Eh... not really how it works. First, the amount of client requests only maps to the amount of Apache processes running when using the prefork MPM. However, with the exception of fairly nasty virtualization environments (like Virtuozzo), threaded MPMs give you much more efficient resource usage. And even *then*, MPM selection and configuration is a topic worthy of volumes of discussion on its own.
Also, linux is *fantastic* at being deceptive about process memory usage. The %MEM column in top is fairly worthless when judging how much memory Apache is really using... as is pretty much every "default" tool. passenger-memory-stats is pretty good at this though!
And even *then* configuring memory limits is **highly** dependent on your production environment's configuration. If you only have one machine, than probably no one component (webserver, DB, mail, etc) should be allowed to use over half your available memory. Things do get when when you add multiple servers to the mix though
And even **then** all this server configuration milarky is the last
(chronological, not emphatically) thing you as the Rails developer
should be be worrying about. The first step to *any* optimization is
finding where the major bottlenecks in your app are. Newrelic is
indeed a great tool for that. Once you've found the major general
areas to work on, you'll need to break out the actual profiler (ruby-
prof is pretty good, unless you're using Ruby 1.9 
) and actually
get dirty. This is where it is *vital* to have thorough test coverage,
else who knows what you'll mysteriously break when optimizing your
code. (and assuming your code is well structured and understandable to
begin with, don't be afraid of making it ugly to speed things up at
this point... thats why you write tests and "elegant" code in the
beginning).
Gah sooo much more to talk about here though, I guess someone else will contribute more to this thread. At this point I guess I'll just plug Enterprise Rails. Seemingly rushed execution aside, it is THE absolute best Rails-geared book that currently exists in regards to application structure and performance.
