New milestone

[Prime95]

Originally Posted by henryzz

If you change the threshold then please make sure that Q6600s fit first-time tests.

Yes, quads will still get first-time tests if the threshold is changed. Under some of the other suggestions posted here a quad would get double-checks.

Your request, though, falls on deaf ears as you can always set your work preference to get exactly the work type you want.

[Uncwilly]

Originally Posted by Prime95

Under some of the other suggestions posted here a quad would get double-checks.

I had suggested earlier (somewhere) that, before actually being assigned a 100M digit L-L, that those requesting such, be assigned a minimum number of DC's (unless they have a history on that CPUID that proves them stable). I was thinking a minimum of 2 per core that match (in a row). If you have a quad core turn in 8 clean and matching DC's, that should prove that it is viable to do a big LL. You may want to tweak the back-end (to account for the first timers being bad or wanting at least a 95% OK rate, which would mean any non-matches would require >10 clean matches.)

This will help the DC problem and prevent wasted cycles on bad 100M digit numbers.

[Prime95]

Originally Posted by Uncwilly

I had suggested earlier (somewhere) that, before actually being assigned a 100M digit L-L, that those requesting such, be assigned a minimum number of DC's.

An overriding rule of GIMPS has always been to let the end user select the work they want to do.

For example, I had one user that insisted on running P-1 with B1=120 and B2=large. Several times I tried to convince him of the foolishness of this, but he persisted. Yet there is still no code in prime95 that prevents the user from doing this.

[NBtarheel_33]

So it seems to me that anyone who happens to just download and install Prime95, and blindly run through the initial dialog boxes without selecting a type of work or anything, should be assigned DCs by default. In other words, make DC the default worktype unless someone knows enough (and takes the action) to select something different.

"Whatever makes sense" should also mean doublechecks. I am assuming that anyone selecting "whatever makes sense" either (1) doesn't care what type of work they contribute, (2) doesn't understand the differences between worktypes, or (3) really wants to do the work that is most needed by GIMPS. Yes, (1) and (2) are taking advantage of apathy and ignorance, respectively, but there is nothing that says that down the line, the user can learn more about or become more interested in GIMPS, and then try out different worktypes. Moreover, I'm betting that many (1) and (2) cases might include the folks who start an LL, only to become frustrated with the length of time it will take, and abandon the assignment. With a doublecheck taking a couple of weeks, we might be lucky enough to get at least one result from the early attriters (is that even a word? LOL).

With regard to silly stuff like the guy wanting to run P-1 with B1 = 120, I think that we really need to work on forming a Web page explaining how the different settings work for the different worktypes. I'm wondering how many users are blindly letting Prime95 run P-1 with 8 MB of RAM allocated, not because they can't afford higher memory, but because they simply don't understand why/how P-1 does better with more memory. It is not really noted anywhere on the GIMPS site what the recommended amounts of RAM are for P-1 (for instance, while 8 MB is really too little, something like 3GB for a single exponent is overkill). Trial factoring is another issue - I know it's in the source code, but it's not really out in the open what bit depth should be used for a given exponent range. Cheesehead wrote a good article on why trial factoring up to sqrt(Mp) is physically impossible - perhaps this could be included to help people understand why you can't just TF an exponent rather than running an LL.

I think that there are many users out there who are interested in the mathematics and the goals of GIMPS, but that are unaware of the finer points of how things work within the project. Along with this fact we must also consider that most users adopt the common paradigm of installing software and leaving the default options in place. I must admit that while I joined GIMPS in 2002, and began LL testing at that point, it was only with the introduction of v5 in 2008 that I began exploring other worktypes and learning more about how the different assignments and settings work. Indeed, I was one of the many who just chose the 8 MB option for P-1 and as a result I handed in very weakly P-1'ed exponents (and probably ran many unnecessary LL tests) for my first five or so years in GIMPS.

[Prime95]

Originally Posted by NBtarheel_33

So it seems to me that anyone who happens to just download and install Prime95, and blindly run through the initial dialog boxes without selecting a type of work or anything, should be assigned DCs by default.

Replace "should" with "could". There is no consensus as to how important DC is vs. first-time LL.

With regard to silly stuff like the guy wanting to run P-1 with B1 = 120, I think that we really need to work on forming a Web page explaining ... It is not really noted anywhere on the GIMPS site what the recommended amounts of RAM are for P-1 (for instance, while 8 MB is really too little

No amount of documentation would have helped Mr. B1=120.

8MB is sufficient for P-1 - just not optimal.

[Primeinator]

Originally Posted by Uncwilly

I had suggested earlier (somewhere) that, before actually being assigned a 100M digit L-L, that those requesting such, be assigned a minimum number of DC's (unless they have a history on that CPUID that proves them stable).
This will help the DC problem and prevent wasted cycles on bad 100M digit numbers.

I like this idea; though would not most of these new computers be stable unless they are being overclocked? Pardon my ignorance on this subject!

Originally Posted by Prime95

An overriding rule of GIMPS has always been to let the end user select the work they want to do.

I think that this should be kept in place, but....

Originally Posted by NBtarheel_33

So it seems to me that anyone who happens to just download and install Prime95, and blindly run through the initial dialog boxes without selecting a type of work or anything, should be assigned DCs by default. In other words, make DC the default worktype unless someone knows enough (and takes the action) to select something different.

"Whatever makes sense" should also mean doublechecks. I am assuming that anyone selecting "whatever makes sense" either (1) doesn't care what type of work they contribute, (2) doesn't understand the differences between worktypes, or (3) really wants to do the work that is most needed by GIMPS. Yes, (1) and (2) are taking advantage of apathy and ignorance, respectively, but there is nothing that says that down the line, the user can learn more about or become more interested in GIMPS, and then try out different worktypes. Moreover, I'm betting that many (1) and (2) cases might include the folks who start an LL, only to become frustrated with the length of time it will take, and abandon the assignment. With a doublecheck taking a couple of weeks, we might be lucky enough to get at least one result from the early attriters (is that even a word? LOL).

This is also a good idea.

Originally Posted by Prime95

There is no consensus as to how important DC is vs. first-time LL.

Perhaps we should take a poll then among active participants? It seems like at least a few on this thread support more attention being put on DCs (with caveats, of course). I think the reason for this was well-stated in the fact that it is going to take an increasingly large amount of time to confirm the positions of M41 to M47. In a practical sense, this may not matter as the work will get done 'eventually' as total throughput between DC and first-time ect is remaining relatively the same. However, some may feel that these gaps need to be completed in a more timely manner to effectively "settle" the matter. The end result will be the same regardless of the path taken; just the time in which they are realized will be different (potentially greatly different).

Kyle

[Prime95]

Originally Posted by Primeinator

Perhaps we should take a poll then among active participants? It seems like at least a few on this thread support more attention being put on DCs

No need for a poll. The important point is I tend to agree with you that DCs should be given somewhat more priority. I'll implement something and we'll see how it goes.

The end result will be the same regardless of the path taken; just the time in which they are realized will be different

Quite true.

[Primeinator]

Okay. I now have two computers working for GIMPS- a dual core and a quad. When my dual core finishes it's current assignments, I will put it onto doing DC work (a drop in the bucket) but my quad core is 'relatively fast' and I will keep it doing first-time LLs at least for the present time. I have another PC 'sort of' working that I will also put onto DC work though it may be months before I resume access to that computer.

Is there anyway to send a message to people currently running the client that would give an explanatory message regarding the DC situation and asking if they would like to temporarily switch over their work type to aid the project? Of course, this information would have to include the prize money, etc. I do not have a good idea about the technical limitations of the new server and this idea may be impractical for other reasons. Again, please pardon my ignorance!

[petrw1]

The same way that you asked for P-1 help last year (and at least got "some" attention) could the same idea work for DC and catch the attention of some of the the big "farms". Someone like me with about a dozen active PC's of various sizes could allocate the entire team to DC and still take 20 years to complete one million range; DC is about 25 ranges behind LL.

I know there are several big farmers doing LOTS of factoring. Maybe(?) in the next few weeks when all the exponents are factored to 64 bits some of these might consider helping out with DC. This way we don't slow down LL in the meantime.

[chalsall]

Originally Posted by petrw1

I know there are several big farmers doing LOTS of factoring. Maybe(?) in the next few weeks when all the exponents are factored to 64 bits some of these might consider helping out with DC. This way we don't slow down LL in the meantime.

OK -- as one of those doing lots of LMHing (currently 110,000 a day)...

I do this partially because I want to have a regular and predictable amount of internet traffic from each machine (75 dual core at at least 2GHz) for monitoring purposes. Thus I will want to continue doing low-level TFing.

However, what I can and will do is allocate one core on each machine to DCing, leaving the other core to TFing / traffic generation.

And, George, I agree with the others above -- don't mess with the stats giving DC work a GHz Days "bonus". Not appropriate; and I don't think needed.

[cheesehead]

Originally Posted by NBtarheel_33

So it seems to me that anyone who happens to just download and install Prime95, and blindly run through the initial dialog boxes without selecting a type of work or anything, should be assigned DCs by default. In other words, make DC the default worktype unless someone knows enough (and takes the action) to select something different.

But it makes much more sense for the default worktype to be "whatever makes sense" than for it to be a specific worktype. That way, the admins can adjust the assignments for "whatever makes sense" to be whatever seems to be in GIMPS's best interest at a particular time. DCs might not always be what is in GIMPS's best interest.

Your proposal would leave the admins less flexibility -- why tie their hands?

"Whatever makes sense" should also mean doublechecks.

... and it can, whenever the admins make it so.

I am assuming that anyone selecting "whatever makes sense" either (1) doesn't care what type of work they contribute, (2) doesn't understand the differences between worktypes, or (3) really wants to do the work that is most needed by GIMPS. Yes, (1) and (2) are taking advantage of apathy and ignorance, respectively,

Not really "taking advantage" in any exploitative sense, I think.

If someone doesn't care -- they don't care, and will be happy for the GIMPS server to make the choice. Isn't that practically synonymous with reason (3)? After all, those people are joining GIMPS because they want to help the project! Allowing the server to decide the worktype is perfectly compatible with that.

I think reason (3) may be much the most common reason when you take that into account.

but there is nothing that says that down the line, the user can learn more about or become more interested in GIMPS, and then try out different worktypes.

... so we don't need to change the assignment system's default, do we?

Moreover, I'm betting that many (1) and (2) cases might include the folks who start an LL, only to become frustrated with the length of time it will take, and abandon the assignment. With a doublecheck taking a couple of weeks, we might be lucky enough to get at least one result from the early attriters (is that even a word? LOL).

That's a good reason to assign DCs to new joiners ... but it can just as well be implemented with a default of "whatever makes sense" -- the algorithm can simply take the user's newness into account, and assign DCs to newbies.

OTOH, reason (2) users might be dismayed to be assigned DCs -- they thought they would be getting a chance to win the EFF prize. However,

(A) they will get that chance later if they complete the initial DCs, and

(B) that's what they get for ignorance ... but it's easily remedied! Once they've learned the difference between a first-time LL and DC, they are completely free to specify first-runs only.

Trial factoring is another issue - I know it's in the source code, but it's not really out in the open what bit depth should be used for a given exponent range.

But TF assignments don't, by default, offer the user the chance to specify bit depth, so what's the problem? A beginning TF worker doesn't need to know the default bit depths -- they're automatically set in the worktodo assignment lines.

Once the user learns more about TF, s/he can learn how to specify bit depths on the TF worktodo lines if s/he desires to override the default bit levels.

Cheesehead wrote a good article on why trial factoring up to sqrt(Mp) is physically impossible - perhaps this could be included

Included in readme.txt, you mean?

to help people understand why you can't just TF an exponent rather than running an LL.

Someone who actually tries TFing to the square root will be effectively educated by the wall clock.

I think that there are many users out there who are interested in the mathematics and the goals of GIMPS, but that are unaware of the finer points of how things work within the project.

We do have this forum ...

Indeed, I was one of the many who just chose the 8 MB option for P-1

Let's not forget that one of the other basic principles of this project, in addition to the one George just mentioned, is to have as small a "footprint" on the user system as possible. That's why we have the programs run at lowest priority ... and why by default we have the user run only stage 1 of P-1 (that's the effect of the 8 MB default).

BTW, there _is_ an explanation of the P-1 memory issues in readme.txt -- it's under the heading "SETTING AVAILABLE MEMORY". Did you read that before, or while, you ran your first assignments? If not, why not?

Not much can be done^* about users who won't bother reading the readme.txt -- adding more explanations to it won't persuade new users who think it's already too long to bother with.

and as a result I handed in very weakly P-1'ed exponents (and probably ran many unnecessary LL tests) for my first five or so years in GIMPS.

... but you also had minimal interference with your system's performance because the software used only as much memory as it had to.

If GIMPS software bogged down systems by allocating gobs of memory once in a while for P-1 stage 2, without an understanding and informed consent (by deliberate memory setting) by the user, our project's software might be kicked out of many places where it now runs quietly with minimal impact.

Getting no LL contribution at all is worse for this project than getting a small percentage of LLs run where a large-memory P-1 stage 2 might have found a factor.

- - - - -

^* We might put some highly-effective introduction at the beginning that will grab new users' interest enough so that they'll want to devour every word that follows -- care to write one, anybody? Shall we make an offer to some best-selling author?? Anyone know a best-selling author who likes to crunch primes in the background as s/he writes?