[6bone] Re: IPv6-only IXP's are absolutely wrong

[Bill Manning]

% On Mon, Oct 21, 2002 at 10:18:42AM -0400, Eric Gauthier wrote:
% > I don't know about the various IX's, but the I2 gigapop that our University
% > uses is concerned about this.  If I remember correctly (not that the exact 
% > numbers are important), but IPv4 is in an ethernet frame with type 0x0800 and 
% > IPv6 is in an ethernet frame with type 0x86dd, so layer 3/4 aware switches 
% > will likely handle these frames differently.  In our case, the Cisco 12,000 
% > and 6500's that we're using are great for IPv4 packets (they handle them
% > in hardware), but IPv6 packets are handled in software so we don't expect 
% > nearly the same type of performance.  I'd imagine that something like this 
% > is what they're alluding to.  
% 
% This is definitely relevant for the individual participants - but for
% the IXP switch (which is strictly layer 2 *only*, at least for all IXes
% that I know), L3/4 forwarding performance should not be an issue.
% 
% One issue that I see is multicast (neighbor discovery etc) which isn't
% seen on an IPv4 unicast exchange switch.
% 
% Gert Doering

	reasons to split v4 and v6:

	multicast - v4 and v6 treat this differently.
	fabric "optimizations" - framing support, buffer sizing, MTU, etc.
	a decent L2 fabric will be able to accomodate the larger MTUs
	of native v6 and won't complain about divergent framing.
	"Smart" fabrics tend to be tuned to IPv4.

	then there are issues wrt RA/ND on an exchange... having all the
	participants trying to "stamp" the fabric with their version of
	which prefix to use is noisy at best and an effective DOS at worst.

	then there are mgmt issues. most L2 fabrics do not have up to date
	v6 mib support, so the stats/traffic collection is not as accurate
	as it should be.

	these things, in addition to the issues with the connecting gear
	(v4 in HW, v6 in SW), tend to argue that a mixed-mode exchange may
	be less stable / harder to troubleshoot than an single use 
	facility.  The offset is the capex/opex costs of connecting to two
	facilities, one for each protocol.

-- 
--bill