Some results from mapbench on josmp (with DEBUG 0 and VM_RADIX):

When TLB shootdowns are disabled (comment out tlbflush in vmap::remove),
mapbench is perfectly scalable (perhaps a little slow, but it's hard to
tell, because I didn't used to run with readaccess=1):

  $ time mapbench 1
  12403435727 cycles
  $ time mapbench 2
  12439895511 cycles
  $ time mapbench 16
  12200449501 cycles

With TLB shootdowns, things are slightly less scalable, and significantly
slower:

  $ time mapbench 1
  40044138778 cycles
  $ time mapbench 2
  44034668320 cycles
  $ time mapbench 16
  57569203944 cycles

Lazy TLB shootdowns improve performance on 1 core, and have some but
not nearly perfect scalability:

  $ time mapbench 1
  12658404885 cycles
  $ time mapbench 2
  29790759720 cycles
  $ time mapbench 16
  39160961925 cycles

Next step is to understand why TLB shootdowns are so expensive and how
to make them cheaper/scalable.