mfaktc: a CUDA program for Mersenne prefactoring

[bcp19]

Originally Posted by flashjh

You're right. I wouldn't put more than one core per instance, but I think his GPU could handle more.

It depends on the setup. I have a Core 2 Quad that runs 2 instances on a GTS 450. Main reason why is that if I run 1 instance, I get 80M/s, where with the 2 running each instance gets 60M/s. Since I don't like to let SievePrimes move from 5000, my CPU's have 10+% of wait time, so I actually run P95 on all 4 cores, but the 2 sharing with mfaktc take about 8-10 times as long to run. My 2400 runs a GTX 560Ti and has the same setup, but the shared cores only take about 3x as long. It runs about 180M/s single instance, 260M/s combined.

[TheJudger]

Originally Posted by nucleon

What I've noticed is that v0.18 needs a little more cpu grunt to max out GPUs.

Yepp, the GPU code is a little bit faster in 0.18 thus you'll need a bit more CPU power (or lower SievePrimes) to keep the GPU busy. There will be another little boost for especially for CC 2.0 chips once CUDA 4.1 is available.
http://www.mersenneforum.org/showpos...postcount=1328
http://www.mersenneforum.org/showpos...postcount=1368
http://www.mersenneforum.org/showpos...postcount=1409

Oliver

[Dubslow]

Heh. I've noticed my cpu wait go up, not down. I'll try and figure it out later.

[MrRepunit]

Hi,

how much effort would it be to extend mfaktc to handle (generalized) repunits? The factors have the same (or similar) properties as Mersenne primes: 2kp+1.
If it is not too much work I would like to put this on the wish-list for mfaktc.

Thanks...

[TheJudger]

Hi!

Originally Posted by MrRepunit

Hi,

how much effort would it be to extend mfaktc to handle (generalized) repunits? The factors have the same (or similar) properties as Mersenne primes: 2kp+1.
If it is not too much work I would like to put this on the wish-list for mfaktc.

Do you have any more information about generalized repunits (math)?

Oliver

[MrRepunit]

Originally Posted by TheJudger

Hi!

Do you have any more information about generalized repunits (math)?

Oliver

Here you go:
http://www.cs.uwaterloo.ca/journals/...NER/dubner.pdf
http://www.maths.tcd.ie/pub/ims/bull59/R5901.pdf
http://en.wikipedia.org/wiki/Repunit

I hope this is what you wanted.

Danilo

[Dubslow]

Originally Posted by TheJudger

Can you try to upgrade the driver: devdriver_4.0_linux_64_270.41.19.run


Well, if everything is right the CUDA SDK and CUDA Toolkit is not needed to run the precompiled mfaktc executable. You'll need a proper driver, nothing more.

Oliver

Here's the issues I've had over the last 4 months or so with drivers.
http://forums.nvidia.com/index.php?showtopic=220802

[TheJudger]

Hi Danilo,

Originally Posted by MrRepunit

Here you go:
http://www.cs.uwaterloo.ca/journals/...NER/dubner.pdf
http://www.maths.tcd.ie/pub/ims/bull59/R5901.pdf
http://en.wikipedia.org/wiki/Repunit

I hope this is what you wanted.

Danilo

OK, might take some days until I look into it.

Oliver

[MrRepunit]

Originally Posted by TheJudger

Hi Danilo,

OK, might take some days until I look into it.

Oliver

Many thanks in advance!

[flashjh]

Originally Posted by TheJudger

Hello!

http://www.mersenneforum.org/mfaktc/mfaktc-0.18.tar.gz
http://www.mersenneforum.org/mfaktc/mfaktc-0.18.win.zip
http://www.mersenneforum.org/mfaktc/...linux64.tar.gz

The executables need at least a CUDA 4.0 capable driver (270 series driver or newer). The Windows zip archive contains both, the 32 bit and 64 bit version. I'll upload new executables once CUDA 4.1 is public available. The sources should compile with older CUDA version, too, but they might be slower. CUDA 4.1 will give another performance improvement for the barrett based kernels on compute capability 2.x GPUs (especially on 2.0).


Compared to mfaktc 0.17 there are "more than usuall" minor changes. Highlights from the Changelog.txt:

• autoadjustment of SievePrimes is now less dependend on the gridsize and
  absolute speed. Instead of measuring the absolute (average) time waited
  per precessing block (grid size) now the relative time spent on waiting
  for the GPU is calculated. In the per-class output "avg. wait" is replaced
  by "CPU wait".
• new commandline option: "-v" (verbosity) let the user decide how many
  informations are printed
  (suggested by aspen on www.mersenneforum.org)
• "has a factor" result lines now contain informations (program name,
  versions, bitlevel, ...) James Heinrich is working on this on the server
  side. This should give more accurate credits for "has a factor" results
  from the primenet server once this is fully implemented.
• mfaktc no longer refuses to load a checkpoint file from a Linux version
  with a Windows version of mfaktc and vice versa. Of course mfaktc still
  refuses to load checkpoint files from other versions than itself
  (identical version string!)
• added a (simple) signal handler (captures SIGINT and SIGTERM).
  1st ^C: mfaktc will exit after the currently processed class is finished.
  2nd ^C: mfaktc will stop immediately
• added a minimum delay between two checkpoint file writes. The user can set
  the delay in mfaktc.ini (CheckpointDelay).
• added a new code path to barrett79_mul32 and barrett92_mul32 kernels, CUDA
  >= 4.1 features multiply-add with carry for compute capability >= 2.0.
  On my GTX 470 (compute capability) this yields up to 15% for
  barrett92_mul32 and up to 7% for barrett79_mul32 extra throughput.

As usuall: finish your current assignments with your current version and do the update after it, mfaktc 0.18 will refuse foreign checkpoint files.

Oliver

Ok, so I'm having some problems. I've been running .18 for some time now, 5 instances with no problems.

Then, last couple of days only instance 1 has started crashing, no error window or anything -- it's just closes. The other problem is that it's duplicating work. For no apparent reason it will do the same factor 2 or 3 times at the same bit level (which is a huge waste) and it always crashes after a duplicate so it's wasting even more time. I tried deleting the .exe and using a fresh copy with no luck. I haven't changed anything else, so I don't know what it could be. The other four instances work fine, no lock-ups or closes or duplicate work. I at a loss here. Anyone having similar problems? Using Win7 64bit with two GTX 580s.

[LaurV]

Originally Posted by flashjh

it's duplicating work. For no apparent reason it will do the same factor 2 or 3 times at the same bit level

smells like no rights to write the worktodo file, is that opened by somebody else? did you change the rights? file attributes? (hidden, system?)
mfaktc needs to access worktodo file to modify the first line, therefore eliminate the work already done. If it can not do that, it will certainly repeat the work.