mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Data

Reply
 
Thread Tools
Old 2014-01-23, 06:55   #45
Batalov
 
Batalov's Avatar
 
"Serge"
Mar 2008
Phi(3,3^1118781+1)/3

906210 Posts
Default

Quote:
Originally Posted by Prime95 View Post
Frankly, I'd prefer that the person who finds the next Mersenne prime knows what a Mersenne prime is.
...or in a parallel universe, "the person who finds the next 321 prime knows what 321 (esp. Proth) prime's properties are".

There's regular talk of "poachers" at the PGrid's forum. Not to say that there's an occasional voice of reason there as well. (Unlike mersenneathome's forum that was just as if it was taken from a "1984" provincial theatrical stage play. There are some pleasant differences.) Anyway. There's always a pliable (and silent) majority, and OTOH there are some activists who think that they have the right to speak for the project - even if they do not. It is a bit amusing to watch.

Quote:
All people who know where the beef is go into politics, the rest get off by trolling on forums.
Batalov is offline   Reply With Quote
Old 2014-01-23, 07:28   #46
TheMawn
 
TheMawn's Avatar
 
May 2013
East. Always East.

11·157 Posts
Default

Assume 1000/6000/12000 based on George's last proposal. There are four "blocks" of assignments:
  1. Top 1000, recycled aggressively, assigned to fast, consenting users.
  2. 1000-6000 recycled passively, assigned to fast, consenting users.
  3. 6000-12000 recycled passively, assigned to fast users
  4. 12000+ recycled passively, assigned to users not qualifying for 1-3 but still meeting today's DC requirements.

The transition from one block to another may present issues. Let us see:

Going from 4 to 3 presents no issue for you. You were deemed a "slow" or "not trusted yet" user but recycling rules did not change. However, you are now taking up one slot from block 3. If block 3 is too small or emptied too fast, "slow" users may occupy too much of block 3, forcing users qualifying for 3 to be put into 4. Overall, this is not a big deal, but isn't ideal either.

Going from 3 to 2 presents no issue for you but the same issue as above occurs for other users. If block 2 is emptied too fast, users from 3 will occupy too many block 2 slots forcing users wishing for block 2 to be sent to block 3 (or even 4 if the system is really out of whack)

Going from 2 to 1 presents a small issue for you because your recycling rules are suddenly much stricter, potentially causing the immediate cancellation of your assignment. This issue is alleviated by requiring you to consent to get assignments in block 2. However, if you were initially placed in block 3 and block 2 is too small (or block 1 is too large) you may end up inside block 1 before you finish, causing the likely immediate cancellation of your assignment.


Conclusions:
  • Block 3 must be sufficiently large to prevent it from becoming too full of block 4 users. The system in imbalanced if a block 3 qualifier is forced to take a block 4 assignment.
  • Block 3 must not be so large as to hold up too many preferred assignments. The system is imbalanced if block 3 has too many vacancies.
  • Block 2 must be sufficiently large to prevent it from becoming too full of block 3 or block 4 users. The system is imbalanced if a block 2 qualifier is forced to take a block 3 assignment, and heavily imbalanced if they must take a block 4 assignment.
  • Block 2 must not be so large as to hold up too many preferred assignments. The system is imbalanced if block 2 has too many vacancies.
  • Block 2 must also be sufficiently large to act as a buffer between 3 and 1. Otherwise, recycling rules may change abruptly. The system is imbalanced if a block 3 user enters block 1.
TheMawn is offline   Reply With Quote
Old 2014-01-23, 07:40   #47
TheMawn
 
TheMawn's Avatar
 
May 2013
East. Always East.

11·157 Posts
Default

Recommendations:

Evaluate the 1000/6000/12000 on a regular basis. Daily would be great. Weekly might be more practical.

Always provide the lowest qualifying exponent.
  • If a B3 user is given a B4 assignment, increase 12000 by 0.5.
  • If B3 has more than 500 vacancies, decrease 12000 by 1.0 per vacancy above 500.
  • If a B2 user is given a B3 assignment, increase 6000 and 12000 by 0.5.
  • If a B2 user is given a B4 assignment, increase 6000 by 1 and 12000 by 2.
  • If a B3 assignment is dropped to B1, ... I have no idea ...
  • If B2 has more than 500 vacancies, decrease 6000 by 1.0 per vacancy above 500.

Upon evaluation, set 1000,6000,12000 to nearest 10.
TheMawn is offline   Reply With Quote
Old 2014-01-23, 07:50   #48
TheMawn
 
TheMawn's Avatar
 
May 2013
East. Always East.

11×157 Posts
Default

Addendum:

I've noticed that the 1000 actually remains fixed by all these rules. I never gave any consideration to how its size affects progress. If it is too big, too many assignments get recycled before they maybe should (because B3 dropped to B1). If it is too small, too many assignments are allowed to carry on for one to two years.

Unfortunately, I can't really think of a well-defined event which clearly shows 1000 being too small. If B3 falls into B1, it's a sign that B2 is too small or that B1 is too large.

However, we don't really define the amount of time before a DC is declared to be taking too damn long.


Also, I think it's possible for an assignment to fall through from B3 to B1 even if there are lots of vacancies in B2, as long as the exponents are done fast. Increasing the size of B2 doesn't help because the vacancy rule will kick in.

EDIT: I'm going to sleep on this. I'll also give you all some time to read this stuff and maybe you'll think of something.

Last fiddled with by TheMawn on 2014-01-23 at 07:54
TheMawn is offline   Reply With Quote
Old 2014-01-24, 07:01   #49
philmoore
 
philmoore's Avatar
 
"Phil"
Sep 2002
Tracktown, U.S.A.

100010111012 Posts
Default

Everything said on this thread should also apply to the first-time LL rules, only the specific limits on the categories may change. That said, Uncwilly has observed that we are currently DC-ing about 39000 exponents per year. Therefore one may conclude that any exponent in the bottom (not "top") 20000 available exponents could become reasonably close to the lower edge in six months. We don't want to change rules on someone just because their exponent has now become close to the lower edge. To me, this is an argument for extending the range where we impose restrictions on completion time, and pushing the relatively more lenient deadlines to a higher interval above the lower edge.

I think this is a great time to make these changes. I don't care how long it takes to do the last 15 double-checks to verify M43, and in lieu of that milestone I would prefer to see progress made on the next milestone towards M44, but I do think that the completion of milestones adds to the community's general sense of achievement, and I applaud changes to make this happen in a more timely and orderly manner.
philmoore is offline   Reply With Quote
Old 2014-01-24, 16:33   #50
TheMawn
 
TheMawn's Avatar
 
May 2013
East. Always East.

11·157 Posts
Default

None of these rules are actually in place. We're in the process of refining them now, and the point you've brought up is a very good one. 12,000 might be a bit low, if our throughput is actually 40,000 per year. The self-adjusting block sizes I proposed would have taken care of that eventually but we should probably start higher than lower.

We're better giving someone assignment #50,000 when #30,000 was appropriate than vice versa.
TheMawn is offline   Reply With Quote
Old 2014-01-24, 16:46   #51
chalsall
If I May
 
chalsall's Avatar
 
"Chris Halsall"
Sep 2002
Barbados

7×1,291 Posts
Default

Quote:
Originally Posted by TheMawn View Post
We're in the process of refining them now, and the point you've brought up is a very good one. 12,000 might be a bit low, if our throughput is actually 40,000 per year. The self-adjusting block sizes I proposed would have taken care of that eventually but we should probably start higher than lower.
For context, this report covering our DC, LL and Factored productivity (over the last six months) might be helpful in this optimization analysis.
chalsall is offline   Reply With Quote
Old 2014-01-24, 17:19   #52
chalsall
If I May
 
chalsall's Avatar
 
"Chris Halsall"
Sep 2002
Barbados

7×1,291 Posts
Default

Quote:
Originally Posted by TheMawn View Post
We're better giving someone assignment #50,000 when #30,000 was appropriate than vice versa.
Hmmmm... Thinking about this a bit more...

Whatever the offset value is communally decided on for the slowest machines, we'll need a little a bit of time to do the appropriate TFing. Very few candidates are already currently appropriately TF above the current DCing wave-front.

Once the decision is made I can bring the appropriate range (high 34M or low 35M, for example) into GPU72, and hopefully some of our GPU TFing workers will step-up and build up an appropriate lead.
chalsall is offline   Reply With Quote
Old 2014-01-24, 17:36   #53
TheMawn
 
TheMawn's Avatar
 
May 2013
East. Always East.

11×157 Posts
Default

The TF'ing is something I hadn't thought about. I like the initiative of trading CPU for GPU.
TheMawn is offline   Reply With Quote
Old 2014-01-25, 00:07   #54
Aramis Wyler
 
Aramis Wyler's Avatar
 
"Bill Staffen"
Jan 2013
Pittsburgh, PA, USA

22×89 Posts
Default

Quote:
Originally Posted by Prime95 View Post
This is IMO the major flaw with my proposal.

...

I'm tempted to say preferred DCs and LLs should be a separate work type. By selecting that work type, the user is volunteering to work on exponents that will be expired more quickly than standard exponents.

The downside is I think this would require a client change.
I'm not arguing for or against anything here, except the client change. You could set them up as different work types by server, rather than per client by setting up a 'preferred' proxy. Only people with trusted computer could add the proxy changes that would put them on the preferred primenet proxy (PPP!), and the regular primenet would not give out preferred assignments - only the PPP would. If someone w/o a trusted computer tried to get assignments from the PPP they'd just get a message to use the regular primenet server till they were trusted.

How would they know when they were trusted you ask? Pfft. Anyone that cares enough about the end goal to add proxy settings in is aiming to be trusted. You don't have to tell them when they're trusted because they're aggressive enough to find out on their own.

Last fiddled with by Aramis Wyler on 2014-01-25 at 00:07
Aramis Wyler is offline   Reply With Quote
Old 2014-01-25, 18:25   #55
chalsall
If I May
 
chalsall's Avatar
 
"Chris Halsall"
Sep 2002
Barbados

7·1,291 Posts
Default

Quote:
Originally Posted by chalsall View Post
Once the decision is made I can bring the appropriate range (high 34M or low 35M, for example) into GPU72, and hopefully some of our GPU TFing workers will step-up and build up an appropriate lead.
So everyone knows, I've tasked "Spidy" with bringing into GPU72 everything in 34M and 35M for TFing.

Given our approximate DC completion rate of 41,000 candidates a year, if the communally agreed "not preferred" assignments given to slower machines is 12 months out, then as of today the first assigned candidate would be 34,290,001. If 18 months out, it would be 35,316,613.

I've got my machine, and will task a few MISFIT machines set to "Let GPU72 decide" to occasionally work these ranges to build up a buffer. I want to make sure the TF'ing is not a block in the critical path once a decision is made.

Anyone who is currently using MISFIT but explicitly only wants to do LL-TF'ing, please either PM me, or just change your setting from "Let GPU72 decide" to "What Makes Sense". Only the "Let GPU72 decide" setting will ever be given this DC-TF work, and not much of it, and only when needed. We definitely don't want to take too much firepower away from the LL-TF'ing work.

Conversely, anyone who wants to help are more than welcome to "throw a few on the barbie..." -- it would be appreciated. I'd like to build up a buffer of a couple of thousand candidates appropriately TFed over the next couple of weeks to ensure we're ready.

Last fiddled with by chalsall on 2014-01-25 at 18:31 Reason: Smelling mistake.
chalsall is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
PrimeNet Assignment Rules S485122 PrimeNet 7 2018-06-08 14:49
Modifications to DC assignment rules Prime95 PrimeNet 74 2017-01-18 18:36
Understanding assignment rules Fred PrimeNet 3 2016-05-19 13:40
Proposed LL assignment and recycle rules Prime95 Data 156 2015-09-19 12:39
Proposed TF, P-1, ECM assignment and recycle rules Prime95 Data 9 2014-02-27 23:52

All times are UTC. The time now is 05:50.

Wed Jul 15 05:50:15 UTC 2020 up 112 days, 3:23, 0 users, load averages: 1.28, 1.29, 1.32

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

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.