I read yesterday that linux passes a "swap token" around to allow the
system to recover when a lot of swapping's going on.  If we had
something like this, we'd be a lot closer to marking yamt's clock-pro
implementation non-experimental (the only significant problem I've
encountered with it is that there's a tendency for "hot" file pages to
consume a majority of available memory to the detriment of "cooler"
exec/anon pages which really shouldn't be paged out).  For more
information, search the mailing list archives for my posts requesting
testers for PDPOLICY_CLOCKPRO. Also, this is a good starting point:

<http://linux-mm.org/SwapTokenTuning>

I'm interested in what those more experienced than I have to say.

Regards,

--Blair