event counting vs. the cache

[David Young <dyoung%pobox.com@localhost>]

It's customary to use a 64-bit integer to count events in NetBSD because
we don't expect for the count to roll over in the lifetime of a box
running NetBSD.

I've been thinking about what these wide integers do to the cache
footprint of a system and wondering if we shouldn't make a couple of
changes:

1) Cram just as many counters into each cacheline as possible.
   Extend/replace evcnt(9) to allow the caller to provide the storage
   for the integer.

   On a multiprocessor box, you don't want CPUs sharing counter
   cachelines if you can help it, but do cram together each individual
   CPU's counters.

2) Split all counters into two parts: high-order 32 bits, low-order 32
   bits.  It's only necessary to touch the high-order part when the
   low-order part rolls over, so in effect you split the counters into
   write-often (hot) and write-rarely (cold) parts.  Cram together the
   cold parts in cachelines.  Cram together the hot parts in cachelines.
   Only the hot parts change that often, so the ordinary footprint of
   counters in the cache is cut almost in half.

I suppose you could split counters into four or more parts of 16 or
fewer bits each, and in that shrink the footprint even further, but it
seems that you would reach a point of diminishing returns very quickly.

Perhaps this has been tried before and found to (not) work reasonably
well?

Dave

-- 
David Young
dyoung%pobox.com@localhost    Urbana, IL    (217) 721-9981