git: HAMMER VFS - Add code to reduce frontend vs flusher locking conflicts

[Matthew Dillon]

commit e2a02b72711fef1968c49644d54e32e1627e4278
Author: Matthew Dillon <dillon at apollo.backplane.com>
Date:   Fri Jul 22 08:55:06 2011 -0700

    HAMMER VFS - Add code to reduce frontend vs flusher locking conflicts
    
    * Implement a Pulse-width modulated time-domain multiplexer on B-Tree
      cursor operations which splits fronttend operations from backend flusher
      operations.
    
      The larger the number of inodes undergoing reclamation, the greater the
      pulse width given over to locking operations initiated by the flusher.
      Frontend operations (typically reads) are given smaller and smaller
      slot widths as the flusher gets more overloaded.
    
      The advantage of this mechanism is that we are not simply imposing a
      delay in the reader, we are imposing a variable-length time slot during
      which the reader is able to grab a B-Tree cursor.  This allows linear
      and partially cached operations to 'burst' many operations within the
      slot, allowing them run at nearly full speed and not imposing an artificial
      performance limitation for linear I/O.
    
    * This solves a major deadlock stalling issue that prevented the flusher
      from being able to actually flush sufficient I/O to keep up with the
      inode backlog that e.g. blogbench creates, causing excessive locking
      stalls throughout the HAMMER filesystem, particularly when many file
      inodes are being cycled through.
    
    * When running blogbench --iterations=150 prior to this change write
      performance would drop to unacceptable levels and read operations
      (particularly things which cycle through inodes like 'find' and 'ls')
      would stall for unacceptably long periods of time, often resulting
      in diagnostic cache_lock messages on the console showing namecache
      blockages in excess of 100 seconds.
    
      With the change namecache blockages are significantly reduced in both
      frequency and duration, find and ls operations are able to run
      concurrently with a heavy multi-file write load, and blogbench shows
      improved write performance while still giving a reasonable priority to
      read operations (which is what we want).
    
    * Rename a few of the sysctls to normalize internal vs external variable
      names.

Summary of changes:
 sys/vfs/hammer/hammer.h         |   11 ++-
 sys/vfs/hammer/hammer_cursor.c  |  166 +++++++++++++++++++++++++++++++++++++-
 sys/vfs/hammer/hammer_flusher.c |  117 +++++++++++----------------
 sys/vfs/hammer/hammer_inode.c   |   36 +++++----
 sys/vfs/hammer/hammer_vfsops.c  |   31 ++++++--
 5 files changed, 259 insertions(+), 102 deletions(-)

http://gitweb.dragonflybsd.org/dragonfly.git/commitdiff/e2a02b72711fef1968c49644d54e32e1627e4278


-- 
DragonFly BSD source repository