Performance monitoring is “the action of collecting information related to how an application or system performs”. The information can be obtained from the code or the CPU/chipset.

Modern processors contain a performance monitoring unit (PMU). The design and functionality of a PMU is CPU specific: for example, the number of registers, counters and features supported will vary by CPU implementation.

The Perfmon interface is designed to be generic, flexible and extensible. It can monitor at the program (thread) or system levels. In either mode, it is possible to count or sample your profile information. This uniformity makes it easier to write portable tools. Figure 7.1, “Architecture of perfmon2” gives an overview.

Figure 7.1. Architecture of perfmon2¶

Each PMU model consists of a set of registers: the performance monitor configuration (PMC) and the performance monitor data (PMD). Only PMCs is writeable, but both can be read. These registers store configuration information and data.

Perfmon2 supports two modes where you can run your profiling: sampling or couting.

Sampling is usually expressed by an interval of time (time-based) or an occurance of a definied number of events (event-based). Perfmon indirectly supports time-based sampling by using an event-based sample with constant correlation to time (for example, unhalted_reference_cycles.)

In contrast, Counting is expressed in terms of a number of occurances of an event.

Both methods store their information into a sample. This sample contains information about, for example, where a thread was or instruction pointers.

The following example demonstrates the counting of the CPU_OP_CYCLES event and the sampling of this event, generating a sample per 100000 occurances of the event:

pfmon --no-cmd-output -e CPU_OP_CYCLES_ALL /bin/ls
1306604 CPU_OP_CYCLES_ALL

The following command gives the count of a specific function and the procentual amount of the total cycles:

pfmon --no-cmd-output --short-smpl-periods=100000   -e CPU_OP_CYCLES_ALL /bin/ls
# results for [28119:28119<-[28102]] (/bin/ls)
# total samples          : 12
# total buffer overflows : 0
#
#                   event00 
#     counts   %self    %cum          code addr 
           1   8.33%   8.33% 0x2000000000007180 
           1   8.33%  16.67% 0x20000000000195a0 
           1   8.33%  25.00% 0x2000000000019260 
           1   8.33%  33.33% 0x2000000000014e60 
           1   8.33%  41.67% 0x20000000001f38c0 
           1   8.33%  50.00% 0x20000000001ea481 
           1   8.33%  58.33% 0x200000000020b260 
           1   8.33%  66.67% 0x2000000000203490 
           1   8.33%  75.00% 0x2000000000203360 
           1   8.33%  83.33% 0x2000000000203440 
           1   8.33%  91.67% 0x4000000000002690 
           1   8.33% 100.00% 0x20000000001cfdf1

To get a list of supported events, use the option -l from pfmon to list them. Keep in mind, this list depends on the host PMU:

pfmon -l
ALAT_CAPACITY_MISS_ALL
ALAT_CAPACITY_MISS_FP
ALAT_CAPACITY_MISS_INT
BACK_END_BUBBLE_ALL
BACK_END_BUBBLE_FE
BACK_END_BUBBLE_L1D_FPU_RSE
...
CPU_CPL_CHANGES_ALL
CPU_CPL_CHANGES_LVL0
CPU_CPL_CHANGES_LVL1
CPU_CPL_CHANGES_LVL2
CPU_CPL_CHANGES_LVL3
CPU_OP_CYCLES_ALL
CPU_OP_CYCLES_QUAL
CPU_OP_CYCLES_HALTED
DATA_DEBUG_REGISTER_FAULT
DATA_DEBUG_REGISTER_MATCHES
DATA_EAR_ALAT
...

Get an explanation of these entries with the option -i and the event name:

pfmon -i CPU_OP_CYCLES_ALL
Name     : CPU_OP_CYCLES_ALL
Code     : 0x12
Counters : [ 4 5 6 7 8 9 10 11 12 13 14 15 ]
Desc     : CPU Operating Cycles -- All CPU cycles counted
Umask    : 0x0
EAR      : None
ETB      : No
MaxIncr  : 1  (Threshold 0)
Qual     : None
Type     : Causal
Set      : None

Use the --system-wide option to enable monitoring all processes that execute on a specific CPU or sets of CPUs. You do not have to be root to do so; per default, user level is turned on for all events (option -u).

It is possible that one system wide session can run concurrently with another system wide sessions as long as they do not monitor the same set of CPUs. However, you can not run a system wide session with any per-thread sessions together.

The following examples are taken from a Itanium IA64 Montecito processor. To execute a system-wide session, perform the following procedure:

Perfmon can also monitor an existing thread. This is useful for monitoring system daemons or programs which take a long time to start. First determine the process ID you wish to monitor:

ps ax | grep foo
10027  pts/1    R     2:23  foo

Use the found PID for the --attach-task option of pfmon:

pfmon --attach-task=10027
3682190 CPU_OP_CYCLES_ALL