>I would expect "down" to not suffer [...]


Don't guess. Far better to go to a library, and look it up (that's
what the libraries are for!), or take an intermediate-level OS class.


For a pragmatic introduction to this particular topic, I would
recommend Harvey M. Dietel, ``Operating System Concepts'', or one of
the several editions of Peterson and Silberschatz' "Dinosaur"
books. Tanenbaum's pre-Minix text, while dated, is pretty good on this
particular stuff.

Those textbooks will point you to the primary literature, which goes
back to the 70s and 0s, if not the 50s.  This stuff is *old*.

I beleive what you will find is pretty much what I said earlier:
C-SCAN was preferred for interactive-database-query-style loads,
as it gives lower variance[*] than SCAN (bidirectional elevator).

Moving from SCAN to C-SCAN will *not* signficantly help Manuel's
problem, because at heart it is a latency problem: pagein-like
requests from interactive processes. C-SCAN would make that latency
more predictable; it wont whack it down by an order of magnitude.

Classes more oriented toward queueing theory, stocahstic modelling, or
OR, where they will assume Poisson arrivals and expect you to
analytically *derive* the mean (expectation for) delay, and variance.
See, e.g., Kleinrock's text on queueing theory and CS applications.
But only if you like maths.