Let GPU72 decide

[chalsall]

Originally Posted by Mark Rose

And now? I dumped another 4THz-days in today haha

Sorry -- missed this question... It is impossible for me to tell exactly, but based on the completion rate of ~34 candidates a day in the DC Cat 4 range, and the fact we have approximately 2100 candidates ready to be assigned in that range, we now have a buffer of approximately two months.

Originally Posted by Mark Rose

Also, is this the first day we found more DCTF factors than LLTF factors?

Yup.

Thanks again for your efforts, and for helping me understand EC2 better.

[chalsall]

Originally Posted by Mark Rose

You'll want to use new hardware that has more CPU instructions exposed to the virtual machine (i.e. C3.large, Xeon E5-2680 v2, 2 cores, about $0.033/hr spot).

Just out of interest, I spun up a C3.xlarge to find out the answer to the question about estimated time and cost (empirical is always good):

[ec2-user@ip-10-45-9-204 prime]$ date ; ./mprime -d
Wed Jul  2 17:27:00 UTC 2014
[Main thread Jul 2 17:27] Mersenne number primality test program version 28.5
[Main thread Jul 2 17:27] Optimizing for CPU architecture: Unknown Intel, L2 cache size: 256 KB, L3 cache size: 25 MB
[Main thread Jul 2 17:27] Logical CPUs 1,3 form one physical CPU.
[Main thread Jul 2 17:27] Logical CPUs 2,4 form one physical CPU.
[Main thread Jul 2 17:27] Using AffinityScramble2 settings to set affinity mask.
[Main thread Jul 2 17:27] Starting worker.
[Work thread Jul 2 17:27] Worker starting
[Work thread Jul 2 17:27] Setting affinity to run worker on logical CPU #1
[Work thread Jul 2 17:27] Setting affinity to run helper thread 1 on logical CPU #2
[Work thread Jul 2 17:27] Starting primality test of M332226749 using AVX FFT length 19200K, Pass1=768, Pass2=25K, 2 threads
[Work thread Jul 2 17:28] Iteration: 1000 / 332226749 [0.00%], ms/iter: 87.206, ETA: 335d 07:48
[Work thread Jul 2 17:30] Iteration: 2000 / 332226749 [0.00%], ms/iter: 87.154, ETA: 335d 02:59
[Work thread Jul 2 17:31] Iteration: 3000 / 332226749 [0.00%], ms/iter: 87.197, ETA: 335d 06:57
[Work thread Jul 2 17:32] Iteration: 4000 / 332226749 [0.00%], ms/iter: 87.102, ETA: 334d 22:08
[Work thread Jul 2 17:34] Iteration: 5000 / 332226749 [0.00%], ms/iter: 87.020, ETA: 334d 14:32
[Work thread Jul 2 17:35] Iteration: 6000 / 332226749 [0.00%], ms/iter: 86.999, ETA: 334d 12:34
[Work thread Jul 2 17:37] Iteration: 7000 / 332226749 [0.00%], ms/iter: 86.986, ETA: 334d 11:18
[Work thread Jul 2 17:38] Iteration: 8000 / 332226749 [0.00%], ms/iter: 86.979, ETA: 334d 10:41
[Work thread Jul 2 17:40] Iteration: 9000 / 332226749 [0.00%], ms/iter: 86.902, ETA: 334d 03:32
[Work thread Jul 2 17:41] Iteration: 10000 / 332226749 [0.00%], ms/iter: 86.898, ETA: 334d 03:09
[Work thread Jul 2 17:43] Iteration: 11000 / 332226749 [0.00%], ms/iter: 86.897, ETA: 334d 03:01
[Work thread Jul 2 17:44] Iteration: 12000 / 332226749 [0.00%], ms/iter: 86.999, ETA: 334d 12:25
[Work thread Jul 2 17:45] Iteration: 13000 / 332226749 [0.00%], ms/iter: 86.935, ETA: 334d 06:29
[Work thread Jul 2 17:47] Iteration: 14000 / 332226749 [0.00%], ms/iter: 87.041, ETA: 334d 16:13
[Work thread Jul 2 17:48] Iteration: 15000 / 332226749 [0.00%], ms/iter: 87.038, ETA: 334d 15:56
[Work thread Jul 2 17:50] Iteration: 16000 / 332226749 [0.00%], ms/iter: 87.069, ETA: 334d 18:45
[Work thread Jul 2 17:51] Iteration: 17000 / 332226749 [0.00%], ms/iter: 87.023, ETA: 334d 14:31
[Work thread Jul 2 17:53] Iteration: 18000 / 332226749 [0.00%], ms/iter: 86.932, ETA: 334d 06:04
[Work thread Jul 2 17:54] Iteration: 19000 / 332226749 [0.00%], ms/iter: 86.944, ETA: 334d 07:12

So, assuming the spot price stayed at around 3.25 cents per hour, it would cost you ~ $260 USD to complete a single LL!

Originally Posted by Mark Rose

Spot instances might last several days, but you're almost guaranteed to be interrupted at some point. When they die, you'll have to manually recover the work (you can do a resumable spot request, but it will launch a new instance with a new filesystem). So there is that to deal with.

Indeed.

This is a non-trivial truth which takes some work to get around automatically.

[TheMawn]

3.25 cents per hour sounds fairly low but that's actually 250W worth, roughly speaking. And that seems to be for one core.

If you built a four-core machine, after a year, you'd have been better off with a computer that draws less than 250W and cost less than $750.

On the other hand, the convenience is really quite nice.

[Mark Rose]

Originally Posted by TheMawn

3.25 cents per hour sounds fairly low but that's actually 250W worth, roughly speaking. And that seems to be for one core.

If you built a four-core machine, after a year, you'd have been better off with a computer that draws less than 250W and cost less than $750.

On the other hand, the convenience is really quite nice.

My electricity costs about $0.15/kwh. I can build a Haswell box that draws 230 watts for four cores (the box I'm typing on). So that works out to $0.008625 per core-hour. That compares to $0.0165 per core-hour for EC2 spot (they're actually hyperthreads, but we'll assume we're getting full cores). So EC2 is twice the price per hour to run.

That Haswell system would cost about $150/core. To come out ahead at $0.007875/hr, I'd have to run for 19000 hours, or about 26 months.

I'd still have to compare the same exponents on both. I usually run Seventeen or Bust work on otherwise idle CPUs.

[chalsall]

Originally Posted by Mark Rose

My electricity costs about $0.15/kwh. I can build a Haswell box that draws 230 watts for four cores (the box I'm typing on). So that works out to $0.008625 per core-hour. That compares to $0.0165 per core-hour for EC2 spot (they're actually hyperthreads, but we'll assume we're getting full cores). So EC2 is twice the price per hour to run.

For me, I pay ~ $0.347 USD per kWh. So EC2 is actually less expensive for me to run (for CPU compute), not even taking into consideration the capital costs for the hardware.

I will later today be in front of the machine which hosts my GTX580 (it's in an office on the other side of the country; about 15 km away... ). I'll take a look at the power draw while running mfaktc, and see where the curves cross for GPU compute compared to a cg1.4xlarge. But I expect it will be similarly more economical for those of us who pay large amounts for our electricity....

[swl551]

as of tonight I am getting error

Sorry... No assignments available which match your criteria.

when trying to fetch assignments where I let GPU72 decide.

[Mini-Geek]

Originally Posted by swl551

as of tonight I am getting error



when trying to fetch assignments where I let GPU72 decide.

Same problem here (fetching via MISFIT), first failed for me at 2014-09-15 15:33 GMT. I was able to get assignments through the web site.

[chalsall]

Originally Posted by swl551

...when trying to fetch assignments where I let GPU72 decide.

Sorry guys. Profoundly SPE... Fixed.