There are many caches in the kernel. Most of these have knobs and adjustments, some exposed and some not, for sizing and writeback rate and flush behavior and assorted other voodoo, and most of the ones that aren't adjustable probably should be.

Currently all or nearly all of these caches operate on autopilot independent of the others, which does not necessarily produce good results, especially if the system is operating in a performance regime different from when the behavior was tuned by the implementors.

It would be nice if all these caches were instead coordinated, so that they don't end up fighting with one another. Integrated control of sizing, for example, would allow explicitly maintaining a sensible balance between different memory uses based on current conditions; right now you might get that, depending on whether the available voodoo happens to work adequately under the workload you have, or you might not. Also, it is probably possible to define some simple rules about eviction, like not evicting vnodes that have UVM pages still to be written out, that can help avoid unnecessary thrashing and other adverse dynamic behavior. And similarly, it is probably possible to prefetch some caches based on activity in others. It might even be possible to come up with one glorious unified cache management algorithm.

Also note that cache eviction and prefetching is fundamentally a form of scheduling, so all of this material should also be integrated with the process scheduler to allow it to make more informed decisions.

This is a nontrivial undertaking.

Step 1 is to just find all the things in the kernel that ought to participate in a coordinated caching and scheduling scheme. This should not take all that long. Some examples include: * UVM pages * file system metadata buffers * VFS name cache * vnode cache * size of the mbuf pool

Step 2 is to restructure and connect things up so that it is readily possible to get the necessary information from all the random places in the kernel that these things occupy, without making a horrible mess and without trashing system performance in the process or deadlocking out the wazoo. This is not going to be particularly easy or fast.

Step 3 is to take some simple steps, like suggested above, to do something useful with the coordinated information, and hopefully to show via benchmarks that it has some benefit.

Step 4 is to look into more elaborate algorithms for unified control of everything. The previous version of this project cited IBM's ARC ("Adaptive Replacement Cache") as one thing to look at. (But note that ARC may be encumbered -- someone please check on that and update this page.) Another possibility is to deploy machine learning algorithms to look for and exploit patterns.

Note: this is a serious research project. Step 3 will yield a publishable minor paper; step 4 will yield a publishable major paper if you manage to come up with something that works, and it quite possibly contains enough material for a PhD thesis.

Add a comment