infiniband: ?
Alexander Bokovoy
ab at samba.org
Sat Nov 18 20:58:54 GMT 2006
Tridge,
Andrew Tridgell wrote:
> Alexander,
>
>> This was what actually led me to suggest doing socket-like
>> interface to Peter, as we anyway will have TCP/IP version and code
>> sharing would be good enough. In case this wouldn't fly most of
>> infrastructure work would still be in use by TCP/IP backend.
>
> ok, though we will have to test to see if using a sockets like API
> adds any latency. The CTDB protocol really wants a datagram style
> interface, where the breaking up of the command stream into packets
> is done below the ctdb code. On TCP I'll be faking that in the usual
> way, but if IB can do it already then it would be nice not to lose
> that.
Agreed.
>> However, I would also like to see this done through SDP (sockets
>> direct protocol) as it bypasses TCP/IP overhead while maintaining
>> socket-compatible API. The reason for that is because Infiniband
>> itself is a reliable transport. May be going through SDP would be a
>> better thing?
>
> I haven't looked at SDP yet, though I did come across a 2004 web page
> saying that Microsoft claimed some rights to SDP, implying that it
> might be patent encumbered. Can you and Pater check on the status of
> that before you commit to SDP ?
After some research: SDP is a successor to Windows Sockets Direct
technology and while being proposed for IETF as standard by RDMA
Consortium, it is a licensed technology offered by Microsoft -- see
http://www.microsoft.com/about/legal/intellectualproperty/standards/ .
Formally, SDP is an optional normative annex to the InfiniBand 1.1
specification.
So making a datagram-based or TCP/IP- (IPoIB) like layer might be a
better option. I'm still looking at the possibilities to get something
common to IB and Myrinetso that both commonly used high-speed networks
would be effectively useful for CTDB.
--
/ Alexander Bokovoy
Samba Team http://www.samba.org/
ALT Linux Team http://www.altlinux.org/
Midgard Project Ry http://www.midgard-project.org/
More information about the samba-technical
mailing list