06-21-2016, 02:59 AM
I agree, 10% less performance on a cluster dedicated to gpu-hashing wouldn't make sense. I think you misinterpreted part of my post. The point was to make public my results on running oclhashcat in a VM. My finding was that using the GPU in passthrough mode probably does not affect performance and that it is therefore a viable way to optimize hardware usage and ease management. I mentioned availability as a benefit in the explanation as to why, generally, IT infrastructure is being virtualized.
First, my result was 2,5% less md5 speed, compared to an unsubstantiated claim from someone using different hardware and probably different oclhashcat version
Second, my result was also 15% faster than the wpa2 crackstation benchmark; as you can see, I lack proper data to compare too, however... since the result was either faster or very slightly less, I concluded that no, or very slight, performance impact is introduced.
To make my stance on this clearer then:
First, my result was 2,5% less md5 speed, compared to an unsubstantiated claim from someone using different hardware and probably different oclhashcat version
Second, my result was also 15% faster than the wpa2 crackstation benchmark; as you can see, I lack proper data to compare too, however... since the result was either faster or very slightly less, I concluded that no, or very slight, performance impact is introduced.
To make my stance on this clearer then:
- If you have several nodes doing nothing but hashing, all (but perhaps one) might as well remain dedicated because you're using so much money on hardware that you can afford to solve your other computing needs better.
- When you are evaluating new drivers/versions/OS, you should perhaps consider using a VM because the shortcuts available will not only save you time, it will make your work less tedious. And you could test it on one GPU, while the other 10 keeps working.
- If you have a small number of nodes and would like to not have a separate set of compute nodes, installing a hypervisor could be a good idea if performance doesn't suffer; as I wrote, more testing is needed.
- If you're like me and have: one good lab server, limited budget, space and time and just want to be able to do some hashing while using the box for other things too: this is definitely the way to go even if performance suffers a bit.