[QUOTE=aketilander;357123]As you remember I did some initial DC. Here are the results from my DC as far as I did it:[/QUOTE]

I had a glance at the attachment and immediately saw something "off":

[quote]Using AVX Core2 type-3 [b]FFT length 19200K[/b], Pass1=768, Pass2=25K, 4 threads. (3 threads)

[Fri Jun 01 04:33:42 2012]
Iteration 10000000 / [b]595999993[/b][/quote]

You do realize that corresponds to over 30 bits per double, thus there is not an iota of hope of the results being correct, yes? Are you running with ROE checking disabled? If so, I strongly suggest you enable it and rerun the first few dozen iterations. If not, I have no idea why you did not see a fatal ROE almost immediately.

My Mlucas code won't even let a user run at such a woefully undersized FFT length - I don't support 19200K so tried at the next-larger FFT length my code supports, 20480K:
[i]
time ./Mlucas -m 595999993 -fftlen 20480 -iters 100 -radset 0 -nthread 4
...
INFO: maximum recommended exponent for this runlength = 374233313.
ERROR: at line 1107 of file ../Mlucas.c
Assertion failed: specified FFT length 20480 K is much too small: Recommended length for this p = 36864 K.[/i]

[QUOTE=ewmayer;374837]You do realize that corresponds to over 30 bits per double, thus there is not an iota of hope of the results being correct, yes? [/QUOTE]

I agree this looks bad. It is possible to force prime95 to use a 19200K FFT. However, when I did this I get a ROUNDOFF error on every iteration.

[QUOTE=LaurV;374754]Just to let you know that 120M iteration is passed (the 34% mark, or one third of the work) and it checks up.
(edit: if it would run continuously, the ETA would be ~39 days, but expect it like August or so, I "[URL="http://www.mersenne.org/report_exponent/default.php?exp_lo=345678877&exp_hi=&full=1"]legally assigned" on Apr 28[/URL], but only work it during the nights, this is a Black-Titan work, air cooled, and I have heat-troubles during the days, initial ETA was 59 days)[/QUOTE]

Here's where I'm at:

Iteration 126430000 / 336004343, 0x6e9fedb8994c8cd7

[CODE]
CUDALucas lines found: 1
Processing result: M( 345678877 )C, 0x6baacca0093e2f58, offset = 5773, n = 19600K, CUDALucas v2.05 Beta, g_AID: D8C..deadbeef0ddf00dbadc0ffee
LL test [COLOR=Red]successfully completes[/COLOR] double-check of M345678877
CPU credit is [COLOR=Red][B]5962.9606[/B][/COLOR] GHz-days.[/CODE]I think this is the largest exponent to date which was [U]independently[/U] double checked (i.e. not the same user or team doing both LL and DC, me and Ake live to totally opposite sides -geographically and climatological - of the world and only talk each-other on this public forum). And of course, I would not start it if not for Ake's list of residues, which helped me a lot along the way (read this thread from the start to see what I am talking about). I started few months ago and only worked few hours each night, during cool hours. This is a GPU test, Titan card, air cooled (yet! I have the water block in the cabinet! but no time for remodeling), it can't stand the heat during the day, and interesting enough, there was no error during the test. Everything went very smooth, in spite of hundreds of interruptions and changing (sometimes on the fly) from DP to SP and viceversa (during the day the card was doing LL-TF for GPU72, where, by the way, I will get no credit for this big work... :sad: Hey Chris can you give me at least few breadcrumbs? :wink:). It did not give any error, all residues matched along the way, so there was no "backtrack" needed.

When we get to the point where it is feasible for us to do 332M exponents, I think we're going to have to triple check some of these. I just really don't trust results on such high numbers that very few people actually care about.

Actually a lot of people do care about these large numbers, because there is where the EFF moneys are. And the tests are very "feasible" for everybody who has a computer half of the computer you have, Prime95 can crunch through 332M LL without any troubles. What the people don't have is the patience to wait months for the result to come, and eventually wait a double time, to do DC in parallel and make sure their calculus is right. So, instead of two tests, LL and DC, you have to do 3, like LL+DC first time, to be sure of the first submission, then having the list of residues, someone else is doing the "proper" DC. This is not for the faint hearts...

[edit: I don't say this is the proper way, but this is what we have done, this exponent was already DC-ed by Ake, and I used his residue list to "sustain my moral" during the DC test]

[QUOTE=TheMawn;380048]When we get to the point where it is feasible for us to do 332M exponents, I think we're going to have to triple check some of these. I just really don't trust results on such high numbers that very few people actually care about.[/QUOTE]

There was a similar spike in errors during the earliest ten-million-digit tests, which were run in the late '90s when typical assigned exponents were closer to 5M-10M. Quite a few triple- (and higher-order) checks were needed once 33M assignments became mainstream.

[QUOTE=TheMawn;380048]When we get to the point where it is feasible for us to do 332M exponents, I think we're going to have to triple check some of these. I just really don't trust results on such high numbers that very few people actually care about.[/QUOTE]

I'm sure you've also seen the strongly held opinion that ECC RAM is virtually mandatory for these long runs.

[QUOTE=kladner;380062]I'm sure you've also seen the strongly held opinion that ECC RAM is virtually mandatory for these long runs.[/QUOTE]

I guess it depends on "how long", really...

[QUOTE=LaurV;380046]Everything went very smooth, in spite of hundreds of interruptions and changing (sometimes on the fly) from DP to SP and viceversa (during the day the card was doing LL-TF for GPU72, where, by the way, I will get no credit for this big work... :sad: Hey Chris can you give me at least few breadcrumbs? :wink:)[/QUOTE]

I (truly) don't understand this statement.

Was it a joke (or a language disconnect), or is there actually an issue?

[QUOTE=chalsall;380078]I (truly) don't understand this statement.

Was it a joke (or a language disconnect), or is there actually an issue?[/QUOTE]

He's (humorously) pointing out that he will get no GPU72 credit for the huge LL test, as GPU72 only tracks *factoring* credit and time saved, rather than straight LL credit, which is already handled by PrimeNet. At least that is what he *seems* to be getting at.