View Single Post
Old 2018-04-23, 08:44   #5
GP2
 
GP2's Avatar
 
Sep 2003

13×199 Posts
Default

Quote:
Originally Posted by VBCurtis View Post
You could run SNFS on it. If you get an Amazon AWS account, you could likely solve it within the year for $10,000 or so. Ok, maybe $20k.

There's a stub of an explanation at http://www.mersennewiki.org/index.php/SNFS, and googling "special number field sieve" will give you more information. SNFS is *not* a "you might find a factor" algorithm; when the process completes, factors will be known. However, we're talking a quite substantial effort, like hundreds of cores for over a year. A cluster would be helpful to solve the matrix; a single machine with 64GB memory might not be sufficient, and would likely take over a year on its own.

There's a reason we haven't tried this job as a forum team; we can marshal 200+ cores to the task, but nobody has the cluster and interest to solve the matrix.
An r4.4xlarge instance has 122 GiB of memory and 8 cores; an r4.2xlarge has 61 GiB and 4 cores. Would that single instance in lieu of a cluster be able to solve the matrix? (and does the number of cores make any difference?)

The spot price is about 15 cents an hour for the r4.4xlarge or about 7.5 cents an hour for the r4.2xlarge (us-east-2, Ohio). Assuming those prices hold steady (no guarantee, but they have for the past 3 months), then that would be $1315 per year, or $660 per year, respectively.

If the matrix solving waits for the preliminary SNFS to complete, then it would begin over a year from now and last over a year, and hourly spot prices would likely be even lower in the future, or there might be r5 instances by then.

Obviously spot instances are interruptible. However, instances with less than 100 GB of memory can be set to hibernate on interruption, rather than stop or terminate. That might be an issue if the larger r4.4xlarge was used, since it's too big to hibernate.

I'm not volunteering to actually do this, just trying to see if I understand correctly. You say that the matrix solving is the stumbling block but it seems that running 200+ cores for over a year is by far the bigger expense, whereas a thousand bucks or two seems doable for a determined individual with money to burn, assuming those hundreds of cores were marshalled for free.
GP2 is offline   Reply With Quote