[QUOTE=Xyzzy]
I haven't been keeping track of the 55 digit stuff I've posted, but hopefully thomasn will continue to keep stats for us![/QUOTE]

I will do that, and I am sure ET_ will be able to take over if neccessary. :smile:

Thomas

:coffee:

[Fri Nov 26 01:50:15 2004]
M1061 completed 200 ECM curves, B1=44000000, B2=4290000000
[Sun Nov 28 23:05:16 2004]
M1061 completed 200 ECM curves, B1=44000000, B2=4290000000

Thomas
:coffee:

[CODE]
Name Curves Multiplier B2 Message No Line total
Xyzzy 5 ? ? #49 5
Xyzzy 50 1 680270182898 #76 50
Marc 1 1 680270182898 #82 1
Xyzzy 1 1 680270182898 #88 1
Xyzzy 105 1 680270182898 #94 105
Xyzzy 17 1 680270182898 #100 17


Grand total 179

[/CODE]

Here is the list of work done for B1=110M. As you can see we have a long way to go if we want to complete that level.

Also: Please remember that we still need about 2000 curves to finish B1=44M.

Thomas
:coffee:

Is the multiplier really 1? After all, these curves have been done with gmp-ecm and higher bounds than 100 * B1...

243 curves using B1=110000000 & B2=680270182898...

Note the interesting time differences:

64-bit install:
Step 1 took 2050755ms
Step 2 took 2291610ms

32-bit install:
Step 1 took 3586745ms
Step 2 took 3287874ms

I installed a 32-bit version of Debian just to test this, and the time difference is pretty amazing... I let both run for several days to make sure that the numbers were stable...

[QUOTE=Xyzzy]243 curves using B1=110000000 & B2=680270182898...

Note the interesting time differences:

64-bit install:
Step 1 took 2050755ms
Step 2 took 2291610ms

32-bit install:
Step 1 took 3586745ms
Step 2 took 3287874ms

I installed a 32-bit version of Debian just to test this, and the time difference is pretty amazing... I let both run for several days to make sure that the numbers were stable...[/QUOTE]


Amazing! :w00t:

Now there's a good reason to exchange S.O.

Luigi

To be fair, the 32-bit install was using the i386 kernel and not a 686 or K7 kernel... I'm not sure how much of a difference is from the kernel and how much is from having the package compiled with a 64-bit only compiler... The 64-bit install was a pure 64-bit install...

Maybe I'll run a quick benchmark with the k7 kernel installed...

[QUOTE=Xyzzy]Maybe I'll run a quick benchmark with the k7 kernel installed...[/QUOTE][code]GMP-ECM 5.0.3 [powered by GMP 4.1.4] [ECM]
Input number is 24707306311927565716857342128774085333197833223161879682238935306082805123046306993647507776054336486228891340858985829027076261887914242781617846672453431386903982455635542158748401823985988322905245077938567513252198179128990807936780194781391547404884040101606295111368825026273254703636026307207764436438929167613951 (320 digits)
Using B1=110000000, B2=680270182898, polynomial Dickson(30), sigma=1217266501
Step 1 took 3578186ms
Step 2 took 3293512ms
Run 2 out of 0:
Using B1=110000000, B2=680270182898, polynomial Dickson(30), sigma=3163474637
Step 1 took 3567709ms
Step 2 took 3284253ms[/code]
Nope, it is just as slow with the K7 kernel...

Edit: Oh yeah, I got 3 more curves done... Same bounds as all the others...

[QUOTE=Xyzzy]Nope, it is just as slow with the K7 kernel...[/QUOTE]
It is probably the libgmp that makes it slow, the pure 64 bit install will have a well optimised libgmp whereas the 32 bit install will have a generic one.

I think it's the other way round: in 32 bit mode, gmp uses the 32 bit K7 optimized code, whereas in 64 bit mode it uses the generic C code with 64 bit types. For small numbers GMP uses the basecase (grammar-school) multiplication code whose time is quadratic in the number of machine words, so 64 bit mode has a factor 4 advantage there. That Xyzzy sees "only" a factor <2 speedup probably comes from lower mul-per-clock throughput the C code gets. That the relative speedup in stage 2 is smaller probably comes from the higher ratio of O(n) opertaions (adds, shifts) over muls in stage 2 compared to stage 1.

Torbjorn Granlund once mentioned that the Hammer (as it was called back then) could in theory get twice the mul throughput of the K7, for a total eightfold speedup when multiplying small numbers. I don't know how optimistic this estimate is, but AMD64 seems to make a good architecture for GMP related work either way - wish I had one. Unfortunately, full AMD64 support in GMP is scheduled only for late 2005.

Alex

Gentlemen, give this man a cigar..... Umm an AMD64 :devil: