21264A Counter Configurations -- A DCPI "one pager" / PJD / 29 November 2001

In addition to ProfileMe, the Alpha 21264A can use its performance counters
in "aggregate" mode. Aggregate mode means that the counters just count the
occurence of user-selected events and the event counts are associated with
specific instructions by PC-sampling. The 21264A can sample and measure
four events:

  Event       Description
  -----       -----------
  cycles      Processor cycles
  retires     Retired instructions
  replaytrap  Memory box (Mbox) replay traps
  bmiss       L2 B-cache misses (or long-latency probes)

Aggregate mode and the specific events to measure must be specified when
starting the DCPI daemon. The events are specified using the -slot option,
which selects a set of event types to monitor simultaneously using the set
of available hardware performance counters. You may measure these events
individually:
  dcpid -slot cycles $DCPIDB
or in the following pair-wise combinations:
  dcpid -slot cycles+retires $DCPIDB
  dcpid -slot retires+bmiss $DCPIDB
  dcpid -slot retires+replaytrap $DCPIDB
Other combinations are supported by "time-multiplexing" (that is, periodically
reconfiguring) the performance counters. Time-multiplexing permits a large
set of events to be monitored, but with a different sampling period, since the
counter measuring the events must be shared. The -slot option may be repeated
to specify a sequence of slots which are time-multiplexed onto the hardware
counters. The command:
  dcpid -slot bmiss -slot replaytrap $DCPIDB
collects both L2 B-cache miss and replay trap events.

ProfileMe mode supports alternative counter configurations, too. The command:
  dcpid -slot pm $DCPIDB
collects ProfileMe samples and measures both inflight cycles and retire delay.
This is the default counter configuration. As shown in the table below, DCPI
and ProfileMe supports four different counter configurations.

  Configuration  Counter 0        Counter 1
  -------------  ---------------  --------------------------
      pm0        Retires          Inflight cycles
      pm         Inflight cycles  Retire delay
      pm2        Retires          L2 B-cache misses
      pm3        Inflight cycles  Memory system replay traps

For example, you can use the command:
  dcpid -slot pm2 $DCPIDB
to measure both the number of retired instructions and L2 B-cache misses. The
dcpiprof and dcpilist command line syntax uses ProfileMe counter names to
modify event names. For example, use the command:
  dcpiprof -pm ret+ret:bcmisses app_prog
to display the number of retired instruction samples and the number of L2 cache
misses taken by retired instructions. Here are the DCPI counter names:

  Counter name  Description
  ------------  ---------------------------------------------
  retdelay      Instruction retire delay
  inflight      Number of cycles the instruction was inflight
  bcmisses      Number of B-cache (L2) cache misses
  replays       Number of memory system replay traps

Remember, you must explicitly enable and collect data for certain events in
order to display the data later!