Clustering already highly available servers

[John Ryland]

Some modern UNIX servers are claimed to be four nines reliable (99.99% -IBM P series) I.e highly available. If I cluster trwo together what improvements in reliability/availability will i get? I presume I will not get mainframe levels (5 x 9's), but can't seem to find out the exact maths to use.

[Peter van Eijk]

The basis statistics goes like this. If the failure modes are completely independent, you can multiply the probabilities for downtime.

So if one server is down 10% of the time, then the probability of two being down at the same time would be 10% times 10% = 1 %, giving an uptime of 99%.

In your case that would lead to an uptime of 99.999999 %, although it is very unlikely that the failure modes are completely independent. Presumably these servers are close to each other, so natural disasters alone are a common failure mode.