[Planetlab-devel] IPv6 support for MyPLC

mef at CS.Princeton.EDU mef at CS.Princeton.EDU
Tue Nov 7 21:36:45 EST 2006


Hi Jim,

> For IPv6 for a link it can support multiple prefixes per link. So lets
> say on PLL vserver is its own link.  There could be multple prefixes for
> one NIC all using the same EUI for IPv6.  Depending on the PLL prefix
> you select will identify what each link vserver uses.

Could you explain this in a bit more detail or refer me to the 
appropriate document.

> Thus I think
> stateless autoconfig can work just need good router that handles the
> node discovery in front of PLL vserver node.

I don't think we can make this assumption @ 300+ sites.  Moreover, at 
most sites IPv6 use will likely be new to most network admins --- 
certainly is the case at Princeton.  From what I gather, Princeton's CS 
department has given us a single prefix for our link.  For this reason, 
we have to do the various autoconf address assignment tricks that I 
mentioned before.  I am not 100% what the ramifications of such a 
limitation are just yet (e.g., does it limit our ability to let 
researchers fully experiment with multicast etc?).

> All the choices marc
> listed for DNS are a decision to be made.  Just so you know years ago we
> moved all the IPv6 node discovery ICMPv6 packets up to user space and
> then from there did DynDNS (yours truly wrote that code :--)).

Neat. I think ICMPv6 ND is now handled in the kernel.  I expect that 
our root context (the "sliver" that manages all other slivers on a PL 
node) will likely implement its own DNS server and directly handle 
DynDNS when it assigns IPv6 addresses to the slivers.  As we "trust" 
the root context, we likely can give it some sort of shared secret key 
that it can use to dynamically register IPv6 addresses with some other 
DNS server.

> Bottom line this is very doable but it will take some tinkering with the
> interface to the vserver incantations and sounds like some code changes?

Code changes: yes.  I am not sure we can use the link prefix trick, as 
it might assume something about the configuration at sites that we 
likely wont be able to mandate.

Marc



More information about the Devel mailing list