New Samba 3 messaging implementations

[Aleksey Fedoseev]

James,

>>We are working on implementing Samba 3 for clusters. I've published the
>>introduction, several clustering problems and proposed decisions (they
>>were inspired by Volker's work) on the Samba wiki:
>>http://wiki.samba.org/index.php/Clustered_Samba.
>>
>>Any cluster implementation has to be based on some kind of messaging.
>>The first step we planned is to implement socket-based (or any
>>non-local) messaging in the trunk/samba-3 branch - upon that we'll be
>>able to extend the messaging with tcp- or any other transport like MPI
>>for cluster version of Samba.
> 
> 
> See branches/tmp/jpeach-cluster for an implementation based on shared
> TDB and TCP messaging. Note that the TCP messaging in this branch is
> obsolete WRT top of tree since that was removed post-3.0.14.
> 
> This also includes auto-configuration and membership tracking which are
> important issues for any clustered file server. You might also want to
> look at the redirect VFS module.
> 

Thank you, I'll check it.

>>Our messaging research is published on the wiki:
>>http://wiki.samba.org/index.php/Messaging. There three messaging
>>implementations (tdb-based, unix-dgram-sockets from Samba4 and
>>unix-stream-sockets) are listed, benchmarked and compared. Patches to
>>the current Samba 3 branch can be downloaded there.
> 
> 
> I presume the TDB messaging results were done mmap enabled? It would be
> interesting to see a result for TDB with mmap disabled, since some
> clusters (eps. heterogeneous ones) might not implement a coherent mmap.

Yes, mmap for TDB was enabled. I just wanted to get the highest speed
from the current implementation to compare with. I guess, TDB-based
version is unusable on most clusters, so I didn't check version w/o
mmaping. But I can do it soon.

> Also, unix domain sockets don't work across the all clusters. Do you
> know of any clusters that implement unix domain sockets?
> 
>>The main quesion is - what messaging extension can be added to trunk
>>branch, so we'll continue clustering research and shall implement
>>messaging-based locking for clusters?..
> 
> 
> My view is that none of the current messaging implementations are
> suitable for a cluster.
> 

Yes, I agree - this implementations can't be used directly on clusters.
But our goal is to replace current samba-3 messaging with any
cluster-friendly code - e.g. when unix domain sockets replace TDB
messaging, it will be easy to extend it to tcp sockets (or any other
kind of cluster-wide transport).

-- 
Aleksey Fedoseev
Linux Center of Competence, IBM EE/A
IBM Linux Technology Center: Samba Team
Phone: +7495 775-8800 x 2133