pmcstat
—
performance measurement with performance monitoring
hardware
pmcstat |
[-A ] [-C ]
[-D pathname]
[-E ] [-F
pathname] [-G
pathname] [-I ]
[-L ] [-M
mapfilename] [-N ]
[-O logfilename]
[-P event-spec]
[-R logfilename]
[-S event-spec]
[-T ] [-U ]
[-W ] [-a
pathname] [-c
cpu-spec] [-d ]
[-e ] [-f
pluginopt] [-g ]
[-i lwp]
[-k kerneldir]
[-l secs]
[-m pathname]
[-n rate]
[-o outputfile]
[-p event-spec]
[-q ] [-r
fsroot] [-s
event-spec] [-t
process-spec] [-u
event-spec] [-v ]
[-w secs]
[-z graphdepth]
[command [args]] |
The pmcstat
utility measures system performance using
the facilities provided by
hwpmc(4).
The pmcstat
utility can measure both
hardware events seen by the system as a whole, and those seen when a
specified set of processes are executing on the system's CPUs. If a specific
set of processes is being targeted (for example, if the
-t
process-spec option is
specified, or if a command line is specified using
command), then measurement occurs till
command exits, or till all target processes specified
by the -t
process-spec options
exit, or till the pmcstat
utility is interrupted by
the user. If a specific set of processes is not targeted for measurement,
then pmcstat
will perform system-wide measurements
till interrupted by the user.
A given invocation of pmcstat
can mix
allocations of system-mode and process-mode PMCs, of both counting and
sampling flavors. The values of all counting PMCs are printed in human
readable form at regular intervals by pmcstat
. The
output of sampling PMCs may be configured to go to a log file for subsequent
offline analysis, or, at the expense of greater overhead, may be configured
to be printed in text form on the fly.
Hardware events to measure are specified to
pmcstat
using event specifier strings
event-spec. The syntax of these event specifiers is
machine dependent and is documented in
pmc(3).
A process-mode PMC may be configured to be inheritable by the
target process' current and future children.
The following options are available:
-A
- Skip symbol lookup and display address instead.
-C
- Toggle between showing cumulative or incremental counts for subsequent
counting mode PMCs specified on the command line. The default is to show
incremental counts.
-D
pathname
- Create files with per-program samples in the directory named by
pathname. The default is to create these files in
the current directory.
-E
- Toggle showing per-process counts at the time a tracked process exits for
subsequent process-mode PMCs specified on the command line. This option is
useful for mapping the performance characteristics of a complex pipeline
of processes when used in conjunction with the
-d
option. The default is to not to enable per-process tracking.
-F
pathname
- Print calltree (Kcachegrind) information to file
pathname. If argument pathname
is a “
-
” this information is sent to
the output file specified by the -o
option.
-G
pathname
- Print callchain information to file pathname. If
argument pathname is a
“
-
” this information is sent to the
output file specified by the -o
option.
-I
- Show the offset of the instruction pointer into the symbol.
-L
- List all event names.
-M
mapfilename
- Write the mapping between executable objects encountered in the event log
and the abbreviated pathnames used for
gprof(1)
profiles to file mapfilename. If this option is not
specified, mapping information is not written. Argument
mapfilename may be a
“
-
” in which case this mapping
information is sent to the output file configured by the
-o
option.
-N
- Toggle capturing callchain information for subsequent sampling PMCs. The
default is for sampling PMCs to capture callchain information.
-O
logfilename
- Send logging output to file logfilename. If
logfilename is of the form
hostname:port, where
hostname does not start with a
‘
.
’ or a
‘/
’, then
pmcstat
will open a network socket to host
hostname on port port.
If the -O
option is not specified and
one of the logging options is requested, then
pmcstat
will print a textual form of the logged
events to the configured output file.
-P
event-spec
- Allocate a process mode sampling PMC measuring hardware events specified
in event-spec.
-R
logfilename
- Perform offline analysis using sampling data in file
logfilename.
-S
event-spec
- Allocate a system mode sampling PMC measuring hardware events specified in
event-spec.
-T
- Use a
top(1)-like
mode for sampling PMCs. The following hotkeys can be used:
A
- Toggle symbol resolution
Ctrl+a
- Switch to accumulative mode
Ctrl+d
- Switch to delta mode
f
- Represent the “f” cost under threshold as a dot
(calltree only)
I
- Toggle showing offsets into symbols
m
- Merge PMCs
n
- Change view
p
- Show next PMC
q
- Quit
Space
- Pause
-U
- Toggle capturing user-space call traces while in kernel mode. The default
is for sampling PMCs to capture user-space callchain information while in
user-space mode, and kernel callchain information while in kernel
mode.
-W
- Toggle logging the incremental counts seen by the threads of a tracked
process each time they are scheduled on a CPU. This is an experimental
feature intended to help analyse the dynamic behaviour of processes in the
system. It may incur substantial overhead if enabled. The default is for
this feature to be disabled.
-a
pathname
- Perform a symbol and file:line lookup for each address in each callgraph
and save the output to pathname. Unlike
-m
that only resolves the first symbol in the
graph, this resolves every node in the callgraph, or prints out addresses
if no lookup information is available. This option requires the
-R
option to read in samples that were previously
collected and saved with the -O
option.
-c
cpu-spec
- Set the cpus for subsequent system mode PMCs specified on the command line
to cpu-spec. Argument cpu-spec
is a comma separated list of CPU numbers, or the literal ‘*’
denoting all available CPUs. The default is to allocate system mode PMCs
on all available CPUs.
-d
- Toggle between process mode PMCs measuring events for the target process'
current and future children or only measuring events for the target
process. The default is to measure events for the target process alone.
(it has to be passed in the command line prior to
-p
, -s
,
-P
, or -S
).
-e
- Specify that the gprof profile files will use a wide history counter.
These files are produced in a format compatible with
gprof(1).
However, other tools that cannot fully parse a BSD-style gmon header might
be unable to correctly parse these files.
-f
pluginopt
- Pass option string to the active plugin.
threshold=<float> do not display cost under specified value (Top).
skiplink=0|1 replace node with cost under threshold by a dot (Top).
-g
- Produce profiles in a format compatible with
gprof(1).
A separate profile file is generated for each executable object
encountered. Profile files are placed in sub-directories named by their
PMC event name.
-i
lwp
- Filter on thread ID lwp, which you can get from
ps(1)
-o
lwp
.
-k
kerneldir
- Set the pathname of the kernel directory to argument
kerneldir. This directory specifies where
pmcstat
should look for the kernel and its
modules. The default is to use the path of the running kernel obtained
from the kern.bootfile sysctl. Modules will also be
searched for in /boot/modules if not found in
kerneldir.
-l
secs
- Set system-wide performance measurement duration for
secs seconds. The argument
secs may be a fractional value.
-m
pathname
- Print the sampled PCs with the name, the start and ending addresses of the
function within they live. The pathname argument is
mandatory and indicates where the information will be stored. If argument
pathname is a
“
-
” this information is sent to the
output file specified by the -o
option. This
option requires the -R
option to read in samples
that were previously collected and saved with the
-O
option.
-n
rate
- Set the default sampling rate for subsequent sampling mode PMCs specified
on the command line. The default is to configure PMCs to sample the CPU's
instruction pointer every 65536 events.
-o
outputfile
- Send counter readings and textual representations of logged data to file
outputfile. The default is to send output to
stderr when collecting live data and to
stdout when processing a pre-existing
logfile.
-p
event-spec
- Allocate a process mode counting PMC measuring hardware events specified
in event-spec.
-q
- Decrease verbosity.
-r
fsroot
- Set the top of the filesystem hierarchy under which executables are
located to argument fsroot. The default is
/.
-s
event-spec
- Allocate a system mode counting PMC measuring hardware events specified in
event-spec.
-t
process-spec
- Attach process mode PMCs to the processes named by argument
process-spec. Argument
process-spec may be a non-negative integer denoting
a specific process id, or a regular expression for selecting processes
based on their command names.
-u
event-spec
- Provide short description of event.
-v
- Increase verbosity.
-w
secs
- Print the values of all counting mode PMCs or sampling mode PMCs for top
mode every secs seconds. The argument
secs may be a fractional value. The default interval
is 5 seconds.
-z
graphdepth
- When printing system-wide callgraphs, limit callgraphs to the depth
specified by argument graphdepth.
If command is specified, it is executed
using
execvp(3).
To perform system-wide statistical sampling on an AMD Athlon CPU with samples
taken every 32768 instruction retirals and data being sampled to file
sample.stat, use:
pmcstat -O sample.stat -n 32768 -S
k7-retired-instructions
To execute firefox
and measure the number
of data cache misses suffered by it and its children every 12 seconds on an
AMD Athlon, use:
pmcstat -d -w 12 -p k7-dc-misses
firefox
To measure instructions retired for all processes named
“emacs” use:
pmcstat -t '^emacs$' -p
instructions
To measure instructions retired for processes named
“emacs” for a period of 10 seconds use:
pmcstat -t '^emacs$' -p instructions
sleep 10
To count instruction tlb-misses on CPUs 0 and 2 on a Intel Pentium
Pro/Pentium III SMP system use:
pmcstat -c 0,2 -s
p6-itlb-miss
To collect profiling information for a specific process with pid
1234 based on instruction cache misses seen by it use:
pmcstat -P ic-misses -t 1234 -O
/tmp/sample.out
To perform system-wide sampling on all configured processors based
on processor instructions retired use:
pmcstat -S instructions -O
/tmp/sample.out
If callgraph capture is not desired use:
pmcstat -N -S instructions -O
/tmp/sample.out
To send the generated event log to a remote machine use:
pmcstat -S instructions -O
remotehost:port
On the remote machine, the sample log can be collected using
nc(1):
nc -l remotehost port >
/tmp/sample.out
To generate
gprof(1)
compatible profiles from a sample file use:
pmcstat -R /tmp/sample.out
-g
To print a system-wide profile with callgraphs to file
foo.graph use:
pmcstat -R /tmp/sample.out -G
foo.graph
If option -v
is specified,
pmcstat
may issue the following diagnostic messages:
- #callchain/dubious-frames
- The number of callchain records that had an “impossible”
value for a return address.
- #exec handling errors
- The number of
exec(2)
events in the log file that named executables that could not be
analyzed.
- #exec/elf
- The number of
exec(2)
events that named ELF executables.
- #exec/unknown
- The number of
exec(2)
events that named executables with unrecognized formats.
- #samples/total
- The total number of samples in the log file.
- #samples/unclaimed
- The number of samples that could not be correlated to a known executable
object (i.e., to an executable, shared library, the kernel or the runtime
loader).
- #samples/unknown-object
- The number of samples that were associated with an executable with an
unrecognized object format.
The pmcstat
utility exits 0 on success,
and >0 if an error occurs.
Due to the limitations of the gmon.out file format,
gprof(1)
compatible profiles generated by the -g
option do not
contain information about calls that cross executable boundaries. The
generated gmon.out files are also only meaningful for
native executables.
The pmcstat
utility first appeared in
FreeBSD 6.0. It is currently under development.
The pmcstat
utility cannot yet analyse
hwpmc(4)
logs generated by non-native architectures.