[Pipet Devel] node-to-node communication: you can have it 3 ways!

jarl van katwijk jarl at casema.net
Sat Mar 18 19:13:32 EST 2000


>

Hi all,

thnx Jeff for all the flowcharts.. you're really putting a lot of work into it.

> Under what circumstances should 'VSH' switch between using Overflow's protocol
> and GMS's?  Well, if we decide that GMS and Overflow will use DIFFERENT
> protocols to communicate between themselves (i.e., Overflow-to-Overflow and
> GMS-to-GMS), we can find these circumstances.

What does overflow-overflow comm. mean: 1) overflow NODE <-> NODE communication
or 2) overflow SUBNET <-> SUBNET communication?

>
> Perhaps Overflow can use a protocol better for small objects: like CORBA.  And
> GMS could use a protocol better for moving large files: straight TCP/IP.
>

GMS will get a specialised data transport protocal, I'm thinking about libgnet.
(http://www.eecs.umich.edu/~dhelder/misc/gnet/) It's a specialised network
library build on top of glib. So it's portable, fast and secure.

I think the 'overflow layer' should not use corba. That would be too slow and is
highly vurnerable to exploiting. It should have a simple communication mechanism,

corba would be overkill, also because the overflow nodes shouldn't transport data

into nodes on non-local systems. I like to see the overflow layer as nodes that
can do anything INSIDE THEIR SUBNET, they can not communicate trough the
overflow layer to another subnet, the dataflow must pass the gms layer.

I do not care about the numers of nodes, what they do, how or why, but they
should
do it all inside their subnet, or addressing space.
This is critical to me, so I want everybody to comment on this.







More information about the Pipet-Devel mailing list