mfaktc: a CUDA program for Mersenne prefactoring

[Karl M Johnson]

Oh, one more thing. Could increasing threads to 512 actually increase GPU load ? And reduce the CPU bottleneck ? Right now, to utilize the gpu close to 92% I need to launch 3 separate instances of mfakt.

[TheJudger]

Hi Karl,

I don't think so. Remind that we're talking about threads per block.
If there are enough resources available (for mfaktc the number of registers is the limiting factor) the GPU can run multiple blocks per multiprocessor.

Oliver

[TheJudger]

Hi Karl,

Originally Posted by TheJudger

Anyway I'll test it with the current version again, but I think I know allready the result...

I just tested THREADS_PER_BLOCK = 64, 128, 192, 256, 384 and 512.
64 is a little bit slower in some cases, all other values are within the measuring inaccuracy on my GTX 470.

Oliver

[Karl M Johnson]

So increasing threads per block to 512 dont give any speed advantage over 256 of them ?
Ok.
But the reason why single instance of mfakt cant utilize GPU up to 99% is that the whole code isnt executed on the GPU ?

[TheJudger]

Originally Posted by Karl M Johnson

But the reason why single instance of mfakt cant utilize GPU up to 99% is that the whole code isnt executed on the GPU ?

Yep, that's the most likely reason.
Another reason might be short running assignments (per class less than a few seconds) because between two classes the internal queues for working blocks run empty.
May I know
- your CPU
- your GPU
- your typical assignments (e.g. M83.xxx.xxx from 2^64 to 2^65)
- your OS
- 32 or 64 bit version of mfaktc
- are you using default settings in mfaktc.ini

Oliver

[Karl M Johnson]

Sure!
Q6600 @ 3.0 Ghz
GTX 480 @ 811/1622/3800
Typical assigments are a bit tricky. Anything PrimeNet assigns via manual request.
Windows 7 x64
Using 32 bit mfaktc.

SievePrimes=5000
SievePrimesAdjust=1
NumStreams=10
CPUStreams=5
WorkFile=worktodo.txt
Checkpoints=1
Stages=0
StopAfterFactor=0

Btw, when compiling mfaktc, is "--use_fast_math" flag used ?

[firejuggler]

Core 2 duo 8300 @ 2.83 Ghz
GTX 460 @ 715/1430/3600
Typical assigments are a bit tricky. Anything PrimeNet assigns via manual request. (from 65,66 to 69,70, for M80XXXXXX speed is around 80M/s at when the assignment is low ,104M/s at high. with 2 assignment, total speed is about 130-140M/s)
Windows XP
Using 32 bit version

[Ungelovende]

2x i7 920@2,66 + 1x i7 2600k@4430 all with one GTX 460@stock
XP64
Manual assignments edited to xx,71
Speed 130M~140M - one app pr GPU

Works like a dream!

[TheJudger]

Hi!

Karl: since you're running a 64bit windows I recommend to try the 64bit version of mfaktc, the sieve (CPU) runs more efficient in 64bit mode (~33% faster on Linux on my C2D and i7). You might to adjust the upper TF limit, too. Take a look at the README.txt.

Maybe your OCed GTX 480 is just a little bit too fast.
...did I really wrote that? Hardware can never be too fast.

The flag "--use_fast_math" isn't used... this only affects floating point numbers an mfaktc primary depends on integers.


Oliver

[Karl M Johnson]

The thing is, I'm currently on most proper drivers, which do not cause slowdown to cuda/opencl apps, intentionally or not.
The good thing, is by doing some curart shenanigans, I can run apps compiled under 26x.xx on my 258.96.
The bad thing, is, well, only 32 bit apps can be run like that.

Regarding the flag thing: heh, well, didnt know that.
I was NOT impressed when I read AVX of Sandy Bridge only works on FPs.
Integers dont get any benefits from it.
Well, at least SB's has a better perf perf clock ratio.

[firejuggler]

a small thing.. i tested mfacktc a bit further and it seems that there is a limit to the exponent you can use. If it is above 4294967295 (yeah, 11 time the billion digit project); it enter the famous loop :
got assignment: exp=4294967295 bit_min=1 bit_max=60
WARNING: exponent is not prime! Ignoring this assignment!

(4294967295=2^32-1)