CTDB - the 'Mainz' plan for clustered Samba
tridge at au1.ibm.com
Mon Oct 2 10:19:59 GMT 2006
> I'm not planning on using the existing messaging code. I want us to be
> able to take advantage of hardware messaging libraries, and I also
> don't want to tie the code to Samba initially.
explaining this a bit more ....
I imagine we'll prototype this with tcp or udp, but we will really
need to support either MPI or another messaging library which can talk
to the sort of fast messaging hardware available on modern clusters.
The trick will be to support those libraries, but also to receive
messages from smbd efficiently at the same time, hopefully via unix
datagram sockets. What I don't know yet is how to combine the two
without polling. It's easy if we use udp or tcp, but I don't yet see
how to do it with something like MPI.
It's also not clear how we handle cluster management with something
like MPI. A MPI program doesn't usually add new nodes while its
running. Maybe a lower level library will make this easier.
Anyway, I'd like to prove the validity of the design first using udp
or tcp, but not tie it to those transports, and not tie it into Samba
in any deep way.
More information about the samba-technical