[ofa-general] Re: Re: IPoIB-CM UC mode

Sean Hefty mshefty at ichips.intel.com
Tue Jul 3 15:45:47 PDT 2007


>>> What real advantages are there for doing this "in-band" as you say?
>> Doing this in-band keeps the entire keep-alive protocol within the ULP.  It can
>> set the keep-alive message size and retry times.
>> LAP messages are fixed at 256
>> bytes, add additional traffic on QP 1, and retries are limited by the CM
>> protocol.
> 
> BTW, I think we might want to avoid retries altogether: if LAP
> timed out, we can just re-create the connection.

The CM currently retries LAP messages based on the value of the REQ max 
CM retries, but I don't see why this couldn't change.

>> (Of course, new CM messages would have these same limits, so it's not
>> clear to me that creating new CM messages are a win.  New CM messages would
>> allow the CM itself to respond directly to keep-alives though.)
> 
> OTOH, using QP1 makes it easier to separate rare keepalives
> from fast-path data packet receive path.

I was thinking more along the lines of whether to use the CM LAP message 
or create a new CM message for handling keep-alive.  The best argument I 
can come up with for creating a new message is that it 'seems' 
cleaner...  Anyway, I agree that using LAP would be the best approach 
for now.

- Sean



More information about the general mailing list