[Snort-devel] Thoughts on threads

Nash nash at ...357...
Fri Apr 6 14:20:38 EDT 2001


I 2nd this emotion. Making the paengines and the decoders as fast as possible
in code (w/out threads) makes good sense. Adding the other functions as threads
that might run slower seems very appropriate. Plus, the pitfalls of such a 
producer / consumer model are fairly well known and there's lots of good example
code. Mmm ...  maybe I'm just lazy. ;-)

In response to Fyodor's suggestion in regards profiling. Well, it just seems
like optimizing for the special case and that seems like a dubious idea. 

My $0.02. 

ciao,

	gnash

On Fri, Apr 06, 2001 at 02:00:10PM -0400, Jon Bentley wrote:
> I haven't received a good beating recently, so I'll through my two cents
> into the ring.
> 
> Threads (nee parallelization) would cause me some concern, as it would
> potentially remove the serial order of received packets.  Perhaps that is
> a concern of only myself, though.  (Packet sequence numbers, with a post-
> process reordering?)
> 
> I'd rather see a thread dedicated to an atomic task, rather than full
> sequence
> execution.  This gives the ability to buffer individual task results, and so
> do
> basic prioritization among tasks.  (To wit, if a single thread processes
> every
> aspect of a packet, and I get hit with a burst of packets, I'm going to lose
> some.
> However, if the front-end task merely traps and buffers packets, with a
> lower
> priority thread doing post-processing, then bursts are no problem.)
> 
> Pthreads are great, but do we care about our W*ndows friends?
> 
> 
> 
> _______________________________________________
> Snort-devel mailing list
> Snort-devel at lists.sourceforge.net
> http://lists.sourceforge.net/lists/listinfo/snort-devel

-- 


   "Babbage himself acknowledged Jacquard's precedence: when
   he presented the concept for his Analytical Engine at the
   Turin conference, he brought  with him a silk portrait of
   Jacquard  that  had been  produced  by an  automatic loom
   programmed  by no fewer than twenty-four thousand  cards. 
   Even by today's standards, that's a lot of code."
                                - Jim Holt,
                                  The New Yorker 2001/3/5






More information about the Snort-devel mailing list