Performance-Monitoring Counters Library, for Intel/AMD Processors and Linux

This example introduces
   rabbit command-line options --mesi m,n --bus_agent m,n --MMX m,n

Return to Main Menu


Command-Line Options

--mesi m,n

  Pentium Pro/II/III and Athlon only
  m, n = 4-bit unsigned integer, default 0xf

  (Intel)
  This option enables or disables, on a bit-wise basis, the counting of four
  L2 cache line events according to the Modified, Exclusive, Shared, Invalid
  protocol.  When used with an input file, the value is applied as a bit-wise
  mask to the unit mask field of each relevant entry in the --event_pairs
  list.

  To see the affected events,
      rabbit -d | grep -e --mesi
      rabbit -g 2 -d
  To see the effect of this option, compare the output of the commands
      rabbit -I
      rabbit -I --mesi 2,4
  The latter selection will only count events affecting cache lines in the
  shared (2) and exclusive (4) states, in counters 0 and 1 respectively.

  (AMD)
  Four counters for three events are available, and the protocol is MOESI
  (Modified, Owner, Exclusive, Shared, Invalid) with a default of 0x1f.

--bus_agent m,n

  Pentium Pro/II/III only
  m, n = 0 or 1, default 1

  This option sets the unit mask for 14 external bus logic events according to
    0: when this processor causes the bus activity,
    1: when any bus agent causes the bus activity.
  The only addendum to this rule is event 0x62 (98), which counts clocks as
    0: bus clocks,
    1: processor clocks.
  When used with an input file, the value is applied as a bit-wise mask
  (appropriately shifted) to the unit mask field of each relevant entry in the
  --event_pairs list.

  To see the affected events,
      rabbit -d | grep -e --bus_agent
      rabbit -g 3 -d
  To see the effect of this option, compare the output of the commands
      rabbit -I
      rabbit -I --bus_agent 1,0

--MMX m,n

  Pentium II/III only
  m, n = 4-bit unsigned integer, default 0xf

  This option enables or disables, on a bit-wise basis, the counting of five
  events concerning the MMX registers and operations.  For more details, see
  the event descriptions for the MMX execution unit.  When used with an input
  file, the value is applied as a bit-wise mask to the unit mask field of
  each relevant entry in the --event_pairs list.

  Pentium III only

  This option enables or disables, on a bit-wise basis, the counting of four
  events concerning the Streaming SIMD Extensions operations and prefetch
  instructions.

  To see the affected events,
      rabbit -d | grep -e --MMX
      rabbit -g 7 -d
      rabbit -g 8 -d
  To see the effect of this option, compare the output of the commands
      rabbit -I
      rabbit -I --MMX 2,4


Example

  Two of the events using --mesi are 41 and 42, L2 data loads and stores.
  A brute-force test is easy to write:

    #!/bin/sh

    for mesi in 0 1 2 3 4 5 6 7 8 9 a b c d e f
    do
      rabbit -s 1 --e 41,42 --m 0x$mesi foo
    done

  On a single-processor system only the odd values of $mesi give non-zero
  event counts.