blob: 4dcddf38159209d1eb23ea2921dc7df020d4cffa [file] [log] [blame]
.TH "numastat" "8" "1.0.0" "Bill Gray" "Administration"
.SH NAME
.LP
\fBnumastat\fP \- Show per-NUMA-node memory statistics for processes and the operating system
.SH "SYNTAX"
.LP
\fBnumastat\fP
.br
.LP
\fBnumastat\fP [\fI\-V\fP]
.br
.LP
\fBnumastat\fP [\fI\<PID>|<pattern>...\fP]
.br
.LP
\fBnumastat\fP [\fI\-c\fP] [\fI\-m\fP] [\fI\-n\fP] [\fI\-p <PID>|<pattern>\fP] [\fI\-s[<node>]\fP] [\fI\-v\fP] [\fI\-z\fP] [\fI\<PID>|<pattern>...\fP]
.br
.SH "DESCRIPTION"
.LP
.B numastat
with no command options or arguments at all, displays per-node NUMA hit and
miss system statistics from the kernel memory allocator. This default
\fBnumastat\fP behavior is strictly compatible with the previous long-standing
\fBnumastat\fP perl script, written by Andi Kleen. The default \fBnumastat\fP
statistics shows per-node numbers (in units of pages of memory) in these categories:
.LP
.B numa_hit
is memory successfully allocated on this node as intended.
.LP
.B numa_miss
is memory allocated on this node despite the process preferring some different node. Each
.I numa_miss
has a
.I numa_foreign
on another node.
.LP
.B numa_foreign
is memory intended for this node, but actually allocated on some different node. Each
.I numa_foreign
has a
.I numa_miss
on another node.
.LP
.B interleave_hit
is interleaved memory successfully allocated on this node as intended.
.LP
.B local_node
is memory allocated on this node while a process was running on it.
.LP
.B other_node
is memory allocated on this node while a process was running on some other node.
.LP
Any supplied options or arguments with the \fBnumastat\fP command will
significantly change both the content and the format of the display. Specified
options will cause display units to change to megabytes of memory, and will
change other specific behaviors of \fBnumastat\fP as described below.
.SH "OPTIONS"
.LP
.TP
\fB\-c\fR
Minimize table display width by dynamically shrinking column widths based on
data contents. With this option, amounts of memory will be rounded to the
nearest megabyte (rather than the usual display with two decimal places).
Column width and inter-column spacing will be somewhat unpredictable with this
option, but the more dense display will be very useful on systems with many
NUMA nodes.
.TP
\fB\-m\fR
Show the meminfo-like system-wide memory usage information. This option
produces a per-node breakdown of memory usage information similar to that found
in /proc/meminfo.
.TP
\fB\-n\fR
Show the original \fBnumastat\fP statistics info. This will show the same
information as the default \fBnumastat\fP behavior but the units will be megabytes of
memory, and there will be other formatting and layout changes versus the
original \fBnumastat\fP behavior.
.TP
\fB\-p\fR <\fBPID\fP> or <\fBpattern\fP>
Show per-node memory allocation information for the specified PID or pattern.
If the \-p argument is only digits, it is assumed to be a numerical PID. If
the argument characters are not only digits, it is assumed to be a text
fragment pattern to search for in process command lines. For example,
\fBnumastat -p qemu\fP will attempt to find and show information for processes
with "qemu" in the command line. Any command line arguments remaining after
\fBnumastat\fP option flag processing is completed, are assumed to be
additional <\fBPID\fP> or <\fBpattern\fP> process specifiers. In this sense,
the \fB\-p\fP option flag is optional: \fBnumastat qemu\fP is equivalent to
\fBnumastat -p qemu\fP
.TP
\fB\-s[<node>]\fR
Sort the table data in descending order before displaying it, so the biggest
memory consumers are listed first. With no specified <node>, the table will be
sorted by the total column. If the optional <node> argument is supplied, the
data will be sorted by the <node> column. Note that <node> must follow the
\fB\-s\fP immediately with no intermediate white space (e.g., \fBnumastat
\-s2\fP). Because \fB\-s\fP can allow an optional argument, it must always be
the last option character in a compound option character string. For example,
instead of \fBnumastat \-msc\fP (which probably will not work as you expect),
use \fBnumastat \-mcs\fP
.TP
\fB\-v\fR
Make some reports more verbose. In particular, process information for
multiple processes will display detailed information for each process.
Normally when per-node information for multiple processes is displayed, only
the total lines are shown.
.TP
\fB\-V\fR
Display \fBnumastat\fP version information and exit.
.TP
\fB\-z\fR
Skip display of table rows and columns of only zero valuess. This can be used
to greatly reduce the amount of uninteresting zero data on systems with many
NUMA nodes. Note that when rows or columns of zeros are still displayed with
this option, that probably means there is at least one value in the row or
column that is actually non-zero, but rounded to zero for display.
.SH NOTES
\fBnumastat\fP attempts to fold each table display so it will be conveniently
readable on the output terminal. Normally a terminal width of 80 characters is
assumed. When the \fBresize\fP command is available, \fBnumastat\fP attempts
to dynamically determine and fine tune the output tty width from \fBresize\fP
output. If \fBnumastat\fP output is not to a tty, very long output lines can
be produced, depending on how many NUMA nodes are present. In all cases,
output width can be explicitly specified via the \fBNUMASTAT_WIDTH\fP
environment variable. For example, \fBNUMASTAT_WIDTH=100 numastat\fP. On
systems with many NUMA nodes, \fBnumastat \-c \-z ....\fP can be very helpful
to selectively reduce the amount of displayed information.
.SH "ENVIRONMENT VARIABLES"
.LP
.TP
NUMASTAT_WIDTH
.SH "FILES"
.LP
\fI/proc/*/numa_maps\fP
.br
\fI/sys/devices/system/node/node*/meminfo\fP
.br
\fI/sys/devices/system/node/node*/numastat\fP
.SH "EXAMPLES"
.I numastat \-c \-z \-m \-n
.br
.I numastat \-czs libvirt kvm qemu
.br
.I watch \-n1 numastat
.br
.I watch \-n1 \-\-differences=cumulative numastat
.SH "AUTHORS"
.LP
The original \fBnumastat\fP perl script was written circa 2003 by Andi Kleen
<andi.kleen@intel.com>. The current \fBnumastat\fP program was written in 2012
by Bill Gray <bgray@redhat.com> to be compatible by default with the original,
and to add options to display per-node system memory usage and per-node process
memory allocation.
.SH "SEE ALSO"
.LP
.BR numactl (8),
.BR set_mempolicy( 2),
.BR numa (3)