Continuing P-1 from mprime savefile with GMP-ECM?

[GP2]

Originally Posted by lycorn

Are you then considering to start looking for first factors? That would be good, as they are becoming increasingly difficult to find for exponents this small. All helping hands are welcome...

Found a first factor for M33871, namely 13749988937505141870311843033629075511 with 38 digits.

[lycorn]

Wow! Good catch.
How long did it take, in actual clock time?
I´ve never used GMP for P-1 but I´ll probably give it give it a go one of these days.

[GP2]

Originally Posted by lycorn

How long did it take, in actual clock time?

204 minutes for stage 1, and 760 minutes for stage 2, so a bit more than 16 hours in all, using a single core.

Edit: I forgot to mention, peak memory usage during stage 2 was 54881 MB.

[Gordon]

Originally Posted by GP2

204 minutes for stage 1, and 760 minutes for stage 2, so a bit more than 16 hours in all, using a single core.

Edit: I forgot to mention, peak memory usage during stage 2 was 54881 MB.

I ran 400 curves at 3M/400M on that one...

How did you get it to use 54 gig without crashing? I have to restrict it to maxmem 14000 on a 32 gig machine (Win 7 Pro).

[GP2]

Originally Posted by Gordon

I ran 400 curves at 3M/400M on that one...

How did you get it to use 54 gig without crashing?

I am using AWS (Amazon cloud) for almost everything, other than one home machine.

In this case I recently started trying out the x1.16xlarge instance, which has 32 cores and 976 GiB of memory. At spot instance prices it costs about $1 / hour.

Needless to say, it is very labor intensive to try to constantly juggle multiple jobs on it to try to get your money's worth out of those cores and memory, so I'm looking forward to the recently announced AWS Batch, which will simply let you specify a batch job, and Amazon themselves schedules and runs it. However it's not yet available.

PS,
there is also an x1.32xlarge instance with double the cores and memory of the x1.16xlarge.

PPS, there are also r3 instances (and very soon r4), for instance r3.2xlarge with 4 cores and 61 GiB of memory, for a correspondingly hourly cost. However, the r3 instances only have about 15 GiB per core whereas the x1 instances have about 30 GiB per core, and in general the x1's have the lowest price per GiB of memory. So they are more cost effective, but you do have to queue up a bunch of work all at once in order to use them cost-effectively. I run a bunch of P−1 stage 1 on other machines beforehand.

PPPS, one huge potential issue is that this is only remotely affordable at spot instance prices, since the on-demand price is about four or five times more. However, spot prices can spike unexpectedly and terminate your instance with only two minutes notice. I don't think gmp-ecm lets you save and checkpoint during stage 2, not to mention the infeasibility of trying to write about 1 TB from RAM to disk in less than two minutes. Fortunately spot prices are relatively flat for the time being, but that can change unexpectedly at any time, and flush several days' worth of incomplete calculations down the drain... again, AWS Batch may be a godsend when it becomes available.

[GP2]

Found another first factor:

M39409 has the 42-digit factor 811625603090654894157972094386309926337383

Peak memory was 63793MB, this time stage 2 took only 250 minutes because it completed after k = 1 of an expected total of 3.

Anyway, any further discussion can go in the thread in the Data forum. For the purposes of the Software forum, it has been established that an P−1 savefile from mprime can be converted into a form suitable for stage 2 by gmp-ecm, in much the same way as for ECM files.