Ideas from GPGPU Forums:

GPGPU forum

woody said:

"Interesting idea. I haven't come across anything on this theme so far, but it has a lot of benefits.
I'm not sure what sort of performance you can achieve, but you could certainly do off-line packet analysis (eg. snort) operations, or do something like encoding 100 or so packets in rows in a texture and processing this texture once a second. So long as the processing time isn't greater then once a second, you're still processing every packet.

There's likely to be a few problems with recording information between rendering passes (say for classification/detection), but with MRT and RenderTextures (espec. with the soon-to-be-supported EXT_framebuffer_object extension), these should be able to be sorted out with minor detriment to performance. "

------------------------------

MRT (Mutliple Render to Texture):

RenderTexture:

EXT_framebuffer_object:

Snort on the GPU !!!

The idea is to try an implement the computationally intense portions in Snort on the GPU. Snort needs to operate at network speeds (gigabytes/sec) so much effort has gone into optimizing the Snort's packet scanner. Hopefully, the GPU can be used to really speed this up.

Some ideas/thoughts:

• The efforts the optimize the AC algorithm have focussed on encoding approaches, such as how the state machine is encoded. These usual rely on sparse-matrix or vector representations.


• the state-machine could be first encoded as a texture which is then accessed in a read-only fashion by texture sampling. Good idea.