[Planetlab-devel] symetric peering

[David E. Eisenstat]

On Wed, 27 Jun 2007, Thierry Parmentelat wrote:

> - a couple other slices use the nm-controller instantiation method, that 
> our code does not currently support

We've made several changes to Node Manager/util-vserver recently that I'd 
like to summarize here. We've either pushed or are about to push these.

- The nm-controller instantiation method is used to inform Node Manager 
that the slice is a special account with shell forward_api_calls that will 
be used to make calls to Node Manager. Previously we had hard-coded an 
account for Emulab, the only user of this feature to date.

- We patched util-vserver to prevent slices from getting wedged in such a 
way that causes pl_chcontext() to fail (thus locking everyone out). This 
was happening regularly on the production PlanetLab.

- We also patched util-vserver not to blow away /var/run when a slice is 
started, which was causing multiple instances of crond and other daemons 
to appear.

- I finally noticed that Node Manager was double-starting some slices and 
changed the logic so that that will not happen.

- Node Manager now re-runs the slice initscripts when they change. This 
was at the request of the Stork project, so that users who opt to use 
Stork after their slices have been instantiated don't have to run the 
initscript manually.

- Instead of redirecting stdout/stderr to /dev/null, Node Manager 
redirects these to a file in /root so that we can tell why instances 
sometimes die.

Also, we're planning to add Sun's Java runtime and "which" to the VServer 
reference image.

Regards,

-David Eisenstat