LMH status

[delta_t]

Originally Posted by TheJudger

here's a version of mfaktc which might be useful for short running assignments: http://www.mersenneforum.org/mfaktc/...essClasses.zip

Hello Oliver,

Thanks for posting this version. Definitely much faster.
I've done some timings on Factor=298241213,64,65 using a GTX 460M, with both the default and this new build.

Here they are timings vs the default win64 0.17 with MORE_CLASSES enabled
SievePrimes=5000, NumStreams=4, CPUStreams=4

GridSize=0
More classes = ~22.5 s
Less classes = ~17.1 s

GridSize=1
More classes = ~26.3 s
Less classes = ~16.7 s

GridSize=2
More classes = ~31.6 s
Less classes = ~17.3 s

GridSize=3
More classes = ~39.8 s
Less classes = ~18.4s

I turned also turned Checkpoints=0 since there is no need for them with such short runs.

[Chuck]

This is excellent for my small TF runs in the 600M range; after switching to the MORE_CLASSES disabled version (LESS_CLASSES ), my time per run has been reduced from 45 sec to 33 sec. Oh and I turned off checkpoints too. I had completely forgotten about that setting.

Chuck

[petrw1]

Originally Posted by petrw1

Could be at 300M by Oct 13 ...

Just squeaked in on the 10th ... 12 days for this last range.

[NBtarheel_33]

Originally Posted by TheJudger

Hello LMH,

here's a version of mfaktc which might be useful for short running assignments: http://www.mersenneforum.org/mfaktc/...essClasses.zip Because I'm no longer bound to the filesize limit in this forum this includes both, Win32 and Win64 executables and the CUDA runtime DLLs.
The only difference to the regular mfaktc version is that the compile option MORE_CLASSES is disabled. Thus there are only 96/420 classes instead of 960/4620. Depending how small the assignments are you might want to lower GridSize in mfaktc.ini, too. (Check runtimes, not the raw rate when comparing different settings!)
Another hint: autotuning SievePrimes might be not optimal for such short running assignments, too. You can try to set it manually in mfaktc.ini.

Oliver

Thanks, I will try this out. This is exactly the behavior I was observing when I was doing 292M runs the other day. SievePrimes=5000 was faster (i.e. 107M/sec) than SievePrimes=25000 (70-80M/sec). I didn't even mess with the CLASSES setting, or the GridSize, so I will have to try that too.

[petrw1]

13 days this time.

[petrw1]

11 more days.

[petrw1]

13 days to complete 330M range.

All partial ranges to go....on the home stretch.

[petrw1]

900M now complete to 64 bits thanks to Linded/Lindee et al.

[NBtarheel_33]

Originally Posted by The Head Bean Counter

900M now complete to 65 bits thanks to Linded/Lindee et al.

That's better.

[petrw1]

35 days to complete the 400M range.

500M Range is over 75% done thanks primarily to Chalsall.
600M Range is about 65% done.
700M Range is long done by monst
800M Range is 60% done.
900M Range is done recently thanks to Lindee/ed

I give each of the 3 remaining ranges about 2 weeks each...aiming still for a end of January completion....16 months total for 64 Bits from 100-999M.

How much less than double (32 months) will the next bit level (65-66 Bits) take due to Hardware/Software advancements?
Or will it take longer because as people get GPUs they move their efforts away from LMH and to deeper bit levels in the LL Range (i.e. GPU to 72 sub-project)?

P.S. I jumped the gun on submitting this post....my intent was to collect stats and post when 400M was actually done...but we are only a couple hours away.

[chalsall]

Originally Posted by petrw1

I give each of the 3 remaining ranges about 2 weeks each...aiming still for a end of January completion....16 months total for 64 Bits from 100-999M.

Sniff... My beautiful art work I spent years creating will be wiped out in a matter of weeks... Like sand castles in a rising tide...

It's been fun, but I'm going to retire from LMH, and move my cluster over to DC work.