[Linux-HA] High Availibility and LoadBalancing

Nicola Ranaldo ranaldo at unina.it
Thu Mar 10 04:38:24 MST 2005


Let's suppose i have an heavy loaded "A" service (for example incoming mail 
with antivirus/antispam) and i want to load balance it over 4 servers.
A simple configuration is 1 frontend (linux virtual server) and 4 backend 
all of them running "A" and storing locally the data (in the examples the 
own mail queue).

Now Let's suppose i want to implement High Availibility over them.
The first thing to do is to cluster the frontend over an active/passive 
heartbeat cluster.
The second is to cluster the 4 backend with other active/passive heartbeat 
cluster and sharing data (the mail queue) over a SAN or with DRBD.
In this case i waste 4 slave server.

The best for me *would be* to have a multinode cluster with 4 active 
servers:  if one of them fails, one of the remaining takes it's services and 
resources.
But I suppose running two istance of the same service "A", on the same 
server could be an heavy configuration problem (if you do not have other 
simpler and transparent solutions :), and i do not want to risk with 
extermely customized config file (in the example, i cannot have 2 sendmail 
binding the same port, of course i can instruct sendmail in order to process 
2 queues -the original, and the queue coming from the failed server-, but 
it's difficult and dangerous).
So the best for me *is* to have a multinode cluster with 4 active servers 
and 1 slave (or more).

Is it possible to implement the above scenario with heartbeat (pheraps in 
the version 2) ?
Can i say to heartbeat to stop the failed service when no more slaves are 
available in order to avoid services collisions?

In a multinode cluster configuration i suppose i cannot use cross-over 
cables and nullmodem cables.
Wath may i do? Need i use 2 local switch (couse 1 of them could fail!) and 
so 3 ethernet cards on every server?

Again, i think i cannot use DRDB to replicate data between more of two 
servers. Need i to use only a SAN?

Finally. Is this scenario a good architectural pattern? Are there some 
alternatives?

Sorry for English, Best Regards

    Nicola Ranaldo
    Federico II  University - Naples (IT) 




More information about the Linux-HA mailing list