mfaktc: a CUDA program for Mersenne prefactoring

[Xyzzy]

To preempt another question - how does affinity work in windows? I had trouble finding suitable help on the topic.

We have never dealt with a quad with HT, so this works for us.

Someday we will get an i7, although it appears that the CPU is far from being the limiter with the GPU cards we have now. We doubt we will be buying a GTX 590 anytime soon.

[Christenson]

I installed the GTX440 tonight; xubuntu isn't doing well with the driver install, and I'm out of energy to work on it tonight. Do you think we can get it to coexist with the on-board VGA graphics which report themselves as ATI Radeon?
Also, lost power over at UVA so won't have any computing from there until tomorrow evening.

[diep]

Originally Posted by Christenson

I installed the GTX440 tonight; xubuntu isn't doing well with the driver install, and I'm out of energy to work on it tonight. Do you think we can get it to coexist with the on-board VGA graphics which report themselves as ATI Radeon?
Also, lost power over at UVA so won't have any computing from there until tomorrow evening.

hi,

for gpgpu we must distinguish 2 diferent things:

a) steer the device as being gpgpu capable while also serving as videocard
b) a device exclusively used for gpgpu

Getting A to work is always possible with mixed manufacturers.

With nvidia i don't know, but with AMD i have a nvidia chipset in the 4 socket box. that means that B is not possible with this videocard.

So the 6970 here is both steering the videocard as well as doing gpgpu. That's far from ideal of course.

Maybe putting in another card, an ATI one, it would be possible for me.
If it's at full production i'll try that - maybe . Maybe not.

It's possible that nvidia has a similar problem and that you need an oldie nvidia card as first device to steer video for a nvidia cuda capable device.

As for the box i have, which i built for under 1000 euro - yes it's an old one it has quad core 8356 AMD's which were $50 a piece on ebay - yes it's faster for all sort of math codes than any Nehalem system as the integer multiplication on AMD cpu's is 2.25 cycles throughput versus intels are 3.75 cycles); the bandwidth to the videocard is crap on such oldie system.

The bandwidth here is something like 2.1GB/s to GPU and 2.2GB/s back.
So that's 2.3GB in total over the pci-e 16x here.

Of course in mainboards from back then in fact pci-e 16x was limited to 8x everywhere.

i7's are really having a superior bandwidth to videocard it seems - didn't test Phenom's that have DDR3 ram, might also be ok.

Note that mfackt needs that bandwidth. For mfockt, the AMD version i'm writing, this is not needed as it's supposed to run on the GPU itself in the long run, the bigger worry there is achieving a fast speed.

You should really check this bandwidth.

[Christenson]

Hi Oliver

After some fooling around, I got the GTX440 (all the power supply wants to support) going on my Xubuntu 10.10 with AMD Phenom II x6 CPU in dual-headed mode; ATI on-board chipset has the primary console and there is a secondary one available on the GTX. Had to do some fooling around with xorg.conf to make this go; the nvidia apps are a bit greedy and don't leave ATI Radeon refernces in xorg.conf.

mfaktc 0.16p1 is now running a detailed test, where it should reproduce my most recent P-1 factoring result.

My question is: How do I check on the GPU loading I am achieving in this setup? How about it's temperature?

[Christenson]

The GTX440 looks *reasonably* hot....
4 hours to do a typical M50M TF to 2^72, that should be worth about 16GHz days.
about 72GHz days/day...I'm going for 2^74, so it will begin doing real work tomorrow some time.

Hey, guys, how would I set mfaktc to TF'ing M1277, which has been TF'ed to 2^60 only? mfaktc complains about a lower limit of M(10^5). Why?
I was thinking about more TF on some of these relatively small mersenne numbers, as these GPUs just might be able to find factors in a reasonable time....

[xilman]

Originally Posted by Christenson

Hey, guys, how would I set mfaktc to TF'ing M1277, which has been TF'ed to 2^60 only? mfaktc complains about a lower limit of M(10^5). Why?
I was thinking about more TF on some of these relatively small mersenne numbers, as these GPUs just might be able to find factors in a reasonable time....

There's no point. Such small numbers have already been attempted by ECM to at least 100 bits.

Paul

[TheJudger]

Hi Christenson,

Originally Posted by Christenson

My question is: How do I check on the GPU loading I am achieving in this setup? How about it's temperature?

you can try 'nvidia-smi -a' on the command line.

Originally Posted by Christenson

Hey, guys, how would I set mfaktc to TF'ing M1277, which has been TF'ed to 2^60 only? mfaktc complains about a lower limit of M(10^5). Why?

As Paul allready said this doesn't make much sense.
Another reason is that I haven't checked if mfaktc produces correct results on that small exponents.
The sieve needs to be adjusted, too. Otherwise the sieving process would remove real factor candidates on that small numbers.

Oliver

[Christenson]

How could I calculate that M1277 has been ECM'ed to an effective TF level of 100 bits? How would I prove that if M1277 had a 62 bit factor, it certainly would have been found by one of the ECM curves?

As for mfaktc, one core is clearly saturating the GTX440(a second instance halves the performance of the first instance)...those sleeps (or buffers) would be appreciated so P95 could also make some progress on that core....there's lots of memory available on both the CPU and GPU for the job.

Benchmark:
Galaxy GEForce GTX440, nvidia driver 270.41.06, mfaktc 0.16p1 bound to CPU#6.
Card at 60C, 10% of memory on GPU used, room is cool -- under 20C, fan speed is 50%, while working on the exponent below to 73 bits. There's a known (due to my second-most recent successful P-1 factoring effort) factor below 2^74, which I'm waiting for the card to find.
no factor for M53953421 from 2^71 to 2^72 [mfaktc 0.16p1 barrett79_mul32]
tf(): total time spent: 4h 16m 16.872s

[xilman]

Originally Posted by Christenson

How could I calculate that M1277 has been ECM'ed to an effective TF level of 100 bits? How would I prove that if M1277 had a 62 bit factor, it certainly would have been found by one of the ECM curves?

Take a look at http://mersenne.org/report_ECM/ where you will find that at least one test to 60-digits has been performed on M1277. 60 digits is over 200 bits, so there are surely no factors under 100 bits still to be found.

(Note for pedants: there is indeed a small possibility that a factor under 100 bits has not been found by the ECM work performed so far. It is so small that I really, really don't recommend looking for it by TF.)


Paul

[Christenson]

Originally Posted by xilman

Take a look at http://mersenne.org/report_ECM/ where you will find that at least one test to 60-digits has been performed on M1277. 60 digits is over 200 bits, so there are surely no factors under 100 bits still to be found.

(Note for pedants: there is indeed a small possibility that a factor under 100 bits has not been found by the ECM work performed so far. It is so small that I really, really don't recommend looking for it by TF.)


Paul

Looked at it -- don't understand the theory very well. You are saying that, to very high probability (parts in 10^6 or 7 on the complement, or better?) the ECM which has been done would have found a factor under 100 bits, right?
What B2 would be used in such curves? How do such curves get run? P95 manual assignments? and is it effective to ask about running curves to 65 digits?

[Christenson]

Next dumb question: Is it worth downloading the mprime source and working on automating the assignment process, fetching TF work from GIMPS?