[url]http://www.sciencedaily.com/releases/2016/01/160120084917.htm[/url]

picked up from ddg'ing: mersenne prime 2016 :smile:

[URL="http://phys.org/news/2016-01-largest-prime.html"]Phys.org[/URL]

Congrats to all!
 

Just showed up on the front page of news.google.com - with the horribly-titled WaPo story in the lead.

[QUOTE=Madpoo;423221]The real test is how many of those will stick it out and finish at least one test. Here's hoping it's a decent amount.[/QUOTE]

Unfortunately, if the experience from M48 is any indication, fewer than 1% of all _assignments_ will be completed (and less than ~5% will even have a single cycle applied).

Many people have no idea just how long this kind of thing takes, and get discouraged quickly.

That being said, those who do take an interest tend to stick around, and after the "surge and churn" (a bit like "surf and turf", but different) stabilizes our net throughput should be higher.

Also, it's really good that George, James and you (Aaron) implemented the new assignment rules, and the "RIP DCTF" team got /really/ busy. It's working out perfectly.

[URL="http://www.theverge.com/2016/1/20/10797052/biggest-prime-number-gimps-missouri"]TheVerge[/URL]

[QUOTE]Although the hardware in this case isn't exactly supercomputer level — the prime was discovered after "31 days of non-stop computing" by Dr. Curtis Cooper of the University of Central Missouri on computer kitted out with just an Intel Core i7 processor.[/QUOTE]


I was thinking 31 days on a Core i7-4790 (Haswell) seems very long and the benchmarks page confirms it should run 18-20 ms/iter at that FFT which fits with 15-16 days, so he must run 2 exponents at once. Can that be confirmed on the server? Maybe that should have been in the press release, that it only used half of its cores.
[url]http://www.mersenne.org/report_benchmarks/?exp_date=2014-01-01&64bit=1&exover=1&exbad=1&exv25=1&exv26=1&specific_cpu=4377497[/url]


Was Curtis Cooper informed exactly which 4 computers found his 4 primes? I wonder if he marked it on them. I also wonder how he got permission to run it on so many computers.

[QUOTE=chalsall;423240]...That being said, those who do take an interest tend to stick around, and after the "surge and churn" (a bit like "surf and turf", but different) stabilizes our net throughput should be higher...[/QUOTE]

Just looked...

So far there are just < 200 assignments from new users that are first-time LL in the 77.6M - 79M range. About 65 new assignments for 100M digit stuff (they don't know the investment in resources for those yet...some have ETAs of 6-7 years from now and some are running multiple 100M workers on each core of one CPU, which is horrible for performance typically).

The surprising bit were the 2300+ DC exponents assigned to new users, anywhere from 39.5M up to just below 44M. Those have a better chance of finishing since the time involved is smaller. DC is pretty far behind the first-time checks so fingers crossed that helps clear the backlog a tiny bit.

[QUOTE=ATH;423243]I was thinking 31 days on a Core i7-4790 (Haswell) seems very long and the benchmarks page confirms it should run 18-20 ms/iter at that FFT which fits with 15-16 days, so he must run 2 exponents at once. Can that be confirmed on the server? Maybe that should have been in the press release, that it only used half of its cores.
[url]http://www.mersenne.org/report_benchmarks/?exp_date=2014-01-01&64bit=1&exover=1&exbad=1&exv25=1&exv26=1&specific_cpu=4377497[/url][/QUOTE]

It may only use one core per worker, but I haven't checked specifically. Right now that same system only has one exponent assigned to it, so perhaps it has all of the cores in the single worker.

If that's the case, the extra lag between assignment and completion is probably that it didn't start working on it right away, which would actually be pretty normal.

[QUOTE]Was Curtis Cooper informed exactly which 4 computers found his 4 primes? I wonder if he marked it on them. I also wonder how he got permission to run it on so many computers.[/QUOTE]

He knows which computer found this most recent one. Hopefully he tracked it down and put a shiny gold star on it. :smile: For the permission thing, I don't know if he's specifically in charge of the computer labs but I imagine he has some pull with whoever that is, and the university itself has received good press as a result of his discoveries over the years. I imagine they enjoy the prestige.

[QUOTE=ATH;423243]I was thinking 31 days on a Core i7-4790 (Haswell) seems very long and the benchmarks page confirms it should run 18-20 ms/iter at that FFT which fits with 15-16 days, so he must run 2 exponents at once. Can that be confirmed on the server? Maybe that should have been in the press release, that it only used half of its cores.
[url]http://www.mersenne.org/report_benchmarks/?exp_date=2014-01-01&64bit=1&exover=1&exbad=1&exv25=1&exv26=1&specific_cpu=4377497[/url][/QUOTE]

On my 3.6 GHz Core i7 4790, I got an estimated ETA of about 7.25 days using all 4 cores. So I expect UCM's computer was using a single core for M49, and running 4 tests at a time.

[QUOTE=cuBerBruce;423247]So I expect UCM's computer was using a single core for M49, and running 4 tests at a time.[/QUOTE]

And / or, the computer wasn't actually running 24/7, and / or was busy doing "real work".

[QUOTE=chalsall;423250]And / or, the computer wasn't actually running 24/7, and / or was busy doing "real work".[/QUOTE]

I'm sticking with the notion that it got the assignment while it still had a week or two to go on the one it was currently working on, just queuing it ahead of time.

Although, that system does have a pretty bad rolling average right now, just 947. So that could indicate sub-optimal performance if it's trying to test 4 exponents at once in the 70-80M range.

Sure, one worker with one core could probably test M49 faster than a month, but what if you had 4 workers each testing exponents around that same size? Still < 1 month for each of them?

[QUOTE=ATH;423188]Piping the output of this into a file took 6 sec:[/QUOTE]

OK, "at least 1000x slower than it should be" - yes, I *know* my decimal-conversion code sucks ... I have spent nearly all my time focusing on number-crunching speed, rather than pretty-printing. (I.e. having a much-faster-than-GMP FFT-mul is worth neglecting the fancy-I/O stuff to me).