Re: A proposal for a General Clustering Framework

[Peter Badovinatz <tabmowzo@yahoo.com>]

--- irbis@orcero.org wrote:
<snip>
> 
> IMHO the proponsal of the paper is trying to solve things of the 6th OSI
> level on  the kernel, moving things that  are now being resolved by the
> applications in the kernel.
> 
>  For me, it sounds strange with OSI glases to put the session layer on the
> kernel clustering framework -this is the overview proponsal main idea-.
> 
>  I also can not find how to but all this XML stuff on HP clusters. Maybe
> is my lack of understanding, but all the problems that is trying to solve
> the proponsal were already solved by the application libs that are being
> used.
> 
>  Cluster transparency and performance -the two main interests of HP
> clustering- are not improved with this proponsal. And performace can be
> penaltied.

By transparency do you mean not having to program to an MPP model?  Or are you
just concerned with administration or job scheduling?

In any case, if you look at the IBM SP running PSSP+RSCT, you have an HP system
(PSSP) with an included HA framework (RSCT, aka Phoenix).  We had to do a
number of things with RSCT to deal with differing environments:
- tight HA, strict failover, shared data, etc.
- HP, with some HA needed due to clustered file system and remote data access

The latter assumed that we did not perform failover and recovery within 10s of
seconds, rather in terms of a minute or two.  This was for two reasons:
- avoid false failure indications due to longer message delays with large
numbers of nodes.
- absolutely minimise performance impacts to allow the programs (usually
distributed MPI jobs) to not be interfered with.  Note that to get this right
did require a fair amount of skill by the system admins and users to work out
job characteristics.  RSCT provided a wide variety of tuning handles, not all
of them sufficiently explained.

Benefits gained with RSCT across an HP cluster were consistent notification of
failures, scalable (500 node) cluster file system, event monitoring (events in
terms of what's been discussed in this thread.)  

But yes, avoiding ANY performance hit was impossible.  The key was to achieve
the level of monitoring and consistency you desired while minimising impact on
the system, a labour-intensive task unfortunately.

This does not say that the framework being proposed/discussed would be
immediately extensible to HP clusters, but it is not out of the realm of
possibility to consider it happening eventually.

Peter
> 
>  Yours:
> 
> David
> 
> ---------------------
> http://www.orcero.org
>   irbis@orcero.org
> ---------------------
> 


=====
These have been the opinions of:
Peter R. Badovinatz -- (503)578-5530 (TL 775)
wombat@us.ibm.com/tabmowzo@yahoo.com
and in no way should be construed as official opinion of 
IBM, Corp.

__________________________________________________
Do You Yahoo!?
Get personalized email addresses from Yahoo! Mail - only $35 
a year!  http://personal.mail.yahoo.com/

Linux-cluster: generic cluster infrastructure for Linux
Archive:       http://mail.nl.linux.org/linux-cluster/