文件操作 - howto.cpuperf.html

文件: /usr/share/doc/pcp-doc/html/howto.cpuperf.html

编辑文件内容

<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN">

<HTML>
<HEAD>
	<meta http-equiv="content-type" content="text/html; charset=utf-8">
	<meta http-equiv="content-style-type" content="text/css">
	<link href="pcpdoc.css" rel="stylesheet" type="text/css">
	<link href="images/pcp.ico" rel="icon" type="image/ico">
	<TITLE>Understanding system-level processor performance</TITLE>
</HEAD>
<BODY LANG="en-AU" TEXT="#000060" DIR="LTR">
<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0 STYLE="page-break-before: always">
	<TR> <TD WIDTH=64 HEIGHT=64><A HREF="https://pcp.io/"><IMG SRC="images/pcpicon.png" ALT="pmcharticon" ALIGN=TOP WIDTH=64 HEIGHT=64 BORDER=0></A></TD>
	<TD WIDTH=1>&nbsp;&nbsp;&nbsp;&nbsp;</TD>
	<TD WIDTH=500><A HREF="index.html">Home</A>&nbsp;&nbsp;&middot;&nbsp;<A HREF="lab.pmchart.html">Charts</A>&nbsp;&nbsp;&middot;&nbsp;<A HREF="timecontrol.html">Time Control</A></TD>
	</TR>
</TABLE>
<H1 ALIGN=CENTER STYLE="margin-top: 0.48cm; margin-bottom: 0.32cm">Understanding measures of system-level processor performance</H1>
<TABLE WIDTH="15%" BORDER=0 CELLPADDING=5 CELLSPACING=10 ALIGN=RIGHT>
	<TR><TD BGCOLOR="#e2e2e2"><PRE><IMG SRC="images/system-search.png" ALT="" WIDTH=16 HEIGHT=16 BORDER=0>&nbsp;&nbsp;Tools 
sar
mpstat
pmchart
pcp-atop
pcp-atopsar
pcp-uptime
</PRE></TD></TR>
</TABLE>
This chapter of the Performance Co-Pilot tutorial provides guidance
on how to interpret and understand the various measures of system-level 
processor (CPU) performance.
All modern operating systems collect processor resource utilization at both
the process-level and the system-level.&nbsp;&nbsp;This tutorial
relates specifically to the system-level metrics.
For an explanation of Performance Co-Pilot terms and acronyms, consult 
the <A HREF="glossary.html">PCP glossary</A>

<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0 BGCOLOR="#e2e2e2">
 <TR><TD WIDTH="100%" BGCOLOR="#081c59">How the system-level CPU time is computed</TD></TR>
</TABLE>

Utilities like mpstat, sar and Performance Co-Pilot (PCP) use a
common collection of system-level CPU performance instrumentation from the kernel. 
This instrumentation is based upon statistical sampling of the state of each
CPU in the kernel's software clock interrupt routine which is commonly called
100 times (PCP metric <TT>kernel.all.hz</TT>) per second on every CPU.

At each observation a CPU is attributed a quantum of 10 milliseconds of 
elapsed time to one of several counters based on the current state of 
the code executing on that CPU.

This sort of statistical sampling is subject to some anomalies, 
particularly when activity is strongly correlated with the clock 
interrupts, however the distribution of observations over several 
seconds or minutes is often an accurate reflection of the true 
distribution of CPU time. The kernel profiling mechanisms offer higher
resolution should that be required, however that is beyond the scope
of this document.

The CPU state is determined by considering what the CPU was doing just 
before the clock interrupt, as follows:
<OL>
 <LI>
 If executing a userspace task (i.e. above the kernel system call
 interface for some program) then
 <UL>
 <LI>
 If executing with modified scheduling priority, the state is
 nice.
 <LI>
 Otherwise the state is user.
 </UL>
 <LI>
 If servicing an interrupt, then the state is irq.
 <LI>
 If this is a virtual CPU being made to wait involuntarily for a real
 CPU (i.e. while the hypervisor services another virtual processor),
 the state is steal. (Linux only)
 <LI>
 If executing a hypervisor guest (Linux only), then
 <UL>
 <LI>
 If executing with modified scheduling priority, the state is
 both nice and guest_nice.
 <LI>
 Otherwise the state is both user and guest.
 <LI>
 If otherwise executing code in the kernel (i.e. below the system call
 interface, or a kernel task), then the state is kernel. 
 <LI>
 Otherwise if some I/O is pending, then the state is wait.
 <LI>
 Otherwise the state is idle.
</OL>

These states are mutually exclusive and complete, so exactly one state 
is assigned for each CPU at each clock interrupt.

The kernel agent for PCP exports the following metrics:
<CENTER>
<TABLE BORDER="1">
 <CAPTION ALIGN="BOTTOM">Table 1: Raw PCP CPU metrics</CAPTION>
 <TR VALIGN="TOP">
 <TH>PCP Metric</TH>
 <TH>Semantics</TH>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>kernel.all.cpu.user</TT></TD>
 <TD>Time counted when in user state.</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>kernel.all.cpu.nice</TT></TD>
 <TD>Time counted when in nice state. (Linux / UNIX)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>kernel.all.cpu.sys</TT></TD>
 <TD>Time counted when in kernel state.</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>kernel.all.cpu.intr</TT></TD>
 <TD>Time counted when in irq state.</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>kernel.all.cpu.steal</TT></TD>
 <TD>Time counted when in steal state. (Linux only)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>kernel.all.cpu.wait.total</TT></TD>
 <TD>Time counted when in wait state. (Linux / UNIX)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>kernel.all.cpu.idle</TT></TD>
 <TD>Time counted when in idle state.</TD>
 </TR>
</TABLE>
</CENTER>

These metrics are all &quot;counters&quot;, in units of milliseconds 
(cumulative since system boot time).
When displayed with most PCP tools they are &quot;rate converted&quot;
(sampled periodically and the differences between consecutive values 
converted to time utilization in units of milliseconds-per-second over
the sample interval).
Since the raw values are aggregated across all CPUs, the time utilization
for any of the metrics above is in the range 0 to N*1000 for an N CPU system; 
for some PCP tools this is reported as a percentage in the range 0 to N*100
percent.

<TABLE WIDTH="100%" BORDER=0 CELLPADDING=10 CELLSPACING=20>
	<TR><TD BGCOLOR="#e2e2e2" WIDTH="70%"> <IMG SRC="images/stepfwd_on.png" ALT="" WIDTH=16 HEIGHT=16 BORDER=0>&nbsp;&nbsp;&nbsp;Using pmchart to display CPU activity (aggregated over all CPUs). 
<PRE>
$ pmchart -c CPU
</PRE>
<CENTER>
<IMG SRC="images/linux-cpu.png" ALT="" BORDER=0>
</CENTER>
</TD></TR>
</TABLE>

<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0 BGCOLOR="#e2e2e2">
 <TR><TD WIDTH="100%" BGCOLOR="#081c59">Interrupts</TD></TR>
</TABLE>
On Linux, the irq state is further subdivided into components
describing different stages of kernel interrupt handling:
<OL>
 <LI>
 If executing deferred interrupt handling code, the state is softirq.
 <LI>
 Otherwise the state is hardirq.
</OL>

The Linux kernel agent for PCP exports the following CPU interrupt
metrics, the sum of which approximately equals <TT>kernel.all.cpu.intr</TT>:
<CENTER>
<TABLE BORDER="1">
 <CAPTION ALIGN="BOTTOM">Table 2: Raw PCP CPU interrupt metrics</CAPTION>
 <TR VALIGN="TOP">
 <TH>PCP Metric</TH>
 <TH>Semantics</TH>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>kernel.all.cpu.irq.soft</TT></TD>
 <TD>Time counted in softirq state.</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>kernel.all.cpu.irq.hard</TT></TD>
 <TD>Time counted in hardirq state.</TD>
 </TR>
</TABLE>
</CENTER>

<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0 BGCOLOR="#e2e2e2">
 <TR><TD WIDTH="100%" BGCOLOR="#081c59">Virtualization</TD></TR>
</TABLE>

On Linux, several additions to the processor utilization metrics from
other platforms have been made, allowing finer-grained observation of
the effects of virtualization.

In a virtual environment processor cycles are shared amongst virtual
guests on a physical host.
If a guest displays a high steal time, then cycles are being taken
from that guest to serve other purposes.
This indicates one of two possible things:

<UL>
 <LI>
 Either the guest is using more than its share of cycles
 <LI>
 Or the physical server (which may be a hosting provider) is overloaded.
</UL>

As described earlier the user and nice processor states are
reported inclusive of virtualization guest state.
There are PCP metrics allowing these states to be observed separately.
<TABLE WIDTH="100%" BORDER=0 CELLPADDING=10 CELLSPACING=20>
	<TR><TD BGCOLOR="#e2e2e2" WIDTH="70%"> <IMG SRC="images/stepfwd_on.png" ALT="" WIDTH=16 HEIGHT=16 BORDER=0>&nbsp;&nbsp;&nbsp;Using pmchart to display separate guest CPU activity (compared to inclusive). 
<PRE>
$ pmchart -c CPU -c vCPU
</PRE>
<CENTER>
<IMG SRC="images/linux-virt-cpu.png" ALT="" BORDER=0>
</CENTER>
</TD></TR>
</TABLE>

<CENTER>
<TABLE BORDER="1">
 <CAPTION ALIGN="BOTTOM">Table 3: Raw PCP virtualization user time split</CAPTION>
 <TR VALIGN="TOP">
 <TD><TT>kernel.all.cpu.vuser</TT></TD>
 <TD>User time when not running a virtual guest, default scheduling priority.</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>kernel.all.cpu.guest</TT></TD>
 <TD>User time when running a guest, default scheduling priority.</TD>
 </TR>
</TABLE>
</CENTER>


<CENTER>
<TABLE BORDER="1">
 <CAPTION ALIGN="BOTTOM">Table 4: Raw PCP virtualization nice time split</CAPTION>
 <TR VALIGN="TOP">
 <TD><TT>kernel.all.cpu.vnice</TT></TD>
 <TD>User time when not running a virtual guest, modified scheduling priority.</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>kernel.all.cpu.guest_nice</TT></TD>
 <TD>User time when running a guest, modified scheduling priority.</TD>
 </TR>
</TABLE>
</CENTER>

<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0 BGCOLOR="#e2e2e2">
 <TR><TD WIDTH="100%" BGCOLOR="#081c59">The per-CPU variants</TD></TR>
</TABLE>

Inside the kernel, most of the metrics described above are actually
accumulated per-CPU for reasons of efficiency (to reduce the locking 
overheads and minimize dirty cache-line traffic).

PCP exports the per-CPU versions of the system-wide metrics with metric 
names formed by replacing <TT>all</TT> by <TT>percpu</TT>, 
e.g. <TT>kernel.percpu.cpu.user</TT>.

Note that on some multiprocessor UNIX platforms with one I/O pending,
all otherwise idle CPUs will be assigned the wait state.
This may lead to an over-estimate of the I/O wait time, as discussed in the 
companion <A HREF="howto.diskperf.html">How to understand measures of 
disk performance</A> document.
This is not the case on Linux, however, where pre-CPU request queues
track counts of processes in I/O wait state.

The pcp-atop and pcp-atopsar tools can be used to report
the per-CPU utilization metrics, the latter in a fashion similar to the
mpstat -I CPU option.

<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0 BGCOLOR="#e2e2e2">
 <TR><TD WIDTH="100%" BGCOLOR="#081c59">Reconciling mpstat, sar -u and PCP CPU performance metrics</TD></TR>
</TABLE>

The mpstat and sar metrics are scaled based on the number of CPUs
and expressed in percentages, PCP metrics are in units of milliseconds per 
second after rate conversion; this explains the PCP metric <TT>hinv.ncpu</TT>
 and the constants 100 and 1000 in the expressions below.

When run without options, sar reports the following:
<CENTER>
<TABLE BORDER="1">
 <CAPTION ALIGN="BOTTOM">Table 5: PCP and sar (default) metric equivalents</CAPTION>
 <TR>
 <TH>sar metric</TH>
 <TH>PCP equivalent (assuming rate conversion)</TH>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%user</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.user</TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%nice</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.nice</TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%system</TT></TD>
 <TD>100 * (<TT>kernel.all.cpu.sys</TT> + <TT>kernel.all.cpu.intr</TT>) / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%iowait</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.wait.total</TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%steal</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.steal</TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%idle</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.idle</TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
</TABLE>
</CENTER>

On the other hand, mpstat, and sar with the -u ALL option,
report the following:
<CENTER>
<TABLE BORDER="1">
 <CAPTION ALIGN="BOTTOM">Table 6: PCP and mpstat (sar -u ALL) metric equivalents</CAPTION>
 <TR>
 <TH>mpstat metric</TH>
 <TH>PCP equivalent (assuming rate conversion)</TH>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%usr</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.vuser</TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%nice</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.vnice </TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%sys</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.sys</TT> / (<TT>hinv.ncpu</TT> * 
 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%iowait</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.wait.total</TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%steal</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.steal</TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%irq</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.irq.hard</TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%soft</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.irq.soft</TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%guest</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.guest</TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%gnice</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.guest_nice</TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
 <TR VALIGN="TOP">
 <TD><TT>%idle</TT></TD>
 <TD>100 * <TT>kernel.all.cpu.idle</TT> / (<TT>hinv.ncpu</TT> * 1000)</TD>
 </TR>
</TABLE>
</CENTER>

<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0 BGCOLOR="#e2e2e2">
 <TR><TD WIDTH="100%" BGCOLOR="#081c59">The load average</TD></TR>
</TABLE>

The &quot;load average&quot; is reported by pcp-uptime, pcp-atop, etc.
and the PCP metric <TT>kernel.all.load</TT>.

<TABLE WIDTH="100%" BORDER=0 CELLPADDING=10 CELLSPACING=20>
	<TR><TD BGCOLOR="#e2e2e2" WIDTH="70%"> <IMG SRC="images/stepfwd_on.png" ALT="" WIDTH=16 HEIGHT=16 BORDER=0>&nbsp;&nbsp;&nbsp;Using pcp-uptime to display historical uptime and load average measures. 
<PRE>
$ source /etc/pcp.conf
$ tar xzf $PCP_DEMOS_DIR/tutorials/cpuperf.tgz

$ pcp --origin 'Tuesday 4:30pm' --archive cpuperf/system -z uptime
 16:30:07 up 5:15, 11 users, load average: 1.59, 1.54, 1.23
</PRE>
</TD></TR>
</TABLE>

The load average is an indirect measure of the demand for CPU 
resources. It is calculated using the previous load average (load) 
the sum of the currently runnable processes (nr_active), and an
exponential dampening expression, e.g. for the &quot;1 minute&quot; average,
the expression is:
<PRE>
load = exp(-5/60) * load + (1 - exp(-5/60)) * nr_active
</PRE>

The three load averages use different exponential constants and are all 
re-computed every 5 seconds.

On Linux, nr_active is computed as the sum across all processors
(i.e. using per-CPU counters) of the number of running tasks and the number
of uninterruptible tasks.

Note that the &quot;run queue length&quot; (reported by the -q option
of sar) uses the aggregation across all processors of the number of
running tasks only (does not include uninterruptible tasks).
This is the <TT>kernel.all.runnable</TT> PCP metric.

<HR>
<CENTER>
<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0>
	<TR> <TD WIDTH="50%">Copyright &copy; 2007-2010 <A HREF="https://www.aconex.com/">Aconex</A> Copyright &copy; 2000-2004 <A HREF="https://www.sgi.com/">Silicon Graphics Inc</A></TD>
	<TD WIDTH="50%"><A HREF="https://pcp.io/">PCP Site</A> Copyright &copy; 2012-2018 <A HREF="https://www.redhat.com/">Red Hat</A></TD> </TR>
</TABLE>
</CENTER>
</BODY>
</HTML>

文件操作 - howto.cpuperf.html

文件: /usr/share/doc/pcp-doc/html/howto.cpuperf.html

编辑文件内容

修改文件时间

删除文件