Re: todolist

Hi,

My 2 ecu's worth...

First off, my understanding is GFS isn't terribly
efficient. Lustre would seem to be the best clustered
filesystem out there. The main drawback of Lustre is
that it only supports ext3.

Support for other filesystems can probably be hacked
in without too much difficulty, and might even improve
Lustre's chances of being accepted by Linus, as it
would mean it is correctly coded to the filesystem
layer rather than to a specific set of underlying
operations.

Next up, shared memory. I believe MAASK's website has
some preliminary distributed shared memory (DSM) code
that works with OpenMOSIX. (Ok, I know they do, 'cos I
entered the record into Freshmeat. :)

Ok, distrbuting other memory resources. RDMA patches
exist for Linux, although they're mostly aimed at
Infiniband. I don't see any good reason not to hack
the interface so it'll work with anything OpenMOSIX
knows about.

Distributing other information. That's what the Plan 9
filesystem is supposed to do in Linux - distribute
access to other resources. It makes no sense
whatsoever to not use it.

Non-TCP/IP networks. GAMMA - "active messages" over
ethernet - look extremely promising there. It would be
good if OpenMOSIX could use GAMMA as an alternative
network device, although GAMMA really could do with
working with a wider range of hardware.

I'm working on some proprietary projects involving
clusters, so I am sticking to information you can pick
up from Freshmeat and then a search of Slashdot. This
is public information. Drawing together these
technologies would give you what I would consider a
fairly baseline clustering system. None of it is hard
to glue together - the only hard part is updating the
glue whenever a component gets changed. That is why
the "ideal" is to get some of that glue into the basic
components.

I furthermore believe that there are many ways to
vastly expand the capabilities of clustering AND
improve the performance of such clusters, with minimal
difficulty. There are numerous ways to instrument the
communications (Web100, for example) and many ways to
distribute data reliably to multiple targets.

Some excellent cluster monitoring projects listed on
Freshmeat are DAKOTA and KOJAK. It should be simple
enough to use these to distinguish between application
issues and kernel issues, and therefore to eliminate
inefficiencies within OpenMOSIX.

There are other possibilities and avenues to explore,
which are worthy of discussion IMHO. These are just a
few that - if OpenMOSIX supported them "out of the
box" - could kick Linux clustering into some serious
overdrive.


__________________________________________________
Do You Yahoo!?
Tired of spam?  Yahoo! Mail has the best spam protection around 
http://mail.yahoo.com 

-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share your
opinions on IT & business topics through brief surveys -- and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV