man pages section 1M: System Administration Commands

Exit Print View

Updated: July 2014



cpustat - monitor system behavior using CPU performance counters


cpustat -c eventspec [-c eventspec]... [-p period] [-T u | d ]
     [-Dmnst] [-A cor|soc|bins] [-k keys] [-o limit]
     [-I statfile] [-O statfile] [interval [count]]
cpustat -h


The cpustat utility allows CPU performance counters to be used to monitor the overall behavior of the CPUs in the system.

If interval is specified, cpustat samples activity every interval seconds, repeating forever. If a count is specified, the statistics are repeated count times. If neither are specified, an interval of five seconds is used, and there is no limit to the number of samples that are taken.


The following options are supported:

–A cor

Aggregate output by core ID. Data rows having the same core ID are aggregated into one row. The columns are replaced with subtotals, by default. The –m option prints column averages, instead.

–A soc

Aggregate output by socket ID. Data rows having the same socket ID are aggregated into one row. The columns are replaced with subtotals, by default. The –m option prints column averages, instead.

–A bins

Aggregate the rows into a lesser number of bins within each sampling period, grouping them in the order in which they appear, and print the columnar subtotal over rows for each bin. The –m option may be used in order to compute the arithmetic mean instead of the subtotal. The –k sorting option may be used to change the row order prior to the binning step. The sze column prints the number of CPUs in each bin. The BIN column replaces the CPU column and prints the ordinal of each bin.

–c eventspec

Specifies a set of events for the CPU performance counters to monitor. The syntax of these event specifications is:


You can use the –h option to obtain a list of available events and attributes. This causes generation of the usage message. You can omit an explicit counter assignment, in which case cpustat attempts to choose a capable counter automatically.

Attribute values can be expressed in hexadecimal, octal, or decimal notation, in a format suitable for strtoll(3C). An attribute present in the event specification without an explicit value receives a default value of 1. An attribute without a corresponding counter number is applied to all counters in the specification.

The semantics of these event specifications can be determined by reading the CPU manufacturer's documentation for the events.

Multiple –c options can be specified, in which case the command cycles between the different event settings on each sample.


Enables debug mode.


Prints an extensive help message on how to use the utility and how to program the processor-dependent counters.

–I statfile

Replay data previously saved in statfile. Create data files for replay by specifying –O. This option is especially useful for analyzing statistics on machines with large numbers of CPUs. The file may be reprocessed multiple times using different sorting and aggregation options.

The –I option is incompatible with an interval and count specification.

Read from the standard input if the file name is (hyphen).

–k key1,...

Sort rows within each sampling period from highest to lowest by key1, then key2, and so on. Each key is a comma-separated list of events. There may be multiple –k options specified.

When cpustat is run with multiple –c event-spec options it produces a report of alternating event-specs. Specify multiple –k options to sort each event-spec differently. For each event-spec, the first –k option whose keys contain a proper subset of the events in the event-spec is used.


Print the arithmetic mean value rather than the sum when the –b or –i is used to aggregate data over multiple CPUs.


Omits all header output (useful if cpustat is the beginning of a pipeline).

–o num

Print only the first num rows within each sampling period, after applying sorting and aggregation options.

–O statfile

Save all data to statfile. This data may be replayed at a later time using –I.

Write to the standard output if the file name is (hyphen).

The purpose of –O is to capture all available data. It is incompatible with the data reduction options: –A, –k, –m and –o.

–p period

Causes cpustat to cycle through the list of eventspecs every period seconds. The tool sleeps after each cycle until period seconds have elapsed since the first eventspec was measured.

When this option is present, the optional count parameter specifies the number of total cycles to make (instead of the number of total samples to take). If period is less than the number of eventspecs times interval, the tool acts as it period is 0.


Creates an idle soaker thread to spin while system-only eventspecs are bound. One idle soaker thread is bound to each CPU in the current processor set. System-only eventspecs contain both the nouser and the sys tokens and measure events that occur while the CPU is operating in privileged mode. This option prevents the kernel's idle loop from running and triggering system-mode events.

–T u | d

Display a time stamp.

Specify u for a printed representation of the internal representation of time. See time(2). Specify d for standard date format. See date(1).


Prints an additional column of processor cycle counts, if available on the current architecture.


A closely related utility, cputrack(1), can be used to monitor the behavior of individual applications with little or no interference from other activities on the system.

The cpustat utility must be run by the super-user, as there is an intrinsic conflict between the use of the CPU performance counters system-wide by cpustat and the use of the CPU performance counters to monitor an individual process (for example, by cputrack.)

Once any instance of this utility has started, no further per-process or per-LWP use of the counters is allowed until the last instance of the utility terminates.

The times printed by the command correspond to the wallclock time when the hardware counters were actually sampled, instead of when the program told the kernel to sample them. The time is derived from the same timebase as gethrtime(3C).

The processor cycle counts enabled by the –t option always apply to both user and system modes, regardless of the settings applied to the performance counter registers.

On some hardware platforms running in system mode using the “sys” token, the counters are implemented using 32-bit registers. While the kernel attempts to catch all overflows to synthesize 64-bit counters, because of hardware implementation restrictions, overflows can be lost unless the sampling interval is kept short enough. The events most prone to wrap are those that count processor clock cycles. If such an event is of interest, sampling should occur frequently so that less than 4 billion clock cycles can occur between samples.

The output of cpustat is designed to be readily parseable by nawk(1) and perl (1), thereby allowing performance tools to be composed by embedding cpustat in scripts. Alternatively, tools can be constructed directly using the same APIs that cpustat is built upon using the facilities of libcpc(3LIB). See cpc(3CPC) .

The cpustat utility only monitors the CPUs that are accessible to it in the current processor set. Thus, several instances of the utility can be running on the CPUs in different processor sets. See psrset(1M) for more information about processor sets.

Because cpustat uses LWPs bound to CPUs, the utility might have to be terminated before the configuration of the relevant processor can be changed.



Example 1 Measuring External Cache References and Misses

The following example measures misses and references in the external cache. These occur while the processor is operating in user mode on an UltraSPARC machine.

example% cpustat -c EC_ref,EC_misses 1 3

    time cpu event      pic0      pic1
   1.008   0  tick     69284      1647
   1.008   1  tick     43284      1175
   2.008   0  tick    179576      1834
   2.008   1  tick    202022     12046
   3.008   0  tick     93262       384
   3.008   1  tick     63649      1118
   3.008   2 total    651077     18204


Example 2 Measuring Branch Prediction Success on Pentium 4

The following example measures branch mispredictions and total branch instructions in user and system mode on a Pentium 4 machine.

 example% cpustat -c \
    emask14=0xf,sys 1 3
    time cpu event      pic12     pic14
   1.010   1  tick       458       684 
   1.010   0  tick       305       511 
   2.010   0  tick       181       269 
   2.010   1  tick       469       684 
   3.010   0  tick       182       269 
   3.010   1  tick       468       684 
   3.010   2 total      2063      3101 

Example 3 Counting Memory Accesses on Opteron

The following example determines the number of memory accesses made through each memory controller on an Opteron, broken down by internal memory latency:

cpustat -c \
   pic0=NB_mem_ctrlr_page_access,umask0=0x01, \
   pic1=NB_mem_ctrlr_page_access,umask1=0x02, \
   pic2=NB_mem_ctrlr_page_access,umask2=0x04,sys \

    time cpu event      pic0      pic1      pic2
   1.003   0  tick     41976     53519      7720
   1.003   1  tick      5589     19402       731
   2.003   1  tick      6011     17005       658
   2.003   0  tick     43944     45473      7338
   3.003   1  tick      7105     20177       762
   3.003   0  tick     47045     48025      7119
   4.003   0  tick     43224     46296      6694
   4.003   1  tick      5366     19114       652
Example 4 Displaying Multiple CPUs with a Filter

The following command displays the three CPUs with the highest DTLB_miss rate.

example% cpustat -c DTLB_miss -k DTLB_miss -n 3 1 1

 time cpu event DTLB_miss
1.040 115  tick       107
1.006  18  tick        98
1.045 126  tick        31
1.046  96 total       236

event DTLB_miss
total       236
Example 5 Aggregating Multiple CPUs into Quartiles by a Filter

The following command aggregates 256 CPUs into quartiles by DTLB miss rate.

example% cpustat -c DTLB_miss -b 4 -k DTLB_miss -m 1 1

 time bin event DTLB_miss sze
1.032   0  tick        46  24
1.021   1  tick         3  24
1.007   2  tick         2  24
1.022   3  tick         0  24
1.045   4 total        51  24

event DTLB_miss
total        51
Example 6 Sorting Multiple Events

The following sequence of commands sorts multiple events.

example% cpustat -O /tmp/OUT -c ITLB_miss,DTLB_miss -c PAPI_tot_ins 1 2
example% cpustat  -I /tmp/OUT -b 4 -k ITLB_miss -k PAPI_tot_ins

 time bin event ITLB_miss DTLB_miss sze
1.020   0  tick       129       673  24
1.009   1  tick         0        61  24
1.005   2  tick         0        79  24
1.039   3  tick         0        64  24
1.082   4 total       129       877  24

 time bin event PAPI_tot_ins sze
2.073   0  tick        51947  24
2.020   1  tick        14976  24
2.076   2  tick        14976  24
2.004   3  tick        14976  24
2.082   4 total        96875  24

event ITLB_miss DTLB_miss PAPI_tot_ins
total       129       877        96875


By running the cpustat command, the super-user forcibly invalidates all existing performance counter context. This can in turn cause all invocations of the cputrack command, and other users of performance counter context, to exit prematurely with unspecified errors.

If cpustat is invoked on a system that has CPU performance counters which are not supported by Solaris, the following message appears:

cpustat: cannot access performance counters - Operation not applicable

This error message implies that cpc_open() has failed and is documented in cpc_open(3CPC). Review this documentation for more information about the problem and possible solutions.

If a short interval is requested, cpustat might not be able to keep up with the desired sample rate. In this case, some samples might be dropped.


See attributes(5) for descriptions of the following attributes:

Interface Stability

See also

cputrack(1), nawk(1), perl(1), iostat(1M), prstat(1M), psrset(1M), vmstat(1M), cpc(3CPC), cpc_open(3CPC), cpc_bind_cpu(3CPC) , gethrtime(3C), strtoll(3C) , libcpc(3LIB), attributes(5)


When cpustat is run on a Pentium 4 with HyperThreading enabled, a CPC set is bound to only one logical CPU of each physical CPU. See cpc_bind_cpu(3CPC) .