john-users - Re: core by core speed

Openwall GNU/*/Linux 3.0 - a small security-enhanced Linux distro for servers

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date: Sat, 20 Mar 2010 01:57:00 +0300
From: Solar Designer <solar@...nwall.com>
To: john-users@...ts.openwall.com
Subject: Re: core by core speed

On Wed, Mar 17, 2010 at 11:01:34PM +0100, websiteaccess@...il.com wrote:
>  JTR 1.7.5 + 2 latest patch compiled x64 (os X)
> 
> 
> http://img52.imageshack.us/img52/4053/picturesy.png
> 
>  As you can see there is a big difference between core speed :(  from 
> 19% to 98 % time processor)

There could be many different things causing this - the Idle setting
working poorly on your system, those processes doing a lot of file I/O
(waiting for disk), or maybe this is proper reporting of CPU usage for
threads sharing a CPU core (your CPU is quad-core with 2 threads per
core, not 8-core).

When I made Idle = Y the default in JtR 1.7.4+, I asked john-users to
benchmark and see if there was any negative effect of this change, then
report back to me.  No one did. :-(  (There were some benchmarks of
other properties of JtR.  BTW, the Idle setting does not take effect
when running with "--stdout".)  I only tested this well on Linux, so I
would not be too surprised if this change had an adverse effect on
another system, especially with SMT (Hyperthreading).  I'd appreciate it
if you benchmark Idle = Y vs. Idle = N now - first on a non-MPI build,
then with 4 processes, and finally with 8 processes (6 tests total).
Record the c/s rates and watch the CPU usage by the processes.  No need
to run them for long - 1 minute will be enough.

Thanks,

Alexander