[Lustre-discuss] Lustre support for ipv6

[Nicolas Williams]

On Fri, Jul 24, 2009 at 07:37:11AM -0400, Chas Williams (CONTRACTOR) wrote:
> In message <1248420309.21455.12.camel at berloga.shadowland>,Alexey Lyashkov write
> s:
> >On Thu, 2009-07-23 at 15:53 -0500, Nicolas Williams wrote:
> >> On Thu, Jul 23, 2009 at 09:07:39AM -0500, Mitch Felton wrote:
> >> > Does anyone know if Lustre 1.8 supports ipv6?
> >> 
> >> I believe it does not.
> >> 
> >this should be work for tcp LND, but never tested before.
> 
> i doubt it.  when i looked at the lustre code i only saw AF_INET.
> if there was any sort of ipv6 support there would have to be atleast
> a little mention of AF_INET6 for some of the socket operations.

Exactly.

Also, LNET has its own addressing scheme using 64-bits, of which 32 are
used to hold an IP address in the case of IPv4.  To add IPv6 support
with addresses other than IPv4-mapped IPv6 addresses would require
significant changes[*].  But IPv6 with IPv4-mapped addresses should be
feasible with changes localized to the sock LND.

We should have a plan for IPv6 though.

[*]  If we want to keep the model of including the IP address in the
     LNET NID, then the LNET NID will have to grow much larger.  That
     would be a huge change.  But we could add a layer of indirection so
     that the LNET NIDs don't need to change size.  In that case most of
     the changes to support IPv6 with non-IPv4-mapped addresses ought to
     be fairly well localized to the sock LNDs + an addressing service
     to manage the indirection.

Nico
--