mfaktc: a CUDA program for Mersenne prefactoring - Page 32

davieddy · 2010-08-04, 11:44

Quote:

Originally Posted by TheJudger

gives most bank for the buck

Point of Anglo-American: the expression is
"bang for buck". Interpret the literal meaning as you like

David

cheesehead · 2010-08-04, 19:32

Quote:

Originally Posted by davieddy

Point of Anglo-American: the expression is
"bang for buck". Interpret the literal meaning as you like

David

From http://www.randomhouse.com/wotd/inde...?date=19971219 :

Quote:

The phrase, which means 'value for one's money', was originally a political one. Its first use was quite literal: With bang referring to 'firepower' or 'weaponry', it really did mean 'bombs for one's money'. The alliteration of bang and buck helps to make the phrase memorable.

The earliest confirmed mention of bang for the buck is found in 1968 in the first edition of William Safire's New Language of Politics. Mr. Safire claims that the phrase was coined in 1954 by Charles E. Wilson, the Secretary of Defense, in reference to the "massive retaliation" policy of John Foster Dulles.

While bang has been used in sexual senses since the seventeenth century, it is unrelated to our phrase.

TheJudger · 2010-08-13, 20:50

Hi David, hi cheesehead,

just a stupid typo...

Since 460GTX are cheaper now they give the most bang for the buck in mfaktc. 768MiB variants are OK, mfaktc doesn't need much memory/memory bandwidth on GPU.

Anyway, here are the performance measurements of a stock GTX 480:

Code:

kernel | M66362159, 2^64 to 2^67 | M3321932839, 2^50 to 2^71
-------+-------------------------+--------------------------
71bit  | 127.0M/s                | 98.7M/s       mfaktc-0.10
75bit  | 227.4M/s         179.1% | 177.8M/s           180.1%
95bit  | 192.2M/s         151.3% | 149.8M/s           151.8%

Oliver

frmky · 2010-08-26, 07:47

I thought I'd give this another try. I'm using a 2.4GHz Core 2 Quad processor with a GTX 480 card. Keeping the GPU busy requires 3 instances of mfaktc. Otherwise, SievePrimes just drops to 5000. Each instance is running at about 58.4 M/s on a >100M digit number, so a total of about 175M/s. This seems consistent with your chart above. This corresponds to taking three >100 digit candidates to 78-bits in 4 days or 79-bits in 8 days.

TheJudger · 2010-08-26, 16:49

Hi frmky,

yep, GTX 480 is a beast.

Did you try to enable MORE_CLASSES in params.h? This lets the sieve running more efficient at the cost of more initializations. So this is usefull for "long runs" where each class takes > 60 seconds without MORE_CLASSES enabled. For a 100M digit to 2^79 it is a clear advantage to enable it.

Hint: currently mfaktc 0.11 is out to some people here in the forum for testing. If everything is right I'll release 0.11 within one week.

You'll like mfaktc 0.11 because the sieve is ~20% faster (measured on i7), but I think you'll still need 3 instances on your machine to keep your GTX 480 busy.
Note: a ~20% faster sieve doesn't mean a 20% improvement on overall performance in general!

Oliver

Lorenzo · 2010-08-27, 11:40

Why there is no Win' 32bit port's? Somebody can make it port for last version?

Sorry for my English

TheJudger · 2010-08-27, 16:55

Hi Lorenzo,

hopefully no port is needed, we need just a win32 build!

64bit is prefered but not necessary simply because they are faster (my sieve runs faster on 64bit).
Actually I can't generate Windows binaries, sorry!
But perhaps someone here in the forum can generate them.

Oliver

Lorenzo · 2010-08-30, 18:17

Quote:

Originally Posted by TheJudger

Hi Lorenzo,

hopefully no port is needed, we need just a win32 build!

64bit is prefered but not necessary simply because they are faster (my sieve runs faster on 64bit).
Actually I can't generate Windows binaries, sorry!
But perhaps someone here in the forum can generate them.

Oliver

Hi Oliver,

I'm trying to install 64-bit system:
1. Windows XP Professional x64 Edition
2. nVidia Forceware v258.96 International (WinXP x64) WHQL
3. CUDA ToolKit 3.1 (Win64)
4. Video Card nVidia Geforce 8500 GT (256 Mb DDR2)
5. mfaktc-0.09-win64-eoc

But when i run mfaktc-0.09-win64-eoc:

Code:

mfaktc v0.09-Win

Compiletime Options
  THREADS_PER_GRID    983040
  THREADS_PER_BLOCK   256
  SIEVE_SIZE_LIMIT    32kiB
  SIEVE_SIZE          230945bits
  VERBOSE_TIMING      disabled
  MORE_CLASSES        disabled

Runtime Options
WARNING: Cannot read SievePrimes from mfaktc.ini, using default value (25000)
  SievePrimes         25000
WARNING: Cannot read SievePrimesAdjust from mfaktc.ini, using default value (0)
  SievePrimesAdjust   0
WARNING: Cannot read NumStreams from mfaktc.ini, using default value (3)
  NumStreams          3
WARNING: can't read WorkFile from mfaktc.ini, using default (worktodo.ini)
  WorkFile            worktodo.ini
WARNING: Cannot read Checkpoints from mfaktc.ini, enabled by default
  Checkpoints         enabled

CUDA device info
  name:                      GeForce 8500 GT
  compute capabilities:      1.1
  maximum threads per block: 512
  number of multiprocessors: 2 (16 shader cores)
  clock rate:                1400MHz

cudaStreamCreate() failed

What i doing wrong?

TheJudger · 2010-08-30, 18:40

Hi Lorenzo,

first of all I recommend that you download the source tarball of mfaktc 0.09.
http://www.mersenneforum.org/showpos...&postcount=295
Download and put the mfaktc.ini into the directory of your mfaktc executable.
This will remove the warnings and gives you some more control over mfaktc.

But I think this won't solve you issue. You could try to set NumStreams to 1 on mfaktc.ini than.

Aillas reported the same(?) issue on Linux with his Nvidia quadro 140M. On his system mfaktc 0.10 works.
I've no clue what happens there, the code which does stream creation is virtually untouched between those two versions, the difference is that mfaktc 0.10 prints the stream number when the stream creation failed.

Oliver

P.S. is this an overclocked GPU? According the wikipedia it should run at 900MHz.

Aillas · 2010-08-31, 08:56

Quote:

Originally Posted by TheJudger

Aillas reported the same(?) issue on Linux with his Nvidia quadro 140M. On his system mfaktc 0.10 works.
I've no clue what happens there, the code which does stream creation is virtually untouched between those two versions, the difference is that mfaktc 0.10 prints the stream number when the stream creation failed.

Hi,

bad news. I can't make it run since one week. In a bunch of Linux update, I also update to the latest nvidia driver. BIG MISTAKE. The computer was unusable.
I remove all nvidia driver and re install manually the 254.35 (I think), but now, mfakt 0.10 doesn't work anymore.

If you want, I can try your 0.11 version;
I think I will upgrade my linux to a 64 bit version too.

TheJudger · 2010-08-31, 12:21

Hi Aillas,

mfaktc 0.11 won't fix this, again the stream creation code is unchanged to the previous version. And don't spent time for searching version 0.11, it is not released yet. This should happen in the next few days.

Sorry for asking again: Can you try the nvidia SDK code samples? http://developer.nvidia.com/object/cuda_download.html

Oliver

2010-08-13, 20:50	#344
TheJudger "Oliver" Mar 2005 Germany 1111₁₀ Posts	Hi David, hi cheesehead, just a stupid typo... Since 460GTX are cheaper now they give the most bang for the buck in mfaktc. 768MiB variants are OK, mfaktc doesn't need much memory/memory bandwidth on GPU. Anyway, here are the performance measurements of a stock GTX 480: Code: kernel \| M66362159, 2^64 to 2^67 \| M3321932839, 2^50 to 2^71 -------+-------------------------+-------------------------- 71bit \| 127.0M/s \| 98.7M/s mfaktc-0.10 75bit \| 227.4M/s 179.1% \| 177.8M/s 180.1% 95bit \| 192.2M/s 151.3% \| 149.8M/s 151.8% Oliver

2010-08-26, 16:49	#346
TheJudger "Oliver" Mar 2005 Germany 1111₁₀ Posts	Hi frmky, yep, GTX 480 is a beast. Did you try to enable MORE_CLASSES in params.h? This lets the sieve running more efficient at the cost of more initializations. So this is usefull for "long runs" where each class takes > 60 seconds without MORE_CLASSES enabled. For a 100M digit to 2^79 it is a clear advantage to enable it. Hint: currently mfaktc 0.11 is out to some people here in the forum for testing. If everything is right I'll release 0.11 within one week. You'll like mfaktc 0.11 because the sieve is ~20% faster (measured on i7), but I think you'll still need 3 instances on your machine to keep your GTX 480 busy. Note: a ~20% faster sieve doesn't mean a 20% improvement on overall performance in general! Oliver Last fiddled with by TheJudger on 2010-08-26 at 16:50

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
mfakto: an OpenCL program for Mersenne prefactoring	Bdot	GPU Computing	1676	2021-06-30 21:23
The P-1 factoring CUDA program	firejuggler	GPU Computing	753	2020-12-12 18:07
gr-mfaktc: a CUDA program for generalized repunits prefactoring	MrRepunit	GPU Computing	32	2020-11-11 19:56
mfaktc 0.21 - CUDA runtime wrong	keisentraut	Software	2	2020-08-18 07:03
World's second-dumbest CUDA program	fivemack	Programming	112	2015-02-12 22:51

2010-08-26, 07:47	#345
frmky Jul 2003 So Cal 100001001010₂ Posts	I thought I'd give this another try. I'm using a 2.4GHz Core 2 Quad processor with a GTX 480 card. Keeping the GPU busy requires 3 instances of mfaktc. Otherwise, SievePrimes just drops to 5000. Each instance is running at about 58.4 M/s on a >100M digit number, so a total of about 175M/s. This seems consistent with your chart above. This corresponds to taking three >100 digit candidates to 78-bits in 4 days or 79-bits in 8 days.

2010-08-27, 11:40	#347
Lorenzo Aug 2010 Republic of Belarus 2·89 Posts	Why there is no Win' 32bit port's? Somebody can make it port for last version? Sorry for my English

2010-08-27, 16:55	#348
TheJudger "Oliver" Mar 2005 Germany 11×101 Posts	Hi Lorenzo, hopefully no port is needed, we need just a win32 build! 64bit is prefered but not necessary simply because they are faster (my sieve runs faster on 64bit). Actually I can't generate Windows binaries, sorry! But perhaps someone here in the forum can generate them. Oliver

2010-08-30, 18:40	#350
TheJudger "Oliver" Mar 2005 Germany 11·101 Posts	Hi Lorenzo, first of all I recommend that you download the source tarball of mfaktc 0.09. http://www.mersenneforum.org/showpos...&postcount=295 Download and put the mfaktc.ini into the directory of your mfaktc executable. This will remove the warnings and gives you some more control over mfaktc. But I think this won't solve you issue. You could try to set NumStreams to 1 on mfaktc.ini than. Aillas reported the same(?) issue on Linux with his Nvidia quadro 140M. On his system mfaktc 0.10 works. I've no clue what happens there, the code which does stream creation is virtually untouched between those two versions, the difference is that mfaktc 0.10 prints the stream number when the stream creation failed. Oliver P.S. is this an overclocked GPU? According the wikipedia it should run at 900MHz.

2010-08-31, 12:21	#352
TheJudger "Oliver" Mar 2005 Germany 11×101 Posts	Hi Aillas, mfaktc 0.11 won't fix this, again the stream creation code is unchanged to the previous version. And don't spent time for searching version 0.11, it is not released yet. This should happen in the next few days. Sorry for asking again: Can you try the nvidia SDK code samples? http://developer.nvidia.com/object/cuda_download.html Oliver