I just make an SHA512 brute force programs that can do 5MH/s on my i5 3.5Ghz (4 cores) which manage to use all CPU resources. If I run the program on a 2PFLOPS hpc , How many hash can it generate per seconds?

  • Find out how many FLOPS your CPU does (your CPU document should tell you), then do the math (knowing it will be a very rough estimate.) Then again, all SHA calculations are bitwise operations and don't use floating point at all, so FLOPS probably isn't the right metric anyway. – John Deters Dec 15 '18 at 18:20
  • Why are you looking at FLOPS (floating point operations per second) when SHA512 does not use floating point operations, it used logical operations. There is a huge performance difference between floating point operations and logical operations. – zaph Dec 16 '18 at 14:19
  • There should be a "flagged as: homework" moderation option. – Royce Williams 2 days ago

Yes, you can probably get a better estimate by using the number and types of nodes, instead of overall FLOPS (as well as being FP, it includes GPU, Xeon Phi and large-memory nodes which won't be any use here).

The Octopus system you linked to has 236 general-purpose nodes and 37 GPU nodes, which are both essentially high-end PCs. Your program won't benefit from the GPU, and both types of node are otherwise identical, so you can treat it as a network of 273 PCs. Each one is a dual socket Intel Xeon Gold 6126, which is a 2.6GHz Skylake. Each of the two CPUs has 12 cores. (And 24 threads - but you can almost always expect hyperthreading to not help on HPC.)

So you have 12x2x273 or 6552 CPUs in total, and a simple brute force approach can probably scale up nearly perfectly. (This is not always the case, but it doesn't have to pass data around, it just gives each CPU a different fraction of the space to search.) Meaning that if the CPU cores are the same speed as yours, you would get a speed up of 6552/4 or 1638 times. They probably aren't the same speed of course, but this will give you a reasonable rough guess.

If you want an more accurate number, you could try comparing benchmark scores of your CPU with the HPC's CPUs and scaling the numbers - but don't expect great results from doing this, as all CPU benchmarks are testing different things and it won't necessarily apply directly to your case. (If you can find a SHA or other bitwise crypto benchmark, it might be closer than most, though.)

The most accurate way though would still be to run it on all CPUs of one machine which is as nearly identical to one of the HPC's nodes as possible.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.