test of the software

[woody]

What testing was done on the Primenet software to be sure it works for large exponents? i raise this question because I looked at the LL results
for exponents from 59,000,000 to 110,000,000
and for those few exponents that were tried twice the residues
never agree and an error was reported in some of the cases.

[Christenson]

How many few? And by what kind of machines? Prime95 (as opposed to Primenet, which is the server and the internet communications) is well beyond the point where actual machine errors can be expected, especially in overclocked machines.

If you want, we can pick an exponent, maybe one of those with multiple LL residues, and make sure it is TF'ed to 72 or 73 bits, and then both run an LL test on it. But it's a project, and it will take a month or two.

[woody]

I count 7 unverified pairs that have unmatched residues.
the display does not reveal the type of machine that was used.
1 pair shows an error.

I should have used the term Prime95 instead of Primenet.
I still would like to know how the software was tested for large exponents.

[Christenson]

And how many verified? Are you interested in attempting LL-D on any of them?

[Uncwilly]

The software works great. Many of those above the current range and below 79,300,000 have been done a while back. The machines are the issue. As a test takes longer (the bigger the number the longer the test), there is a greater chance of an error occurring in the machine (hardware related issues and cosmic rays, etc.) This has been seen before when there was a push by some to try to find the first prime with 10 million digits. Many of those early tests suffered hardware issues.

The basic algorithm to run the LL test is the same. The software is run on smaller numbers, but using the larger FFT sizes to verify that they work. The FFT's are also used in other tests.

The program is ok, it is the actual machines. There are other threads around that talk about this.

[cheesehead]

Originally Posted by woody

What testing was done on the

prime95

software to be sure it works for large exponents? i raise this question because I looked at the LL results for exponents from 59,000,000 to 110,000,000 and for those few exponents that were tried twice the residues never agree and an error was reported in some of the cases.

woody,

Before you use that report again, please carefully examine all the check-box options. When the "Exclude verified results" box is checked (by default), the report excludes all verified results.

When you _do_ specify the verified results (by unchecking the "Exclude verified results" box) and exclude the unverified results (by checking the "Exclude unverified results" box), you get something like:

Verified test results

Exponent,User name,Computer name,Residue

67108763,Stephan Grupp,P4,53023CAC72917105

67108763,Brian J. Beesley,Canopus,53023CAC72917105

67239937,Alexandr Fomin,FAA,F2DC9FBF28E585EB

67239937,Luigi Morelli,Manual testing,F2DC9FBF28E585EB

68808029,wavelet3000,Manual testing,A9878EE6E59BA060

68808029,msft,,A9878EE6E59BA060

69100037,Team_Bundu,bundu5,2C97F61B5F6869BA

69100037,msft,Manual testing,2C97F61B5F6869BA

70022021,Norman Carlson,NormNew,8FC5E5F795149FF8

70022021,msft,Manual testing,8FC5E5F795149FF8

79299821,Ars Technica Team Prime Rib,DSheets_02,C1D3BB65E8D3ED07 

79299821,arnaud,SPR18,C1D3BB65E8D3ED07

100000007,William Christian,StarQwest1,F9042256B193FAA0 

100000007,TeamComputerraRU,Yxine,F9042256B193FAA0

As for the exponents that really do have mismatches:

Have you considered that

1) the prime95 software works in partnership with hardware,

2) that hardware sometimes, in the midst of quadrillions of instructions, has a bit go wrong because of a slight defect or a cosmic ray hit,

and

3) the longer the LL test (i.e., the higher the exponent), the greater the chance that a hardware error will occur sometime during it?

[woody]

Perhaps it would be useful to report for the unverified results
the software version used and the machine type.
If cosmic rays are the explanation, then it would be interesting
to see a plot of exponent range vs number of mismatched exponents in
that range.

[cheesehead]

Originally Posted by woody

Perhaps it would be useful to report for the unverified results
the software version used and the machine type.
If cosmic rays are the explanation, then it would be interesting
to see a plot of exponent range vs number of mismatched exponents in
that range.

This sort of thing is more often discussed in the Data subforum (mersenneforum.org > Great Internet Mersenne Prime Search > Data). Look through some of the threads there.

[patrik]

Originally Posted by woody

[---]
If cosmic rays are the explanation, then it would be interesting
to see a plot of exponent range vs number of mismatched exponents in
that range.

Error rate plot

[S34960zz]

See also:
http://www.mersenneforum.org/showthread.php?t=15542

[woody]

Originally Posted by patrik

Error rate plot

Thanks for posting this plot. Not sure i understand it though.
I am seeing decline error rates generally as a function of exponent size.
If cosmic rays are cited as reason for errors, then with larger
exponents i expect longer running times, so more chance
for the rays to flip a bit and cause an error.
But I don't see that happening.
Could there be another explanation?
I think what is really needed is a graph of error rate by machine
type as a first approximation.