Eric Jennings wrote:
On Mon, 2003-02-10 at 16:28, Jason Lim wrote:
Hi All,
I was wondering if you guys are aware of any solution for multiple
servers
to server 1 domain name?
That is... like those big ISPs that have "user" webhosting.
http://members.isp.com/joe/ (goes to server no. 5)
http://members.isp.com/jane/ (goes to server no. 3)
http://members.isp.com/someone/ (goes to server no. 2)
[...]
Basically the reason for doing this is because the existing single
server
is overloaded, and need to split the workload to 1 or 2
more servers.
Some of the problem might be solved by moving the database to a
dedicated machine. If that doesn't spread the load enough,
doing a DNS
round-robin (having member.isp.com resolve to two different machines
with exactly the same setup) might solve the problem. The web
directories etc. (especially writeable areas for cgi programs) would
need to be shared, with NFS for example (which might create some
locking
problems so you'd need to be careful...).
Of course, this causes requests going to a random server - so PHPs
session tracking ability etc. will have problems. If you use such
things, investigating just upgrading the server first could already
bring relief (where is the performance problem exactly? Disk? CPU?
Memory?)
If you start implementing the server farm architecture, and have the
problem of PHP sessions, SSL sessions, etc... Then you may want to
invest in a hardware web switch, like an F5 or Foundry. Yep, they're
expensive, but super fast and they will hold a particular user's
session to a single server for the duration of that session,
fixing the
problems of the PHP/SSL issues.
Also, they have the added feature of being able to check
heartbeat/pings of each server in the cluster. A particular server
goes down, and the switch will automatically redirect
requests to other
servers, until that broken server comes back online. DNS round robin
will continue to resolve to a broken server, thus making your uptime
availability = numgoodservers / totalservers. Not a good thing if
you're running two servers, and one goes bad. Instant 50%
availability. Furthermore, you can assign weights to each server, so
if you have some old system that you still want to be in the cluster,
you can add them with a lower weight, and they'll get hit
less than the
brand new Dells you just bought (<- that one's for you
Russell Coker. :)
One may also use LVS (http://linuxvirtualserver.org/) for that. It's
free,
and it runs like a charm. There is also connection persistence, one may
choose between different weighting algo's, etc etc. And with keepalived
(http://keepalived.sourceforge.net/) you may set up two LVS directors
in a
high available fashion, and have full control of how and how often the
server's health is being checked (stoopid thigs like tcp connect only,
or
md5-hashes of different pages, or ...).
We have taken this approach, and are very happy with it.
We solved the problem of having only one MySQL server (SPoF) by adding
a
second one, replicating from the main server, but the problem regarding
syncing file systems on the web servers is yet to be resolved. I
dropped the
first idea of using one central file server with NFS - mostly because
of NFS
itself, and because this would be another SPoF. Our tests with coda
were
also stopped, because of (a) problems with lock-ups, and (b) the admin
involved with the tests leaving. For now we stick to one master server
(located on one of the LVS directors), regulary rsync'ing to the web
servers.