Hi Community!
for those who do not know yet, I recently was invited
from the PlanetLab folks to pay a visit to Princeton,
which I did, and besides the fact that it was a lot of
fun, we managed to have a bunch of very productive
discussions ...
I'm going to mention some of the ideas (or solutions)
which will make it into Linux-VServer (or already made
it into the devel branch) sooner or later.
(not all of them are the direct result of my visit,
but I guess it doesn't really matter)
- I/O schedulers come in 4 flavours (noop, deadline,
antcipatory and complete fair queuing (CFQ)), we
decided to focus on the CFQ (which now also supports
priorities) and made that per context (in devel)
- the CPU scheduler will get an overhaul and will
become a hybrid Fair Scheduling partially Work
Preserving multi CPU scheduler :)
this will allow you to do things like:
- hand out CPU 'guarantees' (per unit)
- define a fair share independantly
- restrict a context to a certain maximum
- we will continue to develop ngnet and try to make
it work side by side with the current legacy net
(well, an updated version of that, at least)
this will give you:
- a virtual switch/router like setup, similar to
UML or real machines, on the host (which then
basically becomes the router)
- completely isolated loopback and userspace device
support for tunneling and similar
- ipv6 support inside a guest
- guest per interface accounting (also on the host)
- pid and other types of 'isolation' will be extended
to do full virtualization without increasing the
overhead (in cooperation with folks from columbia)
this is a prerequisite to context migration and
snapshoting in a cooperation with folks from the
Columbia university
- Private Namespaces become hierarchical, i.e. they
start to propagate certain events, like mount or
unmount (if desired) down the hierarchy ...
this is a mainline 'feature', but it will become
very interesting for Linux-VServer I guess ...
- We will look into creating a high speed kernel
userspace interface to query/poll/report status
information for graphing all kinds of stats easily
- Dynamic Context support will be removed from the
kernel, and moved into userspace.
- Persistent Contexts (even without processes inside)
are now possible
- Context Setup can be done from inside (SETUP state)
or from the outside
all folks interested in helping with one or the
other sub-project please contact me and give me
some details about your plans ...
all folks interested in testing one or the other
feature, or willing to provide some infrastructure
for the folks going to test/develop that, please
send an email and let us know ...
everybody interested in the features and/or similar
features please follow up to this thread so that
we can prioritize them properly ...
ah, and I almost forgot, if you want to sponsor any
of those developments (which would probably set
some priority and allow us to get more work done),
please let us know ...
thanks a lot,
Herbert
_______________________________________________
Vserver mailing list
Vserver@list.linux-vserver.org
http://list.linux-vserver.org/mailman/listinfo/vserver
Received on Wed Dec 21 15:25:22 2005