[Swift-devel] a note on running coasters on osg's RENCI Engage site

[Mihael Hategan]

On Mon, 2009-01-26 at 12:28 -0500, Mats Rynge wrote:
> Ben Clifford wrote:
> > Coasters in the release don't work on RENCI Engage.
> > 
> > I fiddled with this a bit before, and just fiddled with it a bit more.
> > 
> > The external IP address of the cluster head node (152.54.1.231) is not 
> > accessible from the cluster worker nodes, which sit on a different 
> > network.
> > 
> > The headnode *is* accessible from its IP address on that network, 
> > 192.168.1.11.
> > 
> > Forcing the URI passed to workers to use that IP address instead of the 
> > automatically determined one is sufficient to make coasters work on the 
> > RENCI Engage site.
> 
> I'm surprised that the private address was not used already. I mean,
> isn't this the use case for putting the coaster forwarder on the
> headnode in the first place? If you have outbound network connectivity,
> the coaster daemons in the worker nodes could connect directly to the
> host where swift is running on.

Automatically figuring out which eth should be used is a tricky thing
given that cluster configurations don't seem to follow some
pre-established rule. Each sites does stuff in their own way, though
most seem to have the head node accessible internally through the public
head node address.

There is a performance and reliability advantage to using a "proxy"
instead of going back directly to the submit site. It's easier to
achieve scalability with a hierarchical approach.