On Wed, 19.09.12 14:10, Matthew Miller (mattdm(a)fedoraproject.org) wrote:
If any of this sounds interesting, please come on over to the Cloud
SIG,
centered around the mailing list
<
https://admin.fedoraproject.org/mailman/listinfo/cloud>
and pitch in. And if the words "strategic" and "planning" make your
eyes
glaze over, don't worry. This is going to be a *doing* thing, not just a
*talking* thing. Computing as a whole is _really_ at one of those big
inflection points, and it's going to take a lot of action to bring us on
over to the other side — where we've got a *lot* to contribute that the
world shouldn't miss out on.
I think there are two crucial requirements to make Fedora fun in the cloud:
a) We should regularly test Fedora in containers. Containers are a core
feature to allow cloud providers overcommit their resources more
drastically. I spent some time on making systemd boot up relatively
cleanly in an nspawn/libvirt-lxc container (simply because it makes
it easier for me to debug systemd), but that's just systemd and not
the whole distro and right now it's a lot more fun to run systemd in
a Debian container on Fedora than on systemd in a Fedora container on
Fedora (audit...). If Fedora cares for the cloud, then we probably
should declare container boots something that is on the level on a
"release architecture", by which I mean that it is regularly tested
and a requirement for release. This really is an area where some
pressure could be needed.
The systemd test suite actually builds an OS image that is booted
once inside of kvm, and once in nspawn, to make sure we get
everything right in systemd, and that the very same image can be
booted both ways unaltered and entirely stateless. I'd like to see a
similar level of testing in all of Fedora, too.
b) We should try much harder to lower the resource cost per
container. If cloud providers need less disk space, less CPU, and
less memory per container, then this allows them to run much more
containers on one system. And that translates directly to cash for
them. For one this means we should try much
harder to reduce the minimal installation set of packages of
Fedora. And that probably means regularly posting blame lists (whose
packages pull in 100MB more this time? Who is the king of adding new
dependencies?) and having somebody who really invests the work in
splitting up packages to minimize deps. But it also means running
services more often by activation on demand (socket
primarily). i.e. if a cloud provider wants to grant SSH access to
customers for their individual containers he
shouldn't have to run one sshd per container, but should just have
the socket listening. i.e. having 5000 sockets listening on a machine
is relatively cheap. Running 5000 sshd instances on a machine is
quite expensive.
In summary: I believe that containers are a big part of being awesome in
the cloud. And right now running Fedora inside a container is not that
great an experience. It kinda works, but it's very very rough around the
edges.
Yes, I know I should have subscribed to the ML and posted that there,
but I am soooo lazy to do that for one mail only... I apologize.
Lennart
--
Lennart Poettering - Red Hat, Inc.