I don't recall this issue being brought up at inception
and as a date for commitment is still unknown, I'd like
to raise this question/issue earlier so that some thought
can be given to it in the preparation for commitment and
hopefully some discussion included in the design document.

The question of what happens with input packets vs output
packets may need separate treatment.

Darren
....

For brevity, I'm referring to the remodeled IPMP architecture
being introduced as IPMPv2 and the current/old model as IPMPv1...

The document presented for inception review of PSARC/2007/272
appears to contain no material that discusses the interaction
of NIC capabilities (such as LSO and hardware checksum offloading)
with IPMPv2.

Specifically, when compared with the existing architecture
and the use of logical IP interfaces, the real NIC that is
to be used in sending packets out of the system is hidden.
So whereas before we might of had IPMP on nxge0:1 and TCP/IP
would be able to see what capabilities nxge0 had and thus
adapt to it, there is no mention in the new design document
of whether or not this capability is exposed to TCP/IP and if
so, how.

Thus, users of IPMPv2 may potentially see a drop in networking
performance, if the hardware was previously providing them with
measurable aid (when either using IPMPv1 or no IPMP.)

If NIC capabilities aren't being propagated up so that TCP can
see them, is there a long term design or plan to correct this?
If they are, can the document please be updated to include how
or when this is done.


Reply via email to