GPU to 72 status... - Page 273

chalsall · 2014-06-04, 14:04

Quote:

Originally Posted by ET_

Can I assue that ALL TF done with mfaktc was completed? Or there may be cases where TF has been interrupted right after finding a factor, without completing the bit level?

The latter is what most people do.

James Heinrich · 2014-06-04, 14:42

Quote:

Originally Posted by chalsall

The latter is what most people do.

Arguably more efficient (in terms of exponents cleared), and yet not the mfaktc default (I believe it was the mfakto default, last time I checked).

Bdot · 2014-06-05, 16:39

Quote:

Originally Posted by James Heinrich

Arguably more efficient (in terms of exponents cleared), and yet not the mfaktc default (I believe it was the mfakto default, last time I checked).

That's correct.

Regarding the performance measurement to be built into mfaktX, Oliver and I were thinking about this:

Implement a test routine that runs a particular kernel with a fix exponent (Oliver's favorite 65M one) and a certain factor candidate size for a fix amount of time (e.g. 3 seconds) and see how far it came.

This routine can be used to loop over kernels and FC size. As we do have a performance dependency to the exponent size as well, I'd allow that to be variable as well.

I will extend the --perftest option of mfakto to do the test as James had suggested (~two pages earlier). Similar to the sieving performance test that I already have, it will read exponents and FC sizes to test from the ini file. I need to figure out how to properly combine these loads of generated data, though.

Example how it is today:

Code:

inifile:
TestSievePrimes=19890,30738,47503,73411,99999,113449,122222,175323,270944,418716,647083,1075766
TestGPUSieveSizes=4,7,12,16,20,36,48,96,101,102,104,120,121,123,124,125,126,127,128

mfakto --perftest:
...
4. GPU sieve, 10 iterations each

 gpusieve_init: 8.685000 ms (CPU work)
 gpusieve_init_exponent: 1.841100 ms (CalcModularInverses)
 gpusieve_init_class: 0.355800 ms (CalcBitToClear)
 gpusieve: 4.861600 ms (SegSieve)
  tf: 39.414000 ms = 3352.123002 M/s (raw rate, cl_barrett15_69_gs)

 GPU sieve raw rate (input rate M/s)
SievePrimes:   19890   30738   47503   73411   99999  113449  122222  175323  270944  418716  647083 1075766
GPUSieveSize
     4 MBit   14108.0 11370.3  6756.6  6382.9  4859.2  4398.7  4139.5  3137.5  2187.4  1434.3   913.2   457.7
     7 MBit   16732.6 15099.3 12317.8  9777.7  7759.9  7098.8  6768.9  5119.1  3674.5  2427.2  1519.5   763.7
    12 MBit   20467.2 18618.0 15609.4 12286.1  9884.3  9147.8  8664.4  6530.8  4511.1  3056.0  1970.5  1052.2
    16 MBit   23049.3 20749.9 17868.8 14529.6 11471.9 10468.8  9777.1  7412.3  5195.1  3472.1  2262.1  1216.6
    20 MBit   24060.3 22617.2 18829.5 15106.1 12188.3 11205.6 10609.4  8223.9  5822.3  3908.1  2536.7  1329.9
    36 MBit   28627.2 26998.6 23338.9 18942.3 15389.7 14035.3 13213.7 10023.2  6860.0  4540.2  2698.0  1165.2
    48 MBit   30928.9 29083.8 25581.4 21601.5 17754.1 15811.9 13833.5 10492.9  7241.4  4719.9  2582.8  1199.7
    96 MBit   34094.0 32564.6 29625.9 23068.5 18816.0 16736.3 15876.7 11714.5  7470.0  4655.9  1841.1   309.5
   101 MBit   34552.9 32000.9 30054.5 23895.0 18857.0 17027.3 16032.5 11663.9  7357.6  4761.7  1646.6   308.4
   102 MBit   33476.8 32197.7 29504.7 23519.2 18718.2 16951.5 16002.2 11640.5  7404.7  4485.3  1898.6   269.2
   104 MBit   34359.6 33008.2 29945.4 23970.6 19012.3 17537.0 16205.4 11688.5  7431.7  4642.9  1935.0   288.1
   120 MBit   34437.5 33495.4 31195.2 24505.2 19296.7 17514.3 16294.4 11727.8  7335.6  4586.1  1262.3   231.2
   121 MBit   35271.0 33778.4 30574.0 24273.9 19359.2 17400.3 16475.6 11664.5  7448.3  4485.8  1306.4   223.6
   123 MBit   35108.7 33998.7 30944.4 24445.0 19096.9 17458.2 16578.0 11906.2  7494.3  4651.0  1175.9   226.3
   124 MBit   35551.9 33176.6 31075.1 24321.1 19549.2 17519.3 16553.2 11973.6  7502.1  4430.9  1719.0   216.8
   125 MBit   34252.3 33766.1 31011.4 24501.4 19032.9 17361.4 16426.7 11899.2  7382.4  4451.5  1240.7   239.7
   126 MBit   35792.7 34077.2 31133.6 24821.3 19502.1 17596.9 16568.5 12003.7  7362.5  4532.9   948.2   244.3
   127 MBit   34532.6 32777.0 30255.3 24371.9 19279.2 17731.3 16766.9 11837.3  7423.6  4535.6  1480.0   232.5
   128 MBit   33707.9 33267.3 31661.8 24512.2 19356.4 17596.0 16445.3 11965.3  7430.9  4559.3   923.9   245.5

Best GPUSieveSize for
SievePrimes:   19766   31030   47414   74038   99894  113206  122422  175670  270902  419382  647734 1075766
at MiB:          126     126     128     126     124     127     127     126     124     101      36      20
max M/s:     35792.7 34077.2 31661.8 24821.3 19549.2 17731.3 16766.9 12003.7  7502.1  4761.7  2698.0  1329.9
Survivors:    22.18%  21.32%  20.57%  19.84%  19.38%  19.19%  19.08%  18.56%  17.98%  17.44%  16.94%  16.41%
removal rate
  average:   27855.6 26812.1 25147.4 19895.9 15760.7 14328.4 13568.5  9776.1  6153.0  3931.2  2240.9  1111.7
  incremental:   n/a  6083.2  3326.5   840.4   427.5   357.1   356.7   219.0   115.1    70.6    31.0    14.0

Another, slightly related idea that I will implement: Allow for dynamic reconfiguration of mfakto during runtime: (GPU)SievePrimes, SieveSize, maybe even select the kernel to be used. This way it would be easier to find the optimal values for a certain task.

chalsall · 2014-06-10, 15:02

This has been discussed before, but just to reiterate...

A GPU'er emailed me last night concerned that the assignments he was getting from GPU72 weren't "owned" by "gpufactor" ("GPU72") according to Primenet, but instead "owned by someone else".

So everyone knows, the only candidates assigned by GPU72 for TF'ing or P-1'ing will be owned by either "gpufactor" ("GPU72") or "wabbit" ("For Research"). GPU72 is not doing any "poaching" nor parallel processing -- the latter account is simply being used to collect recycled candidates from Primenet as they expire. "wabbit's" holdings will appear as either LL or P-1 assignments on Primenet, instead of TF.

Mark Rose · 2014-06-17, 15:24

Quote:

Originally Posted by NickOfTime

Just a few gpu's :-) 2 AMD 290x, amd 260x, gtx 690, gtx 660ti, gtx 560ti

And yesterday I put a spare GTX 520 into service on my work desktop for 28.8 GHz-d/day. I need better cards lol

Mark Rose · 2014-06-17, 16:13

Quote:

Originally Posted by Xyzzy

If anyone is successful in getting CUDA running on Ubuntu 12.04.4 or 14.04 please let us know.

We had it working via the package manager on 12.04.1 but it is horribly broken now.

On a clean 14.04 install I've gotten it to work by installing the nvidia-331 and nvidia-cuda-dev and nvidia-cuda-toolkit packages. That should give you CUDA 6 in the driver and CUDA 5.5 for development. Compiling mfaktc then works.

I've gotten bumblebee to work as well, but I had to modify /etc/bumblebee/bumblebee.conf by making sure the following is set (when using nvidia-331):

Code:

## Section with nvidia driver specific options, only parsed if Driver=nvidia-331
[driver-nvidia-331]
# Module name to load, defaults to Driver if empty or unset
KernelDriver=nvidia-331
PMMethod=none
# colon-separated path to the nvidia libraries
LibraryPath=/usr/lib/nvidia-331:/usr/lib32/nvidia-current
# comma-separated path of the directory containing nvidia_drv.so and the
# default Xorg modules path
XorgModulePath=/usr/lib/nvidia-331/xorg,/usr/lib/xorg/modules
XorgConfFile=/etc/bumblebee/xorg.conf.nvidia

That allows me to use the on-board Intel video for doing work on the machine, and letting the Nvidia card factor away. Just run `cd mfaktc ; optirun ./mfaktc.exe -d 0`. I haven't figured out how to get bumblebee to work with 2+ nvidia cards.

James Heinrich · 2014-06-18, 18:02

Random minor observation: When you request assignments, it presents you a list of assignments and tells you to "cut and paste", but since the assignments aren't in an editable HTML element you can copy, but not actually cut... Perhaps it should say "copy and paste" instead?

Pedantic, but what do you expect from programmers?

chalsall · 2014-06-18, 18:28

Quote:

Originally Posted by James Heinrich

Pedantic, but what do you expect from programmers?

ROFL...

Good point. Language can be so very important...

Unfortunately, changing that single word would involve a bit of work since it is in the "translation database". I would have to add a new (more correct) phrase, update four Perl scripts, and then the translators would have to submit translations. Thus, I'm going to leave it as it is...

If it helps at all, your reported bug about the dates being rendered badly on the individual overall workers' productivity report was fixed a few days ago... (I never, ever, imagined this sub-project would exist for so long!)

James Heinrich · 2014-06-18, 18:31

Quote:

Originally Posted by chalsall

your reported bug about the dates being rendered badly on the individual overall workers' productivity report was fixed a few days ago...

Yay! Thanks

TheMawn · 2014-06-19, 00:42

Quote:

Originally Posted by chalsall

(I never, ever, imagined this sub-project would exist for so long!)

Whaaaaaat? Involving GPU's is a stroke of genius, and so is putting together a system to make it user friendly.

chalsall · 2014-06-19, 01:33

Quote:

Originally Posted by TheMawn

Whaaaaaat? Involving GPU's is a stroke of genius, and so is putting together a system to make it user friendly.

Very kind words. Thank you.

But "The Judger" and "Bdot" were really the ones who brought GPU'ing into this space. And "Mr. P-1" used to coordinate this kind of work manually.

I simply came in afterwards, and stood on the shoulders of giants....

2014-06-10, 15:02	#2996
chalsall If I May "Chris Halsall" Sep 2002 Barbados 2·67·73 Posts	"Ownership" of GPU72 assigned candidates... This has been discussed before, but just to reiterate... A GPU'er emailed me last night concerned that the assignments he was getting from GPU72 weren't "owned" by "gpufactor" ("GPU72") according to Primenet, but instead "owned by someone else". So everyone knows, the only candidates assigned by GPU72 for TF'ing or P-1'ing will be owned by either "gpufactor" ("GPU72") or "wabbit" ("For Research"). GPU72 is not doing any "poaching" nor parallel processing -- the latter account is simply being used to collect recycled candidates from Primenet as they expire. "wabbit's" holdings will appear as either LL or P-1 assignments on Primenet, instead of TF. Last fiddled with by chalsall on 2014-06-10 at 15:04 Reason: Added last sentence.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Status	Primeinator	Operation Billion Digits	5	2011-12-06 02:35
62 bit status	1997rj7	Lone Mersenne Hunters	27	2008-09-29 13:52
OBD Status	Uncwilly	Operation Billion Digits	22	2005-10-25 14:05
1-2M LLR status	paulunderwood	3*2^n-1 Search	2	2005-03-13 17:03
Status of 26.0M - 26.5M	1997rj7	Lone Mersenne Hunters	25	2004-06-18 16:46

2014-06-18, 18:02	#2999
James Heinrich "James Heinrich" May 2004 ex-Northern Ontario 3427₁₀ Posts	Random minor observation: When you request assignments, it presents you a list of assignments and tells you to "cut and paste", but since the assignments aren't in an editable HTML element you can copy, but not actually cut... Perhaps it should say "copy and paste" instead? Pedantic, but what do you expect from programmers?