Which distribution do I use to simulate "random" program crashes?

I want to test a distributed program resistance to random crashes of each node. I need to use some random distribution that would control how often each node would crash - for example, it might be normal dis开发者_运维技巧tribution with an average of 2 hours or something else.

How do I choose that distribution law and how do I decide what values to assign to the distribution parameters?

As Rosh Oxymoron points out, it would be odd if the arrival times of crashes were anything but independent. However, that makes me think that the correct way to model this behavior is as a Poisson process. That means your arrival times would be based on the exponential distribution. The rate parameter of the distribution is something you probably want to come up with empirically by looking at data of your current crash rate.

It would be very strange if the behaviour of the program has a dependence on the specific time of crash. You'd have different outcome depending on what the program is doing when it crashes, but different code points occur at random times, so you have a little chance of hitting something way too often. Just avoid fixed intervals to avoid coincidence with fixed timers, gaussian with average of 2 hours would do, uniform of 1 to 3 hours would probably still do, it shouldn't matter too much. I'd personally make it uniform to make sure that I have an equal chance of hitting a fixed timer or something.