead

Poacher call-out:

"GrunwalderGIMP" poached my TF assignment on 1528291

I had 1528291 reserved through PrimeNet for TF from 61 to 64.

1528291? From 61 to 64? Wouldn't ECM be a better choice on a "small" exponent like that - seems like 61->64 would take forever.

[cheesehead]

Originally Posted by NBtarheel_33

1528291? From 61 to 64? Wouldn't ECM be a better choice on a "small" exponent like that

It depends on one's goal, and the relative balance of how far each method has been used.

Unlike P-1 or ECM, TF is exhaustive to a limit that is simply expressed as a single number, and can be continued without duplication of effort -- (* ahem *) barring poaching -- without having a previous save file (except that one may consider the record of previous power-of-two reached to be a tiny save file). I wanted to exhaustively extend factoring on that exponent. I was not trying to be efficient in any other sense.

P-1 is exhaustive, but to multidimensional limits which are more complicated than a single number. Also, someone who wants to extend P-1 limits from the recorded previous highs has to either duplicate the previous computation up to the previous limits, or have a save file from the previous attempt.

ECM may indeed be the most probabilistically efficient method for factoring, depending on how far other methods have been used, but it isn't deterministically exhaustive.

- - -

Hmm ... I had intended to post about the 1528291 poach in a different thread: "Anyone factoring <5M" at http://www.mersenneforum.org/showthread.php?t=13302 It's really OT here.

Moderators: how about moving these three 1528291 posts to that other thread?

:-)

[10metreh]

I believe GrunwalderGIMP is Graff on this forum. PMing him might be an idea.

[cheesehead]

Thank you. I've sent a PM to Graff asking to confirm whether he is "GrunwalderGIMP", and explaining the duplication.

[cheesehead]

Update: Graff confirmed that he's "GrunwalderGIMP", and explained:

"I grabbed a load of exponents in the 1.5 M range many months back for LMH work. When I tried to register the exponents with the server, it wouldn't let me ... I will check to see if other exponents I currently have are registered to others and will try again to register them if not."

and

"I see that I can register the exponents by adding them to a worktodo.txt file, but that is cumbersome for the number of exponents I work. I'm looking at using curl to register the exponents to prevent future duplication of effort."

I requested of Graff: "If you work out a reliable semi-automated method for registering large numbers of exponent assignments while avoiding duplication, please post a description in the LMH subforum!"

[Graff]

Originally Posted by cheesehead

I requested of Graff: "If you work out a reliable semi-automated method for registering large numbers of exponent assignments while avoiding duplication, please post a description in the LMH subforum!"

Well, my attempt to get all the grabbed 1.5M exponents registered to my
LMH work appeared to work. I registered ~1400 exponents on May 12
using a custom DCL command file that made calls to the server via a simple
curl line.

I just checked my assignments and see that all but four of the 1.5M
assignments have disappeared! I think I know why this happened. Since
this work is "manual testing", there is no machine hardware id that the
assignments can be tied to by the server. So I used the program
self-assigned permanent id from one of my PCs that gets work from the
server. But when that machine did its daily update of completion dates,
the 1.5M exponents were released since it didn't have a record of them
in its worktodo.txt file.

I am currently reregistering the exponents using the self-assigned permanent
id from a machine that no longer reports results. The reregistration should
be completed shortly. I will check again in a few days to make sure that
they are still registered.

Of course, the ideal situation would be to register a "virtual machine"
that would have a id that I could use to register these LMH results.
(Reporting of results is no problem as I use the manual forms for that.)
But I have no yet figured out how to achieve this (I suspect that it
is not possible), as every attempt complains about missing security
hashes, which the on-line documentation implies are optional but that
seem to be mandatory for assignment of new machines.

Gareth (Graff/GrunwalderGIMP)

[Graff]

Originally Posted by Graff

I will check again in a few days to make sure that
they are still registered.

Hmm. They've disappeared again. They were there when I checked
this morning (I think that was the last time I checked). I'm reporting
some more results via the manual reporting forms.

Graff

[cheesehead]

Originally Posted by petrw1

We know that the factoring limits are set based on relative time and potential effectiveness vs LL. http://www.mersenne.org/various/math.php

Yes, and let's remind ourselves that that's versus LL. That is, the TF limits coded into prime95 are for the crossovers between:

A) efficiency of eliminating (re: M^p primeness) a candidate mersenne exponent via finding a factor with TF

and

B) efficiency of eliminating (re: M^p primeness) a candidate mersenne exponent via L-L test.

They were not intended to be the crossovers between:

A) efficiency of finding a factor via TF

and

C) efficiency of finding a factor via ECM

Is there a reference table for the latter comparison anywhere?

I also know that when I try to assign low exponents (mind you, lower than yours) to higher levels of factoring I get an error message something like: "Invalid assignments use ECM instead".

This is because V25 module commonc.c has a hard-coded lower limit of 20,000 for exponents on which it will allow TF. When a worktodo line specifies TF for lower exponents, V25 commonc.c writes the "Error: Use ECM instead of trial factoring for exponent: nn" message.

        if (w->work_type == WORK_FACTOR && w->n < 20000) {
        char    buf[100];
        sprintf (buf, "Error: Use ECM instead of trial factoring for exponent: %ld\n", w->n);
        OutputBoth (MAIN_THREAD_NUM, buf);
        goto illegal_line;
        }

I have ASSUMED from this error that ECM would be more efficient.

Well, at least it means George:

(1) wants folks to do ECM instead of TF on exponents less than 20000 with V25 prime95

and

(2) deprecates V25 worktodo Factor= lines specifying exponents less than 20000 as "illegal".

My guess is that he has indeed figured that ECM is more efficient than further TF for the mass of exponents below 20000.

But there are more important things in life than efficiency ... such as one's obsessive-compulsive drive to TF low exponents to higher powers-of-two, or one's perfectionistically noting that TF is deterministically certain to find factors in a given range, whereas ECM is only probablistic. (I was referring to TF the method there, not necessarily to TF the as-implemented-in-software-by-fallible-programmers code.)

V24 commonc.c did not have this hard-coded restriction. So ... if one were to set up a separate folder for V24 prime95, put the V24 executable in it, and create the proper V24 .ini files (paying attention to the earlier worktodo command line formats), one could have V24 prime95 do TF on exponents under 20000. One would need to remember to use FACTOR_OVERRIDE to specify the top TF level, and not to try TFing exponents to different top levels using the same FACTOR_OVERRIDE setting!

(Hmmm... could someone persuade George to add an undocumented TF_EXPONENT_LOW_LIMIT_OVERRIDE parameter, with default value of 20000 or something like that, to some future version?)

On a multi-CPU system, one could even have both V24 and V25 chugging away separately without bothering each other, as long as one first went to the trouble of modifying one's V25 prime95 local.txt, prime.txt and worktodo.txt parameter files so as to avoid using CPU 0. (One would, of course, have first saved copies of those files, with names such as local_for_V25_when_no_V24_is_running.txt, wouldn't one? And after the modifications, one would save copies with names such as local_for_V25_when_separate_V24_is_on_CPU_0.txt, wouldn't one?)

Then all one would have to do is to solve the problem of how to get V5 PrimeNet to register below-20000 TF assignments. Would a simple manual communication from a V24 prime95 suffice? Who knows?

[garo]

Originally Posted by cheesehead

But there are more important things in life than efficiency ... such as one's obsessive-compulsive drive to TF low exponents to higher powers-of-two, or one's perfectionistically noting that TF is deterministically certain to find factors in a given range, whereas ECM is only probablistic.

How do you square that with your argument that poaching is inefficient? Other people might have obsessive compulsive desires to clean up the trailing exponents. Everything under M39 doublechecked and all that.