What type of workload does your application produce?
Are you mounting and unmounting drives all of the time? That can show the behaviour you're describing... Maybe you don't want the automounter turned on on your server. Maybe you just need to set a longer timeout...
No.
I actually also filed the same bug over at bugzilla.kernel.org and Andrew Morton informs me that it has to do with the SMP/NUMA code and the 4G/4G boundary. By passing a mem=3936MB option to the kernel at boot time, the problem with the filesystem cache getting wiped out goes away.
Bug #2973 if anyone's interested in his explanation (which I don't do justice since I'm not a kernel VM coder).
Philip
