[openib-general] RE: osm_sa_pkey_record.c when VENDOR_RMPP_SUPPORT defined

[Hal Rosenstock]

Hi Eitan,

On Sat, 2005-01-15 at 01:48, Eitan Zahavi wrote:
> The spec. does not require a GetTable response to be RMPP (only
> optionally) so we can easily avoid the RMPP overhead for short
> responses. 

I'd be curious where the optionality of using RMPP for GetTable is in
the IBA spec.

> If you have seen such a requirement please point the exact spec
> paragraph that require an RMPP response.

C15-00.1.18 states that the SA shall respond to SubnAdmGetTable requests
by performing the sender role in a receiver initiated RMPP transmission.

Also, GetTable is not optional unlike GetTraceTable or GetMulti which
are.

> Sending a single mad RMPP response is highly inefficient. (due to the
> protocol overhead).

The categorization of this as highly inefficient is not quite an apples
to apples comparison as RMPP provides reliability which would require
timeout/retransmission without it. For a single packet exchange, isn't
is just a DATA packet in one direction (indicating first and last) and
an ACK in the other.

The efficiency or inefficiency of course comes down to the
implementation. In any case, I believe IBA requires it. It is a
conformance/compliance issue.

> However, the implementation of the RMPP transaction (SAR) can handle
> single mad RMPP transactions too.

But it sounds like a change is required to do this based on how this
indicated currently (based on whether it fits in the MAD). Some other
way would need to be used.

-- Hal