[NBLUG/talk] Uses for GPGPU?

[Aaron Grattafiori]

On 22:21 Wed 14 Jan     , Lincoln Peters wrote:
> On Wed, Jan 14, 2009 at 10:00 PM, Tim Preston <timp at sonic.net> wrote:
> > OK. Am I missing something? We are talking about increased processing power
> > and utilizing the processors on all devices within a system in parallel to
> > accelerate programs that need to crunch lots of code and the only thing
> > we're talking about is hacking?
> >
> > How about improved identification of facial features in streaming video? Or,
> > maybe tracking moving objects in real time with an improved rate of
> > accuracy?
> 
> I know a guy who was working on using a GPU program to analyze a
> recording of someone's voice and determine whose voice it is.  I think
> the algorithm he's using is open-source (it was developed by
> researchers at MIT), but his specific implementation is proprietary.
> 
> I was actually thinking it might be useful for cyber-security, e.g. if
> you used a GPGPU-based application to perform real-time packet
> inspection, you might be able to identify threats that would otherwise
> be way too hard.  One example that comes to mind (although I'm not
> sure how practical it would be) would be to analyze incoming data
> files during transfer (i.e. in real-time) e.g. for the presence of
> valid Intel CPU code (something that I don't think your average data
> file should contain).  If that worked, it could protect against buffer
> overflow attacks, even if the vulnerability was not known to anyone
> but the attacker.
> 


IDS/IPS systems like you're describing are VERY hard to get right and most
client side attacks would not be caught. Most server side "0day" attacks 
would be better prevented with kernel level hardening like grsec/PaX or
SELinux. IDS/IPS systems can also cause other problems.

I also think something like this would be too "generic" to really take
advantage of the GPU's parallel (matrix operation oriented) nature.
If you used an anonomly baised IDS system and then came up with a
complex (and low-false-positive) algorithm to look at data flows and spot 
"strange" behavior. Then ran this on your GPGPU-based system, it might
be better suited for the job. Although also not very useful because its
reactive rather than proactive (like all IDS systems, they are like car
alarms). "Oh shit, I'm already owned".

-Aaron Grattafiori

> 
> -- 
> Lincoln Peters
> <anfrind at gmail.com>
> 
> _______________________________________________
> talk mailing list
> talk at nblug.org
> http://nblug.org/cgi-bin/mailman/listinfo/talk