Ideas for a novice GIMPS client

[Prime95]

Originally Posted by nomadicus

This is not it per se, but perhaps revisiting the discussion on schedulers, games, and novice users is worth while.

Maybe it was in one of the old TPR forum threads.

[cheesehead]

Originally Posted by E_tron

It would be hard for a client to figure out memory usage of a system, so the client would need some fail safe settings.

Basically, the client already has that.

Prime95 already allocates only the minimum amount needed when it's doing trial factoring, P-1 stage 1, or L-L testing.

The only functions for which it uses the memory settings specified by the user are P-1 stage 2, ECM stage 2, and maybe (I don't recall for sure) ECM stage 1, none of which are required in order for any user to satisfy a PrimeNet assignment.

I was thinking that Prime95 could already have the memory settings in it for specific system memory arrays. Like a 128mb system could give 24 to P95. A 256mb system could be 48. A 512mb system could be 96 and so on.

But how could Prime95 determine that those amounts were okay to allocate, without asking the user? Maybe the user is running a system so fully loaded that it can spare only 8MB to Prime 95 without unacceptable levels of swapping, even though it has 256MB RAM total. Prime95 can't determine that by itself.

When a system cannot give enough memory for First LL Tests it can revert to Double-Checking, then factoring, then decommissioned from PrimeNet.

This really isn't a big enough problem to justify making the software more complicated. If a system can't handle what it's assigned, either it releases the assignment or the assignment times out and PrimeNet reassigns it to some other system.

Furthermore, Prime95 and PrimeNet _already have provisions_ to restrict First-time and DC assignments to systems most likely to be able to handle them. The default mode is to automatically request most-suitable assignments. Unless the user _specifically_ requests a first time assignment, first time L-L are assigned only to systems above a certain speed rating, which George raises once in a while. Similarly, there is a lower threshold below which only trial factoring is assigned unless the user _specifically_ requests DC or first-time.

In my mind that would be almost unnoticeable memory settings and we won’t have to worry about 386s getting First time LL tests.

As noted above, there is already a threshold below which PrimeNet will assign only trial factoring unless the user request otherwise. 386s are well below that threshhold.

386s getting LL assignments simply is not a significant problem. It just doesn't happen often enough to make any noticeable difference. When it does happen (which can only be because the 386 user specifically overrode the default setting and requested L-L, or because the 386 user claimed his CPU was a faster model), the PrimeNet assignment system already takes care of it with its timeout procedure.

If the memory settings don’t allow P-1 stage could the stage be skipped all together?

I presume you're referring to P-1 stage 2.

Yes, but this feature (skipping stage 2 if not enough memory is available) is already in Prime95 and does not have to be added.

[Reboot It]

Originally Posted by Prime95

Many times I've thought about eliminating the one-hour test as it seems to be useless.

Unless of course you allow it to write a flag of some kind (based on cumulative time run in self-test and the number of failures) that would prevent Prime95 doing any live work after that point.

[Mr. P-1]

Originally Posted by Reboot It

Unless of course you allow it to write a flag of some kind (based on cumulative time run in self-test and the number of failures) that would prevent Prime95 doing any live work after that point.

Why not just switch to factoring?

[alienz]

what/ where is the torture test?

[patrik]

It is under Options->Torture Test. You would have to stop the exponent you are working on first, if you want to start it manually. If you start it manually it goes on indefinitely until you stop it or an error is found.

The torture test runs a number of iterations of predefined exponents and then compares the result with the known result. If the result doesn't match the known one, there must be some problems with the computer. See the files readme.txt and stress.txt. (stress.txt deals only with these matters.)

[trif]

Originally Posted by Mr. P-1

Why not just switch to factoring?

If it fails the one hour torture test, it would be better off on LL testing than factoring. LL is double checked, factoring isn't. I'd torture test a box intended for factoring much more strictly than a box intended for LL testing.

Who knows, it might just return a good result every now and then, and it will be keeping the user's CPU nice and warm.

[sdbardwick]

Originally Posted by trif

[ LL is double checked, factoring isn't.

IIRC, factors are checked by the server when submitted. Checking a single factor is a low-cost check in terms of CPU time.

[trif]

Originally Posted by sdbardwick

IIRC, factors are checked by the server when submitted. Checking a single factor is a low-cost check in terms of CPU time.

Yes, but finding no factor is not double checked, which means that such a computer can miss an awful lot of factors and cause GIMPS a lot of unneeded work.

[Mr. P-1]

Originally Posted by trif

Yes, but finding no factor is not double checked, which means that such a computer can miss an awful lot of factors and cause GIMPS a lot of unneeded work.

AIUI errors are most likely to affect FP calculations. Also an undetected error rete of 1 in a billion cycles would destroy an LL, but the risk of missing a factor would be negligable.

But if errors ore really such a concern then the solution is not to report negative results, only factors found. Alternatively the program could factor past the current maxumum, thereby searching for factors which would not otherwise be picked up.

Regards

Daran