Have we been having problems with too many flows being revalidated? It would seem to be to be fairly simple to change the tag_type to uint64_t. The extra cost of the datatype would pale in comparison to the cost of revalidating a couple of extra flows.
Ethan On Mon, Mar 28, 2011 at 2:54 PM, Ben Pfaff <b...@nicira.com> wrote: > On Mon, Mar 28, 2011 at 02:36:58PM -0700, Ethan Jackson wrote: >> > Well, in some cases, we *do* use a deterministic hash function to >> > generate tags (see tag_create_deterministic()), but in general I >> > regard randomly selected bits to be better, when we can afford to >> > store them, because to my mind it makes it harder for attackers to >> > force collisions. >> >> Oh sure, I wasn't criticizing the design decision, I hardly think >> storing 32bits of information per flow is an issue, and we probably >> get much better hashing characteristics out of this approach. > > I wasn't taking it as a criticism, for what it's worth. > > I'm actually not all that happy with the design. It's clever, but > that's not really a good thing. I keep hoping that some better design > (that is, one that is precise about what flows need to be revalidated > but still does not require a lot of careful maintenance of lists or > bitmaps that take up a great deal of memory) will magically > materialize. I doubt it, though. > _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
