NAME
perf-top - System profiling tool.SYNOPSIS
perf top [-e <EVENT> | --event=EVENT] [<options>]
DESCRIPTION
This command generates and displays a performance counter profile in real time.OPTIONS
-a, --all-cpusSystem-wide collection. (default)
Event period to sample.
Monitor only on the list of CPUs provided.
Multiple CPUs can be provided as a comma-separated list with no space: 0,1.
Ranges of CPUs are specified with -: 0-2. Default is to monitor all
CPUS.
Number of seconds to delay between
refreshes.
Select the PMU event. Selection can be a
symbolic event name (use perf list to list all events) or a raw PMU
event in the form of rN where N is a hexadecimal value that represents the raw
register encoding with the layout of the event control registers as described
by entries in /sys/bus/event_source/devices/cpu/format/*.
Display this many functions.
Only display functions with more events than
this.
Put the counters into a counter group.
Sort the output by the event at the index n in
group. If n is invalid, sort by the first event. It can support multiple
groups with different amount of events. WARNING: This should be used on
grouped events.
Profile at this frequency. Use max to
use the currently maximum allowed frequency, i.e. the value in the
kernel.perf_event_max_sample_rate sysctl.
Child tasks do not inherit counters.
Path to vmlinux. Required for annotation
functionality.
Ignore vmlinux files.
kallsyms pathname
Number of mmap data pages (must be a power of
two) or size specification with appended unit character - B/K/M/G. The size is
rounded up to have nearest pages power of two value.
Profile events on existing Process ID (comma
separated list).
Profile events on existing thread ID (comma
separated list).
Record events in threads owned by uid. Name or
number.
Collect data with this RT SCHED_FIFO
priority.
Annotate this symbol.
Hide kernel symbols.
Hide user symbols.
Demangle kernel symbols.
Dump the symbol table used for
profiling.
Be more verbose (show counter open errors,
etc).
Zero history across display updates.
Sort by key(s): pid, comm, dso, symbol,
parent, srcline, weight, local_weight, abort, in_tx, transaction, overhead,
sample, period. Please see description of --sort in the perf-report man
page.
Specify output field - multiple keys can be
specified in CSV format. Following fields are available: overhead,
overhead_sys, overhead_us, overhead_children, sample and period. Also it can
contain any sort key(s).
By default, every sort keys not specified in --field will be appended automatically.
Show a column with the number of
samples.
Show a column with the sum of periods.
Only consider symbols in these dsos. This
option will affect the percentage of the overhead column. See --percentage for
more info.
Only consider symbols in these comms. This
option will affect the percentage of the overhead column. See --percentage for
more info.
Only consider these symbols. This option will
affect the percentage of the overhead column. See --percentage for more
info.
Set disassembler style for objdump.
Remove first N entries from source file path
names in executables and add PREFIX. This allows to display source code
compiled on systems with different file system layout.
Interleave source code with assembly code.
Enabled by default, disable with --no-source.
Show raw instruction encoding of assembly
instructions.
Enables call-graph (stack chain/backtrace)
recording.
Setup and enable call-graph (stack
chain/backtrace) recording, implies -g. See --call-graph section in
perf-record and perf-report man pages for details.
Accumulate callchain of children to parent
entry so that then can show up in the output. The output will have a new
"Children" column and will be sorted on the data. It requires
-g/--call-graph option enabled. See the ‘overhead calculation’
section for more details. Enabled by default, disable with
--no-children.
Set the stack depth limit when parsing the
callchain, anything beyond the specified depth will be ignored. This is a
trade-off between information loss and faster processing especially for
workloads that can have a very long callchain stack.
Default: /proc/sys/kernel/perf_event_max_stack when present, 127 otherwise.
Ignore callees of the function(s) matching the
given regex. This has the effect of collecting the callers of each such
function into one place in the call-graph tree.
Do not show entries which have an overhead
under that percent. (Default: 0).
Determine how to display the overhead
percentage of filtered entries. Filters can be applied by --comms, --dsos
and/or --symbols options and Zoom operations on the TUI (thread, dso, etc).
"relative" means it's relative to filtered entries only so that the sum of shown entries will be always 100%. "absolute" means it retains the original value before and after the filter is applied.
Force each column width to the provided list,
for large terminal readability. 0 means no limit (default behavior).
When processing pre-existing threads
/proc/XXX/mmap, it may take a long time, because the file may be huge. A time
out is needed in such cases. This option sets the time out limit. The default
value is 500 ms.
Enable taken branch stack sampling. Any type
of taken branch may be sampled. This is a shortcut for --branch-filter any.
See --branch-filter for more infos.
Enable taken branch stack sampling. Each
sample captures a series of consecutive taken branches. The number of branches
captured with each sample depends on the underlying hardware, the type of
branches of interest, and the executed code. It is possible to select the
types of branches captured by enabling filters. For a full list of modifiers
please see the perf record manpage.
The option requires at least one branch type among any, any_call, any_ret, ind_call, cond. The privilege levels may be omitted, in which case, the privilege levels of the associated event are applied to the branch filter. Both kernel (k) and hypervisor (hv) privilege levels are subject to permissions. When sampling on multiple events, branch stack sampling is enabled for all the sampling events. The sampled branch type is the same for all events. The various filters must be specified as a comma separated list: --branch-filter any_ret,u,k Note that this feature may not be available on all processors.
When displaying traceevent output, do not use
print fmt or plugins.
Enable hierarchy output.
Enable this to use just the most recent
records, which helps in high core count machines such as Knights Landing/Mill,
but right now is disabled by default as the pausing used in this technique is
leading to loss of metadata events such as PERF_RECORD_MMAP which makes
perf top unable to resolve samples, leading to lots of unknown samples
appearing on the UI. Enable this if you are in such machines and profiling a
workload that doesn’t creates short lived threads and/or doesn’t
uses many executable mmap operations. Work is being planed to solve this
situation, till then, this will remain disabled by default.
Don’t do ownership validation.
The number of threads to run when synthesizing
events for existing processes. By default, the number of threads equals to the
number of online CPUs.
Record events of type PERF_RECORD_NAMESPACES
and display it with the cgroup_id sort key.
monitor only in the container (cgroup) called
"name". This option is available only in per-cpu mode. The cgroup
filesystem must be mounted. All threads belonging to container
"name" are monitored when they run on the monitored CPUs. Multiple
cgroups can be provided. Each cgroup is applied to the corresponding event,
i.e., first cgroup to first event, second cgroup to second event and so on. It
is possible to provide an empty cgroup (monitor all the time) using, e.g., -G
foo,,bar. Cgroups must have corresponding events, i.e., they always refer to
events defined earlier on the command line. If the user wants to track
multiple events for a specific cgroup, the user can use -e e1 -e e2 -G
foo,foo or just use -e e1 -e e2 -G foo.
Record events of type PERF_RECORD_CGROUP and
display it with the cgroup sort key.
Only consider events after this event is
found.
E.g.:
Find out where broadcast packets are handled
perf probe -L icmp_rcv
Insert a probe there:
perf probe icmp_rcv:59
Start perf top and ask it to only consider the cycles events when a broadcast packet arrives This will show a menu with two entries and will start counting when a broadcast packet arrives:
perf top -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
Alternatively one can ask for --group and then two overhead columns will appear, the first for cycles and the second for the switch-on event.
perf top --group -e cycles,probe:icmp_rcv --switch-on=probe:icmp_rcv
This may be interesting to measure a workload only after some initialization phase is over, i.e. insert a perf probe at that point and use the above examples replacing probe:icmp_rcv with the just-after-init probe.
Stop considering events after this event is
found.
Show the --switch-on/off events too. This has
no effect in perf top now but probably we’ll make the default
not to show the switch-on/off events on the --group mode and if there is only
one event besides the off/on ones, go straight to the histogram browser, just
like perf top with no events explicitly specified does.
Show callgraph with stitched LBRs, which may
have more complete callgraph. The option must be used with --call-graph lbr
recording. Disabled by default. In common cases with call stack overflows, it
can recreate better call stacks than the default lbr call stack output. But
this approach is not full proof. There can be cases where it creates incorrect
call stacks from incorrect matches. The known limitations include exception
handing such as setjmp/longjmp will have calls/returns not match.
INTERACTIVE PROMPTING KEYS
[d]Display refresh delay.
Number of entries to display.
Event to display when multiple counters are
active.
Profile display filter (>= hit
count).
Annotation display filter (>= % of
total).
Annotate symbol.
Stop annotation, return to full profile
display.
Hide kernel symbols.
Hide user symbols.
Toggle event count zeroing across display
updates.
Quit.
OVERHEAD CALCULATION
The overhead can be shown in two columns as Children and Self when perf collects callchains. The self overhead is simply calculated by adding all period values of the entry - usually a function (symbol). This is the value that perf shows traditionally and sum of all the self overhead values should be 100%.void foo(void) { /* do something */ } void bar(void) { /* do something */ foo(); } int main(void) { bar() return 0; }
Overhead Symbol ........ ..................... 60.00% foo | --- foo bar main __libc_start_main 40.00% bar | --- bar main __libc_start_main
Children Self Symbol ........ ........ .................... 100.00% 0.00% __libc_start_main | --- __libc_start_main 100.00% 0.00% main | --- main __libc_start_main 100.00% 40.00% bar | --- bar main __libc_start_main 60.00% 60.00% foo | --- foo bar main __libc_start_main
SEE ALSO
perf-stat(1), perf-list(1), perf-report(1)2024-06-21 | perf |