Statit
From Siwiki
We often use the [statit command] as a general statistics snapshot.
STATIT(8)
NAME
statit - show cpu, memory, network and IO statistics
SYNOPSIS
statit [ -bhnstxy ] [ -c core_file ] [ -f user_stats_file ] [ -i
interval ] [ -k kernel ] [ -o output_file ] [ -u
binary_output_file ] [ -v binary_input_file ] [ command ]
DESCRIPTION
Statit reads and prints kernel resource utilization statistics
from the memory or core image of a kernel. It may be used to look
at how a program utilizes kernel resources or to print continu-
ously updated statistics about kernel variables.
OPTIONS
The options are:
-b bare mode. Do not print any identifying text. All variables
are printed on one line. At present statit has two output
modes: the default mode, which prints a two column screenful
of statistics and identifying text; and the -b or -h (see
below) modes which print raw data without identifying text
or just a line of description associated with each variable
in the bare mode. Output from the -b mode is in a form
suitable for digestion by a post processing program. There
are presently some bugs in the bare mode output formatting.
-c core_file
specify a core file to read instead of the default in memory
kernel image /dev/kmem.
-f user_stats_file
specify a configuration file which contains a list of vari-
ables that the user is interested in. The variables must be
offsets from a symbol name appearing in the kernel symbol
table. For example, the following lines are used to observe
the characters input to the tty and the cpu user, nice, sys-
tem and idle times:
# TTY statistics
_tk_nin 0 long total tty characters input
# CPU statistics
_cp_time 0 long user time (ticks)
_cp_time 4 long nice time (ticks)
_cp_time 8 long system time (ticks)
_cp_time 12 long idle time (ticks)
The format of the file is as follows: words must start at
the beginning of a line and use a space as a separator. If
the first letter in a line is a pound (#), space or tab,
the rest of the line is ignored. The first word in a line
must be a valid symbol name. The existence of a symbol can
be verified by using the nm(1) command. For example, to
check if the symbol _tk_nin occurs in /vmunix one can use
the command:
example% nm /vmunix | egrep tk_nin
If _tk_nin exists in the symbol table you would get an output like this:
0f099868 B _tk_nin
The second word in user_stats_file must be an offset from
the symbol specified in the first word. In the above exam-
ple _cp_time is really the symbol name of an array declared
as follows in /usr/include/sys/dk.h:
long cp_time[4];
where cp_time[0] contains the user time and cp_time[3] con-
tains cpu idle time. The third word is the C type of the
variable you want to read at the specified offset from the
given symbol. At present only the types
char, double, float, int, long, short and string (max of 7 chars).
are supported. The fourth and subsequent words in a line
are taken to be description lines and are output when the
-fh options are specified.
-h header mode - prints a description of each of the bare mode
variables. In the header mode the description of each vari-
able is output on a separate line. The description is pre-
ceded by the field number of the variable as it appeared in
the bare mode. The statit command to output the headers
should exactly resemble the associated bare mode command
except that the -h option substitutes for the -b option. No
commands are executed by statit in the header mode.
-i interval
interval mode - print the difference in kernel variables
every interval seconds. Start with a snapshot of the kernel
state till now. A command may be given with the -i mode, to
enable command monitoring. If a command is given statit
will exit upon printing the next scheduled output after the
completion of the command. Note that an interval of less
than 3 seconds may cause high cpu overhead. Specifying the
-i mode will make output variables appear in bare mode,
unless the -u flag (described below) is used.
-n Show nfs statistics in the manner of NFSSTAT(8C).
-k kernel
report the statistics of a kernel other than the default
/vmunix.
-o outfile
append statit output to outfile rather than to the default
stdout.
-s snapshot mode. Print kernel statistics up to the present
time since the last boot.
-t Print output as totals rather than the default per second
rate.
-u binary_output_file
Used with the -i flag above to store data in binary form
instead of bare mode. This is much more efficient in disk
space and CPU overhead. Data can be converted back to ascii
form using the -v flag below.
-v binary_input_file
Converts binary data captured using the -u flag to default
ascii mode of presentation. The -u and -v flags may be used
together to monitor a system for extended periods. The com-
mand
statit -ni 10 -u /tmp/binout sleep 3600
This will run the "sleep" command for 1 hour and will take a
snapshot of all system statistics (including NFS statistics)
every 10 seconds during that hour. The data is saved in
binary form in the file /tmp/binout. These statistics can
then be converted to standard statit output format by the
command
statit -nv /tmp/binout
-x binary snapshot mode. Store a binary snapshot of the kernel
statistics in the file /tmp/Xstat. To be used in conjunc-
tion with the -y option below.
-y binary snapshot mode. Take a snapshot of kernel statistics.
Read a set of previously saved statistics from /tmp/Xstat
and then remove it. Compute the difference in the two
snapshots and output the result. The -x and -y flags are
useful when it is inconvenient to fork a command from sta-
tit. In that case the commands to be measured are enclosed
between a statit -x and a statit -y.
NOTES
If a command is specified without the interval mode, statit exe-
cutes the WAIT(2) command and thus uses minimal resources while
the command completes. When a command is specified with the
interval mode, statit will snapshot statistics every interval
seconds (starting with an initial snapshot) until the command
completes. Note that running statit changes the kernel statis-
tics, and a small interval (< 3 seconds) will cause high over-
head.
When the following is typed:
example% statit -o temp1 ls
statit executes the command "ls" and outputs the default set of
statistics to the file temp1. The output generated by "ls" goes,
as usual, to the standard output. The contents of temp1 look
like this:
Hostid: 5100bbd2 Hostname: "firefly" Version: 2.04 Command: ls
Elapsed Time Statistics
1.09 time (seconds) 100.00 % Start time: Thu Mar 14 13:19:08 1991
0.58 user time 53.21 %
0.00 nice time 0.00 %
0.33 system time 30.28 %
0.18 idle time 16.51 %
System Statistics (per second)
55.05 cpu context switches 133.94 traps
291.74 system calls 165.14 device interrupts
0.00 swapins 0.00 swapouts
0.00 pages swapped in 0.00 pages swapped out
0.00 page ins 1.83 page outs
0.00 pages paged in 1.83 pages paged out
3.67 page reclaims from free list 112.84 total page faults taken
0.00 intransit blocking pg flts 3.67 totl pg reclms (includes pgout)
0.00 swap txt pgs fnd on free lst 3.67 inode txt pgs fnd in free lst
82.57 pgs xmnd by the clock daemon 0.00 revs of the clock hand
19.27 pgs zero filled on demand 31.19 pages freed by clock daemon
Hat Statistics (per second)
2.75 ctxt allocations 1.83 ctxts stolen from other as's
133.94 pmg_allocs that have pmg 12.84 pmg allocations
0.00 pmgs stoln frm as's w/no ctx 0.00 pmgs stoln frm as's with ctx
0.00 smg_allocs that have smg 0.00 smg allocations
0.00 smgs stoln frm proc w/no ctx 0.00 smgs stoln frm proc with ctx
3.67 pmgs found on as at ctx_alloc
Virt Addr Cache Flush Statistics (per second)
1.83 context flushes 0.00 segment flushes
207.34 complete page flushes 0.00 partial page flushes
0.00 non-supervisor flushes 0.00 region flushes
Directory Name Cache Statistics (per second)
18.35 cache hits ( 83 %) 3.67 cache misses ( 16 %)
1.83 enters into cache 0.00 enters when already cached
0.00 long names tried to enter 0.00 long names tried to look up
0.00 LRU list empty 0.00 purges of cache
Buffer Cache Statistics (per second)
37.61 total bread requests 37.61 bread hits ( 100 %)
0.00 times aged buf was alloced 0.00 times lru buf was alloced
0.00 times had to sleep for buf
TTY Statistics
0.00 ttyin/sec 4389.91 ttyout/sec
Device: sd0
6.42 % time disk busy 15031.19 bytes/sec
5.02 % time disk seeking 2.75 disk transfers/sec
0.00 disk reads/sec 0.00 seeks/sec
Device: sd1
1.83 % time disk busy 3757.80 bytes/sec
1.49 % time disk seeking 0.92 disk transfers/sec
0.00 disk reads/sec 0.00 seeks/sec
Device: sd3
3.67 % time disk busy 18788.99 bytes/sec
1.92 % time disk seeking 2.75 disk transfers/sec
0.00 disk reads/sec 0.00 seeks/sec
Interface: le0
19.27 input packets/sec 0.00 input errors/sec
7.34 output packets/sec 0.00 output errors/sec
0.00 collisions/sec
Interface: lo0
1.83 input packets/sec 0.00 input errors/sec
1.83 output packets/sec 0.00 output errors/sec
0.00 collisions/sec
Rusage Statistics
0.23 user time 0.22 system time
66 max resident set size (pgs) 0 integral shared text size
2334 integral unshared data size 0 integral unshared stack size
94 page reclaims (no IO) 0 page faults (required IO)
0 swaps 0 block input operations
0 block output operations 0 messages sent over sockets
0 messages recvd from sockets 0 signals received
9 voluntary context switches 0 involuntary context switches
Most of the output of statit shown above should be self explana-
tory. Refer to a Unix architecture book such as "BSD 4.2 Unix"
by Leffler, McKusik, et. al. for detailed explanations of the
meanings of most of these statistics or look at SunOS source.
Interface: le0 refers to the lance ethernet interface # 0 on the
machine. Device: sd0 refers to a SCSI drive, and the statistics
that pertain to it. The Rusage statistics are derived from the
getrusage (2) system call. getrusage (2) returns statistics per-
taining to all the child processes (commands) spawned by statit.
Therefore Rusage Statistics may differ slightly from the statis-
tics above them because the latter include the overhead of fork-
ing off the commands. The block IO statistics reported by
getrusage (2) refer to the number of complete 8K byte blocks
input or output from the file system. In this respect the block
statistics are different from the Device statistics which report
only raw IO numbers.
In the following example:
example% statit -bo temp2 ls
the bare mode output of statit is output in a single line
and stored in temp2. The bare mode output consists of
approximately 90 (the exact number depends on the command
and machine configuration) numbers and words. The first few
numbers output by statit look like this:
590797572.76 50 19 0 13 ...
To make sense of the bare mode numbers it is necessary to
use the header mode as follows:
example% statit -ho temp3 ls
Part of the output is shown below. Each statistic is
printed one to a line with a field number followed by a
short description.
1 start time (seconds)
2 hertz (ticks/sec)
3 user time (ticks)
4 nice time (ticks)
5 system time (ticks)
6 idle time (ticks)
7 Sum_Statistics Header
8 cpu context switches
9 traps
10 system calls
11 device interrupts
12 pseudo dma interrupts VAX only
.
.
.
81 output packets
82 output errors
83 collisions
84 Hostid Header
85 Hostid
86 Hostname Header
87 Hostname
88 Version Header
89 Version Number
90 Command Header
AUTHOR
Varun Mehta
varun@firefly.Eng.Sun.com
Sun Microsystems, Inc.
FILES
/vmunix default kernel object file.
/dev/kmem kernel memory.
/tmp/Xstat temporary statistics stored in binary form (-x option only)
BUGS
There are presently some bugs in the -b and -h modes, I suggest
that they not be used at present.
In cases where the program being measured is very short (under 3
seconds) or statit is called with a short interval its output may
not be accurate. For example, the statistics that relate to vir-
tual memory are updated once every second. Statit would thus
show a large number of zero statistics for programs that take
less than a second to run. Statit causes some overhead of its
own, and in short benchmarks these can be appreciable.
Note that statit output can be misleading due to other reasons.
In many cases kernel or device driver implementations fail to
update statistics that statit reads. Using commands like
vmstat(8) or iostat(8) to verify the performance of statit may
not be useful since statit employs the same mechanism as the
latter commands. Other ways can be devised to verify certain
statistics. For example, the characters input to or output from
ttys can easily be counted. To verify certain disk statistics
use a command such as
statit -o temp dd if=/dev/rsd0a of=/dev/null bs=128w count=100
In this case the dd(1) command will read a 100 blocks of size 512
bytes from the raw I/O device sd0. The output of statit can be
checked to see that a 100 reads were done on sd0 and that 51200
bytes of data were transferred. Similarly the spray(8C) command
can be used to verify ethernet interface statistics by including
it in a command such as
statit -o temp spray remote_host -c 1000 -l 1024
This command will send a 1000 packets of length 1024 bytes from
your host to the remote_host. Note that statit will report a
lower average packet throughput than spray(8C) because statit
measures the extra overhead of forking off spray(8C).
To verify statistics concerning page and swap rates is a little
harder. The onset of swapping, for example, depends on a number
of factors such as the number of contexts available in the MMU
and the number of free pages in physical memory. A program that
does rapid context switching between a large number of processes
would be a good candidate to see if the context switching and
process swapping statistics behave as expected.
Please report all bugs/suggestions to Varun Mehta
(varun@firefly).
SEE ALSO
vmstat(8), iostat(8), netstat(8), nfsstat(8C), pstat(8),
getrusage(2)
Modified 10 July 1989 STATIT(8)
