I want to offer UML as a service, including fail-over etcetera. The idea is
that people will be able to get more functionality (running on RAID, fail-over
of your node if a machine breaks, etcetera) for less money than a phyiscal
host. Some of our bigger competitors are doing this with Linux mainframes, but
we'll have to make do with Intel boxen for the time being :-).
The basic idea is to have an UML instance pull as much as possible from
remote filestores, where my inventory of remote filestores contains
OpenAFS, NFS, and DRBD in various configurations. Heartbeat
should make sure that a UML node neatly fails over. UML instances would run on
pizzaboxen, and the files/filesystems all reside on 'real servers' (as in
Xeon, Raid, redundant power supplies, etcetera).
All in all there's a bewildering amount of setups possible: NFS root, DRBD
root, locally installed root with daily rsync; mount OpenAFS from within UML
or underneath UML (the first necessitates an OpenAFS port to UML); etcetera.
Due to the, err, 'distinctive' ;) performance characteristics of UML they are
likely to have wildly differing performance figures as well.
What would be a reasonable starter configuration for a setup like this? I am
more interested in flexibility (growing filesystems) and maintainability (I'm
looking at initially some 10 UML setups, hopefully growing fast) than ultimate
speed (at the moment, we pull about 2Mbit bandwidth for our whole server
TIA for any thoughts you want to share on this.
Cees de Groot http://www.cdegroot.com <cg@...>
GnuPG 1024D/E0989E8B 0016 F679 F38D 5946 4ECD 1986 F303 937F E098 9E8B
Get latest updates about Open Source Projects, Conferences and News.