[rds-devel] ***SPAM*** Re: [ofa-general] [RFC] dropping RDS over TCP support
Richard Frank
richard.frank at oracle.com
Tue Sep 2 15:42:05 PDT 2008
Jon Mason wrote:
> On Wed, Sep 03, 2008 at 12:48:53AM +0300, Or Gerlitz wrote:
>
>> On Tue, Sep 2, 2008 at 11:04 PM, Andy Grover <andy.grover at oracle.com> wrote:
>>
>>
>>> We've been discussing dropping RDS's support for using TCP as a transport,
>>> and just focusing on RDS as a IB and iWARP-focused protocol.
>>>
>> do we have any results that compare Oracle IPC using rds/bcopy/tcp vs udp?
>>
>> This would simplify the RDS codebase, allow easier inclusion of more
>>
>>> IB-centric features, and also give RDS an easier path towards mainline Linux
>>> kernel inclusion. Also, the imminent RDS iWARP support will address non-IB
>>> use cases.
>>>
>> So just to make sure, do IB and iWARP share the same transport code today?
>> if yes, does removing TCP means the transport abstraction would not be
>> needed any more, or you still want to maintain it for the loopback case?
>> Generally speaking, the loopback transport also uses IB, correct? and if it
>> doesn't I am quite sure it can.
>>
>
> Not all iWARP devices can do loopback, so having it done in the IB
> specific code could be painful. I believe there is a loopback module in
> RDS which can handle this though.
>
Currently, loop back (connecting to local ip:port) is handled by the
transport (with IB it uses a local IB RC) - we moved to this to simplify
the driver and support things like performing local process comm
including rdma'ing between processes - which is a key feature in of itself.
If a particular transport can not support this - then it should either
emulate operations when possible - or fail them. For example, getting a
key for rdma may fail if the connection is loop back.
>
>> I tend to agree with Jon that removing TCP might help with mainline
>> inclusion or might create damage...
>>
>> Roland, maybe you have more definitive intuitions re the netdev people
>> potential feedback on rds as a new socket type applicable to RDMA cards such
>> as IB and iWARP using a verbs/rdmacm native transport AND to non RDMA cards
>> with TCP transport, vs the case of RDS being "just" a ULP under
>> drivers/infiniband/ulps that defines a new socket type, etc.
>>
>>
>> Or
>>
>
>
>> _______________________________________________
>> general mailing list
>> general at lists.openfabrics.org
>> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
>>
>> To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
>>
>
> _______________________________________________
> rds-devel mailing list
> rds-devel at oss.oracle.com
> http://oss.oracle.com/mailman/listinfo/rds-devel
>
More information about the general
mailing list