[Linux-cluster] Re: [cgl_discussion] Re: [dcl_discussion] Cluster summit materials

[Lars Marowsky-Bree]

On 2004-08-11T14:24:49,
   Daniel McNeil <daniel at osdl.org> said:

> How can the DLM go to Andrew without a membership layer to
> provide membership?

I'd agree with this question. Membership is really the first and
foremost question, then the DLM can go in.

Fencing turns out to be a more difficult beast, because the way how the
GFS stack handles it's recovery (a static priority list) is somewhat
fundamentally incompatible with the way how a more powerful dependency
based cluster recovery manager might wish to handle things. We've just
run into this discussion ourselves, and as soon as we have an idea, will
propose that adequately for discussion...

> I think John really does mean communication.  For high availability,
> the cluster should have no single point of failure. 

Exposing the communication APIs begs a ton of questions regarding the
semantics; atomic, causal or total ordering?; communication groups;
access controls to those; sync or async; broadcast, multicast or
pair-wise channels?

All of these and some more can/should be supported, however most systems
just provide subsets. How to expose that, how to handle it?

That's a bit more difficult than answering the question about
membership, which is even complex enough - do you get to see membership
before or after fencing, with or without quorum etc.

Don't rush this. Don't get sidetracked. (And trust me, I've been there
at OCF for that one.) Concentrate on the slightly more palatable ones
like membership and DLM, and after we've established prior art, then
lets tackle the bigger issues.

Nobody denies that communication, recovery coordination etc are required
and very important, just that we don't wish to start there.

> Does CMAN provide this kind of functionality?  If so, then it
> really is a communication service.

It provides a very limitted subset of it which is, for example, not even
useable to the low requirements SCRAT (heartbeat's new recovery/resource
manager) has, as far as I can see, because it's not performing well
enough. And it's not meant to, because they architect their stack
differently (around DLM + TCP etc), but it means we'll need to work on
this area some more first.


Sincerely,
    Lars Marowsky-Brée <lmb at suse.de>

-- 
High Availability & Clustering	    \ Philosophy proclaiming reason to be 
SUSE Labs, Research and Development | the supreme human virtue is falling
SUSE LINUX AG - A Novell company    \ prey to self-adulation.