[ofa-general] Re: Re: IPoIB-CM UC mode
Or Gerlitz
ogerlitz at voltaire.com
Tue Jul 3 22:54:35 PDT 2007
Tziporet Koren wrote:
> Or Gerlitz wrote:
>> Dror did not mention the HW, my understanding is that this aspect is
>> fine... now, assuming the need for liveness protocol is behind us, and
>> if not, it can be implemented as you suggested. the problem is
>> narrowed to have the FW support SRQ/UC.
> We still don't have a solid plan for this in the FW
The current implementation of IPoIB-CM uses RC, where Michael has
admitted that the RC ACKs are actually used as keep alives, no more.
-> under high packet rate there are some ten (hundred?!) --thousands--
keep alive messages/second.
This is --very-- poor architecture, the code must move to UC.
The only actual barrier here is FW support for UC/SRQ, the IBTA
signature can come later (and is on its way).
Or.
More information about the general
mailing list