Another new prime (M42-Related)

[T.Rex]

I'll see next Monday if I can be authorized to use a 16x or a 32x machine with new ia64 processors. With the GLucas program of Guillermo Ballester Valor, verifying this new prime candidate should not take very long. One week ?

Look at this paper: Tera10 SuperComputer and imagine how many Mersenne primes could be found in one year with such a machine if a clustered Mersenne crunching program were available. (and if some fool would authorize us to use it !). Just a dream.

Guillermo, do you think it is possible to modify GLucas so it can be spread over several machines by means of MPI ? If I remember well, this kind of machine can exchange 4 GB of data per second.

Tony

[gbvalor]

Guillermo, do you think it is possible to modify GLucas so it can be spread over several machines by means of MPI ? If I remember well, this kind of machine can exchange 4 GB of data per second.

Frankly, I don't know. I must to know more about MPI. Any good and easy link about it?

Guillermo

[Jeff Gilchrist]

Originally Posted by gbvalor

Frankly, I don't know. I must to know more about MPI. Any good and easy link about it?

Actually MPI would be very useful for me as well. There are limited SMP machines at my disposable but we have clusters with hundreds of machines.

Main MPI page: http://www-unix.mcs.anl.gov/mpi/index.html

MPI Complete Reference: http://www.netlib.org/utk/papers/mpi-book/mpi-book.html

Jeff.

[Peter Nelson]

Originally Posted by T.Rex

Look at this paper: Tera10 SuperComputer and imagine how many Mersenne primes could be found in one year with such a machine if a clustered Mersenne crunching program were available. (and if some fool would authorize us to use it !). Just a dream.

Tony

Well, as you're located in France I can see why you might be happy about this being a French project.

Unfortunately they're outclassed already.

See the top 500 supercomputer list at www.top500.org

This shows IBM Blue Gene having about 70 TFlops as compared to the Tera10 proposal of about 60 TFlops.

What is more important - it's here now whereas Tera10 has to wait for new dual core Itaniums to be shipped, and then in some volume. They need 544 x 8 =4352 of the new chips.

All the other Intel based clusters in the list could simply replace their current single core cpus with dual core too, perhaps almost doubling their performance which would keep many of them well ahead of the French project.

As regards message passing, etc low latency network is good and many of the top500 use myrinet, infiniband or similar high speed interconnect.
The Tera also seems to use 8 way Itanium (as SMP) which potentially has shared memory faster than the net interconnect.

A parallel implementation of the FFT might be possible but the aggregate throughput would remain similar per processor because Lucas-L is a sequential iteration task.

Such a supercomputer would simply be acting like lots of PCs. It might be better applied to a task which requires the interconnect and SMP, like fluid dynamics, weather forecasting, nuclear simulations etc.

Although, not fastest in the world, Tera10 will be fastest in France.

By the time it gets built I imagine quite a few developments both in new cpus and other organisations building bigger things. eg what about clusters of Sony's Cell processor?

[wpolly]

DB-Library: Unable to connect: SQL Server is unavailable or does not exist. Specified SQL server not found.
Net-Library error 2: ConnectionOpen (CreateFile()).

Primenet database seems to be down --- So I can't do the data analysis. :(

[PrimeCruncher]

Originally Posted by ixfd64

I think it's 25964951 because the residue is very similar:
7EC80125B6DB7E (24036583)
7EC80136C8136C

I agree with this assessment. A search of hrf3 and lucas_v returns one other residue that contains this string.

[T.Rex]

Originally Posted by Peter Nelson

Well, as you're located in France I can see why you might be happy about this being a French project.
Unfortunately they're outclassed already.
....
what about clusters of Sony's Cell processor?

I'm very proud of this Tera10 project not only because it is built by my company but also because one year and a half ago the Bull HPC team knew very few about what a super-computer is. They made terrific progress. They have been chosen not only because of processor performance, but also for large bandwith and reduced latency between 2 machines, and for many other reasons.
About Cell, I've quickly read some papers and yes it seems this processor is terrific !
About using MPI within GLucas, the performance clearly depends on the memory dependencies between all GLucas threads and also on the latency and bandwidth between each computer in the cluster. Would a MPI-based GLucas be useful ?

Tony

[T.Rex]

Originally Posted by gbvalor

Any good and easy link about it?
Guillermo

I'll ask my colleagues next week. They should recommend me some paper.
Tony

[jinydu]

Well, don't we have anything more to say about this?

[jinydu]

I've found something!

http://mathworld.wolfram.com/news/2005-02-18/mersenne/

Nothing on www.mersenne.org yet...

[Prime95]

Originally Posted by jinydu

I've found something!

http://mathworld.wolfram.com/news/2005-02-18/mersenne/

Nothing on www.mersenne.org yet...

Today's update:

No reply yet from the discoverer. Heck of a time for a vacation! Or maybe the email address was a work address.

The Unix double-check is up to iteration 3 million.

I haven't updated mersenne.org because I'm being cautious. The last two primes I was able to rerun the last iterations using the discoverer's save file to be vertually certain of the discovery. I won't be able to do that this time. I haven't even seen the results.txt file to see if there were any errors reported during the run. On the plus side this particular computer has done several LL tests all error free.

Based on rather limited historical data, pre-v23.4 prime95 reported 6 Mersenne primes and only 5 were really prime. So, I was at least waiting until I saw the results.txt file before updating the main web page. I'm even a little uncomfortable with the preliminary Wolfram and slash-dot coverage.