ctdb/tcpip: duplicate connections?
Peter Somogyi
psomogyi at gamax.hu
Tue Feb 13 17:47:25 GMT 2007
Hi Tridge,
Currently I'm integrating ib into ctdb and noticed the following (=> I'm
unsure how to connect logically a transport connection and node):
I suspect you're making duplicate tcp/ip connections.
I mean that you make a separate tcp/ip connection from all the N (=here 2)
nodes to each other:
127.0.0.1:9001
127.0.0.2:9001
For infiniband the same model is a sure redundancy (a connection is duplex)
and waste of resources (2x). Of course, in ib it's also possible that you can
have 2 connections for a node (1 for send + 1 for receive).
For accept()-like functionality in ib you get a different connection. For
tcpip's accept() I suspect you also get a different fd...
If yes: after you receive a message in ctdb_recv_pkt, in ctdb_queue_packet you
will answer on a different tcpip connection (different fd?)...
My questions:
- is it intended?
- will it change in the future?
- wouldn't be better to answer to a request on the same connection (like in a
normal client-server scenario)?
One notice: I'm still missing a node removal scenario... I mean e.g. can't
find ctdb_tcp_node_disconnect. For ib it's ready and tested.
Peter
More information about the samba-technical
mailing list