NAME

SYNOPSIS

pfmon [OPTION] [PROGNAME]

DESCRIPTION

With pfmon, it is possible to monitor a single thread or the entire system. It is also possible to monitor multi-process and multi-threaded programs. For each, it is possible to collect simple counts or profiles.

The set of events that can be measured depends on the underlying processor. Similarly certain options are specific to a processor model. In general pfmon gives access to all processor-specific monitoring features.

GENERIC OPTIONS

-h or --help

display list of available options and exit

-V or --version

print pfmon version information and exit

-l[regex] or --show-events[=regex]

If regex is not provided, pfmon lists the names of all available events for the current processor. Otherwise only the events matching the regular expression are printed.

-L or --long-show-events[=regex]

If regex is not provided, pfmon lists all available events for the current processor with all their unit masks using the event_name:unit_mask_name notation. Only one unit mask per line is printed, thus multiple lines may be printed. If regex is provided, it is applied only on the event name and not on the unit masks. All event names matching the pattern are printed.

-i event or --event-info=event

Display detailed information about an event. The event parameter can either be the event code, the event name, or a regular expression. In case multiple events match the expression, they are all printed.

-u, -3, or --user-level

Monitor at the user level for all events. By default, this option is turned on.

-k, -0, or --kernel-level

Monitor at the kernel level for all events. By default, this option is turned off.

-1

Monitor execution at privilege level 1. By default, this option is turned off.

-2

Monitor execution at privilege level 2. By default, this option is turned off.

-e ev1,ev2,... or --events=ev1,ev2,...

Select events to monitor. The events are specified by name or event code. If there are multiple events, they must be passed as a comma separated list without spaces. The maximum number of events depends on the underlying processors. Events requiring unit mask can be specified using the notation: event_name:unit_mask1:unit_mask2... Each unit mask can be specified by its name or its numerical value. Pfmon also supports passing combination of unit masks as a single numerical value. For instance, if event A supports unit masks M1 (0x1) and M2 (0x40), and both unit masks are to be measured, then the following event specifications are valid: "A:M1:M2", "A:M1:0x40", "A:M2:0x1", "A:0x1:0x40", "A:0x41".

Each -e option forms a set of events, multiple sets can be defined by specifying the -e option multiple times. Events related options always apply to the last defined sets. All events from a set are measured together. Pfmon uses the perfmon interface to multiplex the sets on the actual processor. In case multiple sets are used, pfmon scales the final counts to provide estimates of what the actual counts would have been had all the events been measured throughout the entire run. Pfmon does not re-arrange events between sets in case they cannot be measured together. It is not possible to use multiple sets when sampling.

-I or --info

Print information related to the pfmon version, the supported processor models and built-in sampling modules.

-t secs or --session-timeout=secs

Duration of the monitoring session expressed in seconds. Once the timeout expires, pfmon stops monitoring and prints final counts or profiles.

-S format or --smpl-module-info=format

Display information about a sampling module.

--debug

Enable debug output (for experts).

--verbose

Print more information about the execution of pfmon.

--outfile=filename

Print final counts in the file called filename. By default, all results (count or profiles) are printed on the terminal.

--append

Append results (counts or profile) to the current output file. If --outfile or --smpl-outfile are not provided results are printed on the screen.

--overflow-block

Block the monitored thread when the sampling buffer becomes full. This option is only available in per-thread mode. By default, this option is turned off meaning that the monitored thread keeps on running, with monitoring disabled, while pfmon is processing the sampling buffer. In other words, there may be blind spots. Note that this option may not work with thread that rely on signals.

--system-wide

Create a system wide monitoring session where pfmon measures all threads running on a set of processors. By default this option is turned off, i.e., pfmon operates in per-thread mode. By default, system-wide mode measures the same events on all available processors. Available processors may depend on the CPU-set pfmon is executed in. It is possible to restrict to a subset of processor using the --cpu-list option.

--smpl-outfile=filename

Save profiles into the file called filename. By default, profiles are printed on the terminal.

--long-smpl-periods=val1,val2,...

Set the long sampling period to reload into the overflowed counter(s) after a buffer full notification. The values must be passed in the same order as the events they refer to. For instance, if the events are passed as -eev1,ev2 then the sampling periods for ev1 must be the first, followed by the period for ev2. It is possible to skip a period, by providing an empty element in the list, e.g., --long-smpl-periods=,val2. Sampling periods are expressed in the same unit as the event they refer to. If an event counts the number of instructions retired, then the sampling period is using the same unit, i.e., instructions retired. To sample every 100,000 instructions, you can pass --long-smpl-periods=100000. If this option is not set but --short-smpl-periods is set, then the short reset values are used for both periods and vice-versa.

--short-smpl-periods=val1,val2,...

Set the short sampling to reload into the overflowed counter(s) after a sample is recorded into the buffer and the buffer does not become full. The option functions exactly like --long-smpl-periods. If this option is not set but --long-smpl-periods is set, then the long reset values are used for both periods and vice-versa. is

--smpl-entries=n

Selects the number of samples that the kernel sampling buffer can hold. The default size is determined dynamically by pfmon based on the size of a sample and system resource limits such as the amount of locked memory allowed for a user process (as reported by ulimit). When this resource limit is set to unlimited then pfmon uses 2048 entries.

--with-header

Generates a header before printing counts or profiles. The header contains information about the configuration of the host system and about the measurement.

--cpu-list=num,num1-num2,...

For system-wide mode, this option specifies the list of processors to monitor. Without this option, all available processors are monitored. Processors can be specified individually with their index, or by range. Pfmon takes into account CPU-sets, so cpu identification must be relative to the set.

--aggregate-results

Aggregate counts and profiles output. By default, this option is off meaning that results are per-thread or per-CPU. This option may not be supported by all sampling modules.

--trigger-code-start-address=addr

Start monitoring the first time code executes at address addr. The address can be specified in hexadecimal or with a symbolic name. This option is not supported in system-wide mode.

--trigger-code-stop-address=addr

Stop monitoring the first time code executes at address addr. The address can be specified in hexadecimal or with a symbolic name. This option is not supported in system-wide mode.

--trigger-data-start-address=addr

Start monitoring when the data address at address addr is accessed. By default, this is for any read or write access. This option may not be supported on all processors. This option is not supported in system-wide mode.

--trigger-data-stop-address=addr

Stop monitoring when data address at address addr is accessed. By default, this is for any read of write access. This option may not be supported on all processors. This option is not supported in system-wide mode.

--trigger-code-repeat

By default, the start and stop code triggers are activated only the first time they are reached. With this option, it is possible to repeat the start/stop behavior each time the execution crosses the trigger address. This option is not supported in system-wide mode.

--trigger-code-follow

Apply the start/stop code triggers to all monitored threads. By default, triggers are only applied to the first thread. This option has no effect on system-wide measurements.

--trigger-data-repeat

By default, the start and stop data triggers are activated only the first time they are reached. With this option, it is possible to repeat the start/stop behavior each time the data address is accessed. This option has no effect on system-wide measurements.

--trigger-data-follow

Apply the start/stop data triggers to all monitored threads. By default, triggers are only applied to the first thread. This option has no effect on system-wide measurements.

--trigger-data-ro

Data trigger are activated on read access only. By default, they are activated on read or write access. This option may not be supported on all processors. This option is not supported in system-wide mode.

--trigger-data-wo

Data trigger activated on write access only. By default, they are activated on read or write access. This option may not be supported on all processors. This option is not supported in system-wide mode.

--trigger-start-delay=secs

Number of seconds before activating monitoring. By default, monitoring is activated immediately, except when code/data triggers are used.

--priv-levels=lvl1,lvl2,...

Set privilege level per event. The levels apply to the current set, i.e. the last -e option. The levels are specified in the same order as the events. Accepted values for privileges are: u, k, 0, 1, 2, 3 or any combinations thereof.

--us-counter-format

Print counts using commas, e.g., 1,024.

--eu-counter-format

Print count using points, e.g., 1.024.

--hex-counter-format

Print count using hexadecimal, e.g., 0x400.

--smpl-module=name

Select the sampling module. By default the first module that matches the PMU model is used. This is typically the detailed-* module. To figure out which modules are supports, use the -I option.

--show-time

Show real,user, and system time for the command executed in per-thread mode.

--symbol-file=filename

ELF image containing the symbol table for the command being monitored. By default, pfmon uses the binary image on disk.

--check-events-only

Verify combination of events and exit. No measurement is performed.

--smpl-periods-random=mask1:seed1,...

Apply randomization to long and short periods. For each period, a seed and a mask value must be passed. The mask is a bitmask representing the range of variation for randomization. As of perfmon v2.3, the seed value is now ignored.

--smpl-print-counts

When sampling, the final counts for the counters are not printed by default. This option forces counts to be printed at the end of a sampling measurement.

--attach-task pid

Attach to thread identified by pid that is already running. User must have permission to attach to the thread. The pid really refers to the kernel thread identification (tid). When attaching to a multi-threaded program, only the designated thread is monitored, unless the --follow-pthread is also specified. In that case, all threads will be monitored, along with any newly created thread thereafter.

--reset-non-smpl-periods[=a,b-c,..]

At the end of a sampling period, reset PMD registers non used as sampling periods. When no parameter is passed to this option, all non sampling PMD registers are reset. Otherwise, the list of specified registers is reset. For instance, if the list is 0,4-6, then registers 0, 4, 5, 6 will be reset when any counter overflows.

--follow-fork

Monitoring continues across fork(). By default monitoring is not propagated to child processes. This option has no effect in system-wide mode.

--follow-vfork

Monitoring continues across vfork(). By default monitoring is not propagated to child processes. This option has no effect in system-wide mode.

--follow-pthread

Monitoring continues across pthread_create(). When attaching to a thread in a multi-threaded process, only the designated thread is monitored. However, if this option is specified with --attach-task, all other threads will also be monitored along with any newly created thread thereafter. This option has no effect in system-wide mode.

--follow-exec[=pattern]

Monitoring follows through the exec*() system call. By default monitoring stops at exec*(). It is possible to specify a regular expression pattern to filter out which command gets monitored. Without the pattern all commands are monitored.

--follow-exec-exclude=pattern

Monitoring follows through the exec*() system call. By default monitoring stops at exec*(). This option is the counter-part of --follow-exec in that the pattern specifies the command which must be excluded from monitoring. Depending on the monitored workload, it may be easier to specify the commands to excludes rather than the commands to include.

--follow-all

This option is equivalent to specifying all of --follow-fork, --follow-vfork, --follow-pthreads, --follow-exec.

--no-cmd-output

Redirect all output of executed commands to /dev/null.

--exec-split-results

Generate separate results output for execution before and after exec*().

--resolve-addresses

Resolve all code/data addresses in profiles using symbol table information. If the symbol information is not present, the raw address is printed. By default, only raw addresses are printed.

--extra-smpl-pmds=num,num1-num2,...

Specify a list of extra PMD registers to include in samples. Those PMD registers are typically non counting PMD registers.

--saturate-smpl-buffer

Stop collecting samples the first time the sampling buffer becomes full. In other words, simply collect the first N entries when --smpl-entries=N. By default, this option is off.

--pin-command

Pin executed command on the CPUs specified by --cpu-list. This option is only relavant in system-wide mode.

--switch-timeout=milliseconds

The number of milliseconds before switching from one event set to the next. Depending on the granularity of the underlying operating system timer, the timeout may be rounded up. If the difference with the user provided timeout exeeds 2%, pfmon prints a warning message.

--dont-start

Do not activate monitoring. This option is useful on architectures where it is possible to start/stop counters directly from the user level.

--cpu-set-relative

With this option, CPU identifications for --cpu--list are relative to CPU-set affinity. By default, they are relative to actual CPU0.

--print-interval=msecs

With this option, intermediate results can be generated when counting in a system-wide session. Pfmon prints the delta for each event since the last print. The interval is expressed in milliseconds. This option is not supported in per-thread mode. No total counts are printed at the end.

--smpl-per-function

For sampling modules which produce an histogram, aggregate samples per function as opposed to per sample address which is the default.

--smpl-ignore-pids

In system-wide sampling, there is currently no symbol correlation possible. However, many formats do print the program name and thread identification for each sample. With other formats, e.g., pebs, such output is not possible and to ensure users are aware of the problem, pfmon fails unless this option is passed. It is ignored in all other situations.

--smpl-show-top=n

When sampling, only show the top n samples. The default is to show all samples. This option can be combined with --smpl-cum-thres, in which case, output stops when one of the two limits is reached.

--smpl-cum-thres=p

When sampling, only show the samples up to the point where the cumulative number of samples reaches p percent. The p argument must be between 1 and 100. By default all samples are shown. This option can be combined with --smpl-show-top in which case, output stops when one of the two limits is reached.

--smpl-eager-save

By default when sampling in per-thread mode, pfmon waits until the last thread terminates before processing the samples to produce the final profiles. The motivation is to avoid introducing noise while the measurement is running. However, this can incur high pressure on system resources such as memory and file descriptors. This option causes profiles to be generated when a thread session terminates, thereby potentially minimizing the amount of system resources used.

SPECIAL CONTROLS

AUTHOR