Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <f629a972016cca474e873c0d5f506b73@smtp.hushmail.com>
Date: Fri, 2 Sep 2016 19:20:03 +0200
From: magnum <john.magnum@...hmail.com>
To: john-users@...ts.openwall.com
Subject: Re: Question regarding CPU load on GPU tasks

On 2016-09-02 13:53, Sebastian Heyn wrote:
> Now I noticed when I run a wpapsk-opencl vs a wordlist, the cpu load on a single core is 100% solid.

I presume we're not talking about running OpenCL on CPU here. Are you 
using an nvidia GPU? A silly property of their OpenCL runtime is it 
busy-waits for GPU events. So it looks like it's pegging the CPU but in 
reality it does no work on it.

> When I fork the task to 4, it is 4x100% (its a quad core cpu) - however the amount of passwords per second is the same. I dont know how to interpret that result. Is the cpu 100% flat out due to some loop in the code, or is it really 100% busy? If so the forking should show some improvement, right?

For WPAPSK there's no point in forking a run with OpenCL formats (unless 
you have several GPU's - then you'll fork one process per GPU). If you 
really have 4 GPU's, it's the same answer as before: Busy-waits. No real 
bottleneck.

> I am asking because I wonder what would happen if I bought a larger gpu (or 2) - would the cpu become a bottleneck?

Not for this format anyway. It runs almost entirely on GPU(s).

> Second question: Is openmpi also able to run gpu tasks on several machines?

Sure.

magnum

Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.