Security question?

[Jim Gettys]

On Thu, 2004-11-04 at 10:53 +0000, Alan Cox wrote:
> On Iau, 2004-11-04 at 03:11, Jim Gettys wrote:
> > Not to mention the fact that Unix domain sockets on Linux are really,
> > *really* fast; performance data of that age is in general suspect; it is
> > several major releases old (at least).  And Futexes are indeed a new
> > very interesting possibility.
> > 
> 
> It would suprise me greatly to see much difference. All your real cost
> should be "put bits in cache" "context switch". That comes down to the
> efficiency of message sizes and write batching.

Which also hasn't been carefully analyzed for years, if not decades.

I'm not aware of any serious work there since I did work in that area
in X6 days (but as there is a serious gap in my X work, I may not be
reliable.  Machines are *very* different (as is X's pattern of use) than
they were when I did that on MicroVAX II's.
> 
> > work on X is now *very* stale in general; we have better facilities
> > than select/poll these days as well (and a typical X server now
> > has many more connections open; 25 as I compose this message).
> 
> We have epoll() but I've seen no evidence it makes any difference at
> such a low connection count. OTOH at 1000 connections (eg big irc
> servers) its very visible. I believe the BSD folks have a similar
> interface to epoll for event based waits
> 

Yeah, the first thing to do is to get some hard data about where time is
going.  A fun thing for the performance inclined (as I've been in the
past).  Anyone up for inserting data into the discussion?
				- Jim