Re: RFC 1812, section

Tony Li <> Sun, 17 September 1995 07:49 UTC

Received: from by IETF.CNRI.Reston.VA.US id aa06183; 17 Sep 95 3:49 EDT
Received: from CNRI.Reston.VA.US by IETF.CNRI.Reston.VA.US id aa06179; 17 Sep 95 3:49 EDT
Received: from by CNRI.Reston.VA.US id aa19758; 17 Sep 95 3:49 EDT
Received: from by (5.65c/5.61+local-22) id <AA06404>; Sun, 17 Sep 1995 00:23:49 -0700
Received: (tli@localhost) by (8.6.8+c/8.6.5) id AAA22468; Sun, 17 Sep 1995 00:23:48 -0700
Date: Sun, 17 Sep 1995 00:23:48 -0700
Sender: ietf-archive-request@IETF.CNRI.Reston.VA.US
From: Tony Li <>
Message-Id: <>
In-Reply-To: <>
Subject: Re: RFC 1812, section

   I find it difficult to believe a 25% performance hit considering the other
   sanity checks that probably have to be made, but then I've never counted
   cycles through our fastest single-processor path...

I don't know exactly what Craig's working on, but I suspect that it
causes him to pull in at least two extra longwords, thus screwing up
whatever clever memory controller (or pipeline?) that he's thinking
of.  Note that if you take interpretation 2, the fast path need only
reference the version/IHL byte, the TTL and checksum longword, and the
destination address longword.

You can ALSO decide that you want to play games with the TOS byte, and
still only be referencing 50% of the IP header.

And before someone flames me, yes, I realize that conventional memory
controllers are pulling in big blocks efficiently.  I _assume_ that
Craig is doing something more interesting than the conventional.  ;-)