mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   Marin's Mersenne-aries (https://www.mersenneforum.org/forumdisplay.php?f=30)
-   -   Strategic Double Clicking (https://www.mersenneforum.org/showthread.php?t=20372)

Madpoo 2015-07-22 21:47

Strategic Double Clicking
 
So here's the thought, if you missed the discussion in the "Lounge" thread:

I'm doing an analysis of [I]computers[/I] to see which of them have had a sketchy track record, and return unusually high #'s of bad residues.

I've done my own pilot program, picking some 34M exponents that had only been tested once by machines with spectacularly low good/bad ratios and it's been pretty effective. Off the top of my head I'd say that in the first 15 or so of these that I've targeted, ~12 of them ended up with my result NOT matching that first test by the iffy machine.

Here's the plan then:
I generate lists of exponents that have only been checked once by these suspect machines, and people volunteer to take them on.

George has suggested either doing a manual assignment request which should work okay for exponents above 40M. For smaller exponents (I'll figure out the exact threshold) his suggestion is to just go ahead and add a "DoubleCheck=xx" line to your worktodo and get it going. It'll apparently create the assignment next time it contacts the server (is that correct, George?)

I'm not quite sure the best way to hand out any exponents... I could just post a list here like I was doing for the triple-checking stuff, and let people request blocks of work (first come, first serve)? Or would people prefer if they volunteer for X exponents and I just PM those, to avoid problems like multiple people accidentally starting on the same things? Not sure how things like that have been done previously.

I probably shouldn't generate a big list ahead of time because it might get auto-assigned to a random participant in the meantime.

So, any comments, questions, ideas?

petrw1 2015-07-22 21:50

How you handled the several hundred triple checks last time seemed to work....

Unless GPU72 can create a new "reserved" work-type for this. :)

Madpoo 2015-07-22 21:52

[QUOTE=petrw1;406321]How you handled the several hundred triple checks last time seemed to work....

Unless GPU72 can create a new "reserved" work-type for this. :)[/QUOTE]

That's kind of my thought too. Not make it too complicated, and trust people to lookup the exponent first and see if it's assigned to someone before they start any work on it.

If two people did end up working on it, that may just be okay because I predict a good amount of these will need 2 more checks anyway, considering I'm guessing the first one is bad... so it wouldn't be the end of the world, but then again some of these will be right.

As the saying goes, even a broken clock is right twice a day, and sometimes these computers manage to get it right.

chalsall 2015-07-22 22:23

[QUOTE=petrw1;406321]How you handled the several hundred triple checks last time seemed to work....

Unless GPU72 can create a new "reserved" work-type for this. :)[/QUOTE]

I would rather not get involved with the DC'ing assignments for two reasons. First, work that doesn't have long-term value. And second, I don't want this effort to be constrained to only GPU72 workers.

On the other hand, it would be trivial for me to ensure the candidates are appropriately TF'ed. The code already exists to prioritize certain candidates.

Aaron, if you PM me the list (or post it here) I'll put them in the system for workers to process.

Madpoo 2015-07-22 22:54

1 Attachment(s)
[QUOTE=chalsall;406323]I would rather not get involved with the DC'ing assignments for two reasons. First, work that doesn't have long-term value. And second, I don't want this effort to be constrained to only GPU72 workers.

On the other hand, it would be trivial for me to ensure the candidates are appropriately TF'ed. The code already exists to prioritize certain candidates.

Aaron, if you PM me the list (or post it here) I'll put them in the system for workers to process.[/QUOTE]

I'm attaching a rough list here. These are all of the currently unassigned exponents, and I used a factor of "bad > good * 1.5" which is pretty loose. But I wanted to get as many in here *solely* for purposes of making sure they're all TF'd to a high enough level.

This also only includes exponents below 58M since I started this out as a way to fill in some vital gaps between where the DC'ing is, and M(48).

I wouldn't start on doing any DC work yet... wait on the TF'ing and then we'll have new lists of the DC eligible stuff.

For those lists, I'll include the # bad, # good, # unknown, # suspect that the same machine has done which people can use as a guide to picking the exponents they're interested in checking.

Mark Rose 2015-07-22 23:56

[QUOTE=chalsall;406323]On the other hand, it would be trivial for me to ensure the candidates are appropriately TF'ed. The code already exists to prioritize certain candidates.[/QUOTE]

What "Option" is needed to get these prioritized assignments? It would be nice if there were a new one that wouldn't return any assignments if the whole list is assigned, so I can prioritize the work I do.

chalsall 2015-07-23 01:17

[QUOTE=Madpoo;406328]I wouldn't start on doing any DC work yet... wait on the TF'ing and then we'll have new lists of the DC eligible stuff.[/QUOTE]

OK. This is now underway. Should take about a week or so.

Note that the first 356 in your list (to 41228987 inclusive) are already appropriately TF'ed, so DC'ing could start on these immediately.

chalsall 2015-07-23 01:26

[QUOTE=Mark Rose;406331]What "Option" is needed to get these prioritized assignments? It would be nice if there were a new one that wouldn't return any assignments if the whole list is assigned, so I can prioritize the work I do.[/QUOTE]

No special option needed; these are simply handed out first to those who choose "WMS" or "LG72D". You, in fact, are already being assigned some of these to take from 70 to 71.

Prime95 2015-07-23 02:29

I've taken these:


48435221,72
48465463,72
48587491,72
48659089,72
49229179,72
49270057,72
49380893,72
49450519,72
49451071,72

All reserved by prime95 after prefixing with "Test=" and adding a suffix of ",1"

Prime95 2015-07-23 02:43

I took these:

Test=37110529,71,1
Test=37126321,71,1
Test=37127599,71,1
Test=37133377,71,1
Test=37139287,71,1
Test=37170901,71,1
Test=37188343,71,1
Test=37188829,71,1
Test=37191487,71,1
Test=37207999,71,1
Test=37227653,71,1
Test=37240271,71,1

Mark Rose 2015-07-23 03:25

[QUOTE=chalsall;406336]No special option needed; these are simply handed out first to those who choose "WMS" or "LG72D". You, in fact, are already being assigned some of these to take from 70 to 71.[/QUOTE]

Awesome. I've now switched 2/3 of my current power to take things to 72 for this project.


All times are UTC. The time now is 14:53.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.