Easy question first: The munin server just spits out a bunch of HTML and PNG files every 5 minutes, which can then be statically served by a web server of your choice. Easy breezy. The parts that interface with the web server are statistics-gathering plugins on the munin client (aka munin-node); obviously, if you don't run Apache, you won't be using the Apache plugins
shameless plug:
http://github.com/rtucker/munin-lighttpd
Harder question: The strange thing about Xen is
there is no host. On the old UML platform, there was a clear and obvious delineation between the "guests" (Linodes) and the "host" (the thing running the ~40 Linodes as individual processes). The closest we get to that is "dom0", which is, essentially, a guest with some additional privileges above and beyond the normal "domU" guests.
In practice, the dom0 instance is the "gatekeeper" between the domU instances and the hardware, specifically the disks and the network interfaces. This is from where the Linode platform does all of the stuff it needs to do, like launching Linodes, setting up virtual network interfaces, managing disk I/O priorities, etc.
Unlike the UML platform, however, the dom0 doesn't have the guests "within" it. Under UML, the host load meter was a decent metric of what was going on for the whole host, since each Linode was its own process and thus impacted the load average, showed in "top" and "ps", etc. With Xen, the host load meter just looks at the dom0's performance, which is pinned at "idle" unless there's a lot of platform-related activity going on (deploys, backups, mikegrb's seti@home clients, etc).
In other words, anything you can do from within your Linode is unlikely to get that needle above "idle", so monitoring it will be really boring.
And yup, /proc/loadavg is just your Linode's load average. Remember, your kernel has almost no idea it's virtualized. Heck, with the newer/experimental paravirtualized kernels, it's so realistic your clock will drift unless you run ntpd. (Weird, huh?)