[216] in athena10
Re: Cluster machine reset idea
daemon@ATHENA.MIT.EDU (Kenneth Arnold)
Wed May 28 20:02:45 2008
Message-ID: <483DF274.7090504@mit.edu>
Date: Wed, 28 May 2008 20:01:56 -0400
From: Kenneth Arnold <kcarnold@MIT.EDU>
MIME-Version: 1.0
To: Timothy G Abbott <tabbott@mit.edu>
CC: athena10@mit.edu
In-Reply-To: <Pine.LNX.4.64L.0804301714350.29779@mega-man.mit.edu>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
Here's an even more drastic idea:
The cluster machines run virtual Athena 10 and WinAthena images
side-by-side. (Ctrl-Alt-F1 for Linux, Ctrl-Alt-F2 for Windows). When you
log off of either, they just return to a saved state. Software upgrades
are then handled by rsyncing the hard disk image (clone, rsync, kill VM,
replace old image, start VM).
And dreaming even more... it would be cool if users could easily move a
session from one computer to another. I think VMware does that on
servers side, but I don't see an obvious solution on their site for
doing it with desktops.
The VMs could be VMware (they'd be happy, but I don't know if it has the
interfaces we'd need -- basically a "reset-me-now" guest OS op) or
something else; do cluster machines have hardware-assisted virtualization?
I do like the LVM idea too; maybe Linux could run native with that. I
think I even saw support for LVM snapshots in schroot, for whatever that
means. Tim wins in simplicity.
-Ken
Timothy G Abbott wrote:
> One problem that we will probably experience with running Debian-based
> cluster machines is that users will su to root and then apt-get
> install some packages containing programs that they want to run for
> that session. The cluster maintainance code would then have to be
> responsible for removing any such packages cleanly.
>
> I thought of the idea of having (most of) the filesystem tree that you
> see when you login graphically be a chroot containing an LVM snapshot
> of the actual Athena source filesystem, which is then destroyed when
> you log out. Directories that want to survive past the user logging
> out, like /home, /tmp, various parts of /var, etc. would be
> bind-mounted from the source filesystem, and thus preserved when users
> log out.
>
> I would not intend this to be a security measure, but instead a
> mechanism for making it difficult for users to accidentally
> reconfigure cluster machines.
>
> I'm not convinced that this idea doesn't have serious problems, but
> some variation on it might be a good way to support temporarily
> installing software on cluster machines using apt.
>
> -Tim Abbott
