Many (all?) of you may have seen this, but I thought I would point out a paper on implementing fast pattern matching for Snort using CUDA called Gnort that came out over a year ago: <a href="http://www.ics.forth.gr/dcs/Activities/papers/gnort.raid08.pdf">http://www.ics.forth.gr/dcs/Activities/papers/gnort.raid08.pdf</a> . I tried to get the source code from one of the authors last year, but the main author was away in the military at the time and so the author I was communicating with couldn't send me the source code for the project. However, the paper may lend a hand in designing the architecture for getting real performance out of the current Suricata CUDA code, as Gnort claimed exponential performance increases. Also, Google points out many other articles that cite that paper, including ones from IBM's SCAMPI project which appears to attempt to do the same thing as Gnort. It may be too high-level to be helpful in tweaking the Suricata implementation, but I thought I could at least make sure everyone was aware of some the current work on the subject. Also, I notice that GPU's like the GeForce GTX 295 come with 480 stream processors for around $550, so it probably wouldn't be bad to assume there could be over 1024 stream processors available to the CUDA API, in case that makes a difference.<br>
<br>Thanks,<br><br>Martin<br>