The Next Big Development for GIMPS

[kriesel]

Something about the combination of PrimeNet and prime5 V30.2 seems to be ignoring "Days of work to queue up" in the case where existing work in progress does not match the worker window setting. I've seen this on multiple systems immediately after updating from V30.1b1 to V30.2b1. On one, an existing PRP DC had 19 days to run yet and days of work to queue up was set to 1, but it obtained a new world record PRP assignment immediately upon launch.
Most occurrences were on first launch, with new LL DC assignments, on (already booked for a week to a month) prime95 V30.2b1 instances.

[chalsall]

Originally Posted by Prime95

The bug was in my parsing of the proof file header using fscanf. If the first byte of the first residue was ASCII whitespace, it thought the header was on byte longer than it really was. Error handling reading the last residue was not good, resulting in the 0.5 roundoff error.

This was in C? I'd forgotten (until recently) how much of a pain-in-the-butt string parsing is in C. I actually "made friends" with a RegEx C library just to do what I regularly do in Perl out-of-the-box.

Two quick points (I know you're juggling lots of stuff ATM):

1. rdt1 has been given (and successfully completed) three (3#) CERT jobs, even though the machine is not yet "trusted" (read: "Reliability, Confidence: 1.00, 0.0").

2. Might these heuristical metrics be influenced by such successful CERT runs? (I know; modeled, not yet coded).

Thanks George. This is ***so*** cool to observe!

[preda]

Originally Posted by chalsall

This was in C? I'd forgotten (until recently) how much of a pain-in-the-butt string parsing is in C. I actually "made friends" with a RegEx C library just to do what I regularly do in Perl out-of-the-box.

Two quick points (I know you're juggling lots of stuff ATM):

1. rdt1 has been given (and successfully completed) three (3#) CERT jobs, even though the machine is not yet "trusted" (read: "Reliability, Confidence: 1.00, 0.0").

2. Might these heuristical metrics be influenced by such successful CERT runs? (I know; modeled, not yet coded).

Thanks George. This is ***so*** cool to observe!

The problem was not "string parsing in C", it's the un-intuitive behavior of scanf() to keep consuming whitespace after any whitespace format character.
E.g. scanf("\n") will eat the newline *and* any follow-up whitespace (such as spaces, tabs, etc) that may follow that newline! Not exactly intuitive.

[Runtime Error]

I just got assigned 107805641 as a cert. I turned in the PRP via manual testing earlier. Is it working as intended that I can certify my own proof? Thanks!

Edit: Looks like the same thing happened here 107796071

[Prime95]

Originally Posted by Mark Rose

It looks like readme.txt from the mprime tarball hasn't been updated in a while.

I thought Windows and Linux had the same readme. I've updated the Windows readme so that it applies to both OSes.

[Chuck]

How does Prime95 manage the scheduling of a downloaded CERT when there is an existing PRP or DC running? Does it wait for some period of time to interrupt the running PRP/DC?

[Prime95]

30.2 checks for CERT work to do once an hour

[preda]

Originally Posted by Runtime Error

I just got assigned 107805641 as a cert. I turned in the PRP via manual testing earlier. Is it working as intended that I can certify my own proof? Thanks!

Edit: Looks like the same thing happened here 107796071

Yes the same person doing proof and CERT on the same exponent is fine. No weakness there.

PS: we should put together a FAQ on the new proof/cert thing. The above would be question No. 1 :)

[ixfd64]

Will the ability to obtain manual LL assignments be deprecated once 30.2 goes gold?

[Prime95]

S00030 where are you?

You have a lot of proofs to upload. Please grab v30.2 to start the process. Thanks.

[Prime95]

Originally Posted by ixfd64

Will the ability to obtain manual LL assignments be deprecated once 30.2 goes gold?

No decisions have been made. It's really 2 questions:

1) Do we want to be issuing first-time LL tests that will require double-checks?
2) Do we want to DC LL tests with LL or PRP-with-proof.

As to the later, the PRP-with-proof savings are small (we save an occasional triple-check). Do users that have submitted LL tests over the years have a desire to see if those first time tests were good or bad? If so, we should probably stick to LL-DC.

As to the first, what are your circumstances that make PRP with proof undesirable?
How much time do we want to give prime95 users to upgrade to v30?
If a user has limited free disk or limited upload bandwidth, is it fair for us to deny them first time tests?