[openib-general] TCP/IP connection service over IB

[Caitlin Bestler]

> -----Original Message-----
> From: openib-general-bounces at openib.org 
> [mailto:openib-general-bounces at openib.org] On Behalf Of Steve Wise
> Sent: Friday, October 21, 2005 1:35 PM
> To: Sean Hefty
> Cc: swg at infinibandta.org; openib-general at openib.org
> Subject: Re: [openib-general] TCP/IP connection service over IB
> 
> > Random thought... if the src and dst IP addresses will always be on 
> > the same network, the data could be layed out as:
> >
> > network addr (x)
> > src host addr (y)
> > dst host addr (y)
> >
> > This could save enough space to provide 64 bytes of user 
> private data. 
> > Although my preference would be to keep it simpler.  (I'm not that 
> > familiar with IPv6 addressing.  How does it define network 
> versus host
> > addressing?)
> 
> I don't think you want to make the assumption that src and 
> dst addrs are on the same IP network.  While that may be true 
> for a set of IB hosts on a common IB switch set as a single 
> IP subnet, there may be TCP/IB bridge/gateway products that 
> allow remote IP hosts to connect into an IB cluster and those 
> could certainly be on a remote subnet.
> 
> IPv6 addrs define networks in a similar manner to IPv4:  IE 
> some number of the bits in the address define the network 
> number, and the remaining define the host number.
> 

More relevantly, GIDs are syntactically identical to IPv6.
The only real difference between a GID and an IPv6 address
is who/how the network portion is assigned.

While IPv4-only applications exist, they are not supposed to.
Certainly no new API or protocol should encourage an application
to be IPv4 dependent.

The rationale for using only the IPV4 format would be that
GIDs *could* be assigned that are valid IPv6 addresses. Hence
no translation would be needed.

Relying on assignment of IPV6 compatible GIDs may be undesirable,
however, because HCAs generally cannot accept a large number of
assigned GIDs.

The amount of private data supported does vary on network characteristics.
A responsible application should be allowed to piggy-back additional data
when the larger size is support (512 bytes over IP networks). Forcing the
application to have to use an additional round-trip over the network would
not be network friendly. However, it should be clear to application
developers that they are expected to make their applications work
in the minimum size guaranteed -- and that the minimum size is adequate
for the core purpose of enabling the QP to be selected/configured.