[SATLUG] Next month's presentation

[Jeremy Mann]

> I agree, this would be a very interesting presentation.  Personally my
> company  (Global Gaming Innovations) is working with wide area
> clustering.  My  interest would be peaked and I *might* actually make it
> to my first SATLUG  meeting.

Come out anyway ;)

> That's true that creating a cluster is
> easy, although to be honest,  you DONT need MPI or PVM to do it.  The
> tools are already present in TCP/IP  protocol programing.  Building a
> cluster doesn't have to require installing  or configuring anything!!!
> However, if you wanna get into fun things like multi-threaded parrallel
> computing, you're gonna want to use the kitchen sink of distributed
> parallel  computing (MPI).

This is true with multi-threaded parallel programs. But, what openMosix
does is implement this at kernel level so there is no need to reprogram
your binaries for MPI parameters. I like OpenMosix because it is kernel
level so it load balances regardless of IF the binary is compatible. Now
there are requirements as I mentioned before, but most applications WILL
balance across nodes that need the extra resources. There are tweaks you
can do and even in my experience I can't get *true* balancing (50/50), but
I can get at least 60/40 (60% on the node that started the app, 40 for the
rest of the nodes). There is more but I wanna save it for the demo ;)


> There's one major draw back in today's tech
> that is on the  market.  That is, try starting up your program and pull
> the plug on ANY of  the nodes in the LAM, whops all gone, with no
> signal.  There's ways to send  heartbeats and Duke University is
> developing methods of gracefully shutting  down your distributed program
> with error codes.  However, this is currently a  major pain in the arse.
>  Also for more about wide area distributed processing  check out the
> Albatras (sp?) project.  It's looking pretty good (4  universities
> making up shared clusters).

With OpenMosix you only need an entry in its config file (mosix.map) to
span across networks to access resource nodes.

> Oh and if there is interest, my company is about to release its first
> beta  test (will work natively in linux) of a FPS game that will use the
>  distributed processing tech to create massive multiplayer online games.
> (read  as 500-1000 or more quakster types fragging each other on one
> server). I'll  post an anouncement incase anyone is interested in seeing
> this type of stuff  work.

I am definitely interested in hearing about this. I have heard Doom3 will
P2P resources for extra bandwidth, BUT in my experience with balancing you
need a fat pipe to move the data between nodes.



-- 
Jeremy Mann
jeremy at biochem.uthscsa.edu

University of Texas Health Science Center
Bioinformatics Core Facility
Phone: (210) 567-2672