Diff for /wikisrc/projects/project/improve-caching.mdwn between versions 1.2 and 1.3

version 1.2, 2011/11/06 21:08:23 version 1.3, 2014/02/27 08:04:16
Line 1 Line 1
 [[!template id=project  [[!template id=project
 title="Improve caching"  title="Coordinated caching and scheduling"
 contact="""  contact="""
 [tech-kern](mailto:tech-kern@NetBSD.org)  [tech-kern](mailto:tech-kern@NetBSD.org)
Line 8  contact=""" Line 8  contact="""
 category="kernel"  category="kernel"
 difficulty="hard"  difficulty="hard"
 duration="4 months"  duration="4 months and up"
 description="""  description="""
 Replace all of the knobs and wacky algorithms for sizing the many separate LRU  
 caches in our kernel (for pages and for various kinds of objects like metadata  There are many caches in the kernel. Most of these have knobs and
 buffers, mbufs, etc.)  with a generalized (N-way) algorithm like IBM's ARC  adjustments, some exposed and some not, for sizing and writeback rate
 ("Adaptive Replacement Cache") to size all the caches for optimal hit 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
   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.
 """  """
 ]]  ]]

Removed from v.1.2  
changed lines
  Added in v.1.3

CVSweb for NetBSD wikisrc <wikimaster@NetBSD.org> software: FreeBSD-CVSweb