[QUOTE=akruppa]I, ummm, kinda misremembered the date :alex: it is on the 10th in fact - monday.

Alex[/QUOTE]
Oh well, at least you misremembered in the correct direction. Turning up on the 10th for an exam that actually occurred on the 2nd would be embarassing.


Paul

The exam is over and went much better than I had expected. Many thanks to Prof. Ernst W. Mayr for a very relaxed and successful examination! :wink:

And I just got an email saying that I'll get an account on the Opteron cluster!!! :banana:

I like how this week is starting.

Alex

I've got the account on the cluster now. It's a 36 node cluster with quad-Opteron 850 nodes - 144*2.4 GHz of Opteron power. That's one bad-ass piece of number cruncher. And the crazy part is: according to the reservation schedule, this monster is idle most of the time. Sitting there, doing nothing... I can't believe it.

Time for some [I]serious[/I] factoring work.

Alex

Sweet! Enjoy this week while it lasts!

[QUOTE=akruppa]I've got the account on the cluster now. It's a 36 node cluster with quad-Opteron 850 nodes - 144*2.4 GHz of Opteron power. That's one bad-ass piece of number cruncher. And the crazy part is: according to the reservation schedule, this monster is idle most of the time. Sitting there, doing nothing... I can't believe it.

Time for some [I]serious[/I] factoring work.

Alex[/QUOTE]

Nice. What are your plans (after 5,349-?)? NFSNET will do 2,764+ after
2,761-.

I could do a lot with that kind of hardware....... :grin:

Do you have a parallel B-L running???

I just started 2,1366M. With my current hardware it will take ~10 days
to sieve, using about a dozen machines, and another week to do the LA
on a single machine... I plan on doing all 2,LM up to 2,1402L.

I asked Peter and Herman for the parallel block-Lanczos version of gauss. When I get that, I'll probably spend a few days compiling it for the Opteron/Infiniband cluster and testing. I once tried to compile their line siever for Opteron, but it didn't recognise the 64 bit architecture. If gauss has a similar problem, I may have to tweak a few things to get it to full speed, so this phase may take a while.

Once the parallel block-Lanczos runs satisfactorily, I can offer matrix solving service to other factorers.

I test filtered the 5,349- relations today (dupe removal only), there are ~110M unique relations. I expect I'll need about 150M, so sieving will take a bit more time. After that, I don't know yet - the Athlon64 and Opteron are excellent at GMP work, so I'll probably do some ECM work. I'm kinda attracted to the OPN research right now, I may do 3221^73-1 later if there's no ECM factor. This could be done with volunteer sievers. Other than that, I may do a bit of ECM and NFS on the base 3 Cunningham numbers. It would be nice to clear the tables up to exponent 500.

Alex

Edit: Oh, and I'm working with Paul Zimmermann on Aliquot sequences. Those will need a lot of ECM and GNFS, too.

Bummer. The development of CWI's parallel Block Lanczos was largely sponsored by Microsoft Research and is not available to outside parties. Maybe I can use Jens Franke's.

Alex

Did a test pruning today, there are ~2.5M more ideals than relations. With the partial output files that weren't included in this test, there should probably be enough for a dependency, but I'll go for a lot of excess.

Nothing new on the parallel block-Lanczos matter. Now I have access to the perfect hardware for the matrix step, but no software to run on it. :sad:

Alex

Just for satisfying my curiousity:

What is the status of the 5,349- SNFS factorization, Alex?

I did another test filtering a few days ago and it seemed that I had enough relations for a dependency. I'll sieve a while longer, though - I want a lot of excess to get a smaller matrix. I still have no access to a parallel Block-Lanczos implementation and apparantly won't get one... I may have to do the matrix on a single quad-Opteron node, provided that I may reserve one for a few months in one piece. So nothing dramatically new so far... I'll post a report when I start post-processing.

Alex

Did another test filtering today. I modified the CWI filter program to run on Opterons in 64 bit mode, it can handle primes up to 2^31 on those (on x86 something broke with primes greater than approx. 1G). I needed that since I used lp=2^31 for the sieving, as this considerably increased yield.

CWI's filter program produces much more accurate ideal counts than the home-grown code I used for earlier test filter runs. The bad news: the excess is much less than I had hoped, only about 4M. After clique removal, there are about 35M ideals left, so merging will reduce this to a matrix of maybe (7M)^2 at best - probably larger than that. This is too large in any case.

Need to do more sieving. I used to sieve on a couple of Athlons, but I think I'll move everything over to the Opteron cluster. It's time to get this over with.

Alex