libpfm_intel_bdx_unc_pcu - support for Intel Broadwell Server
Power Controller Unit (PCU) uncore PMU
#include <perfmon/pfmlib.h>
PMU name: bdx_unc_pcu
PMU desc: Intel Broadwell Server PCU uncore PMU
The library supports the Intel Broadwell Server Power Controller
Unit uncore PMU. This PMU model only exists on various Broadwell server
models (79, 86).
The following modifiers are supported on Intel Broadwell server
PCU uncore PMU:
- e
- Enable edge detection, i.e., count only when there is a state transition
from no occurrence of the event to at least one occurrence. This modifier
must be combined with a threshold modifier (t) with a value greater or
equal to one. This is a boolean modifier.
- t
- Set the threshold value. When set to a non-zero value, the counter counts
the number of HA cycles in which the number of occurrences of the event is
greater or equal to the threshold. This is an integer modifier with values
in the range [0:15].
- i
- Invert the meaning of the threshold or edge filter. If set, the event
counts when strictly less than N occurrences occur per cycle if threshold
is set to N. When invert is set, then threshold must be set to non-zero
value. If set, the event counts when the event transitions from occurring
to not occurring (falling edge) when edge detection is set. This is a
boolean modifier
- ff
- Enable frequency band filtering. This modifier applies only to the
UNC_P_FREQ_BANDx_CYCLES events, where x is [0-3]. The modifiers expects an
integer in the range [0-255]. The value is interpreted as a frequency
value to be multiplied by 100Mhz. Thus if the value is 32, then all cycles
where the processor is running at 3.2GHz and more are counted.
There are 3 events which support frequency band filtering, namely,
UNC_P_FREQ_BAND0_CYCLES, UNC_P_FREQ_BAND1_CYCLES, UNC_P_FREQ_BAND2_CYCLES,
UNC_P_FREQ_BAND3_CYCLES. The frequency filter (available via the ff
modifier) is stored into a PMU shared register which hold all 4 possible
frequency bands, one per event. However, the library generate the encoding
for each event individually because it processes events one at a time. The
caller or the underlying kernel interface may have to merge the band filter
settings to program the filter register properly.
Stephane Eranian <eranian@gmail.com>