/usr/share/doc/pcp-doc/html/lab.containers.html

<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN">
<!--
 (c) Copyright 2015 Red Hat.
 Permission is granted to copy, distribute, and/or modify this document
 under the terms of the Creative Commons Attribution-Share Alike, Version
 3.0 or any later version published by the Creative Commons Corp. A copy
 of the license is available at
 http://creativecommons.org/licenses/by-sa/3.0/us/ .
-->
<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>Analysis of Linux Containers</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><FONT COLOR="#000080"><A HREF="http://pcp.io/"><IMG SRC="images/pcpicon.png" ALT="pmcharticon" ALIGN=TOP WIDTH=64 HEIGHT=64 BORDER=0></A></FONT></TD>
	<TD WIDTH=1><P>&nbsp;&nbsp;&nbsp;&nbsp;</P></TD>
	<TD WIDTH=500><P ALIGN=LEFT><A HREF="index.html"><FONT COLOR="#cc0000">Home</FONT></A>&nbsp;&nbsp;&middot;&nbsp;<A HREF="lab.pmchart.html"><FONT COLOR="#cc0000">Charts</FONT></A>&nbsp;&nbsp;&middot;&nbsp;<A HREF="timecontrol.html"><FONT COLOR="#cc0000">Time Control</FONT></A></P></TD>
	</TR>
</TABLE>
<H1 ALIGN=CENTER STYLE="margin-top: 0.48cm; margin-bottom: 0.32cm"><FONT SIZE=7>Analysis of Linux Containers</FONT></H1>
<P><BR></P>
<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><I>Tools</I>
docker
lxc-info
lxc-ls
pmcd
pmdalinux
pmdaproc
pmdaroot
pminfo
pmprobe</PRE></TD></TR>
</TABLE>
<P>Container engines like Docker, LXC, Rocket and others build on
two Linux kernel facilities - <B>cgroups</B> and <B>namespaces</B>.
In order to understand the performance characteristics of containerized
environments, we need some background kernel knowledge to see how these
concepts affect both the system itself, and system-level analysis tools
like PCP.</P>
<P>This tutorial aims to build or refresh your knowledge in these areas,
while at the same time showing you how to extract performance data from
individual containers using the PCP tools.</P>
<P>For an explanation of Performance Co-Pilot terms and acronyms, consult 
the <A HREF="glossary.html">PCP glossary</A>.</P>
<P><BR></P>
<UL>
    <LI>
    <A HREF="#started">Getting Started</A> 
    <LI>
    <A HREF="#cgroups">Control Groups</A> 
    <LI>
    <A HREF="#namespaces">Namespaces</A> 
    <LI>
    <A HREF="#details">Containers and PCP</A>
    <UL>
        <LI>Core Extensions
        <LI>Elevated Privileges
        <LI>Container-specific Metric Values
    </UL>
    <LI>
    <A HREF="#nextstep">Next Steps</A> 
    <UL>
        <LI>Console Tools
        <LI>Graphical Tools
        <LI>Domain Agents
    </UL>
</UL>
</P>

<P><BR></P>
<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0 BGCOLOR="#e2e2e2">
        <TR><TD WIDTH="100%" BGCOLOR="#081c59"><P ALIGN=LEFT><FONT SIZE=5 COLOR="#ffffff"><B><A NAME="overview">Getting Started</A></B></FONT></P></TD></TR>
</TABLE>
<P>Creating and administering containers is the job of a container runtime
engine, and each operates differently.
If you don't have some handy already, create and start one or two containers
for experimentation - the steps involved are widely documented, and differ
differ between engines, so we will skip over them here.
</P>
<UL>
  <LI>Using <A HREF="http://docker.io/">Docker</A>, the <B>docker ps -a</B> command can then be used to observe running containers.
  <LI>With <A HREF="http://linuxcontainers.org/">LXC</A>, the <B>lxc-ls</B> and <B>lxc-info</B> commands are used for this purpose.
</UL>
<P>Once these commands are correctly reporting running containers, we
are in a position to begin analysing those containers with PCP tools.
</P>
<TABLE WIDTH="100%" BORDER=0 CELLPADDING=10 CELLSPACING=20>
	<TR><TD BGCOLOR="#e2e2e2" WIDTH="70%"><BR><IMG SRC="images/stepfwd_on.png" ALT="" WIDTH=16 HEIGHT=16 BORDER=0>&nbsp;&nbsp;&nbsp;Check local PCP collector installation (requires the <I>pcp-verify</I> utility):<BR>
<PRE><B>$ pcp verify --containers</B></PRE>
</TD></TR>
</TABLE>
<TABLE WIDTH="100%" BORDER=0 CELLPADDING=10 CELLSPACING=20>
	<TR><TD BGCOLOR="#e2e2e2" WIDTH="70%"><BR><IMG SRC="images/stepfwd_on.png" ALT="" WIDTH=16 HEIGHT=16 BORDER=0>&nbsp;&nbsp;&nbsp;Request networking metrics for a host and then a container running on the host:<BR>
<PRE><B>$ pminfo --fetch containers.name containers.state.running

containers.name
    inst [0 or "<SPAN>f4d3b90bea15</SPAN>..."] value "<SPAN>sharp_feynman</SPAN>"
    inst [1 or "d43eda0a7e7d..."] value "cranky_colden"
    inst [2 or "252b56e79da5..."] value "desperate_turing"

containers.state.running
    inst [0 or "f4d3b90bea15..."] value 1
    inst [1 or "d43eda0a7e7d..."] value 0
    inst [2 or "252b56e79da5..."] value 1

$ pmprobe -I network.interface.up
network.interface.up 5 "p2p1" "wlp2s0" "lo" "docker0" "veth2234780"

$ pmprobe -I --container <SPAN>sharp_feynman</SPAN> network.interface.up
network.interface.up 2 "lo" "eth0"

$ pmprobe -I --container <SPAN>f4d3b90bea15</SPAN> network.interface.up
network.interface.up 2 "lo" "eth0"</B></PRE>
</TD></TR>
</TABLE>
<P><B>Note</B>: these commands all query the same <I>pmcd</I> process
- from the host running the container engine.  In other words, <I>there
is no need to install any PCP software inside the monitored containers</I>.</P>

<P><BR></P>
<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0 BGCOLOR="#e2e2e2">
        <TR><TD WIDTH="100%" BGCOLOR="#081c59"><P ALIGN=LEFT><FONT SIZE=5 COLOR="#ffffff"><B><A NAME="cgroups">Control Groups</A></B></FONT></P></TD></TR>
</TABLE>
<P>Control Groups are a Linux kernel mechanism for aggregating
or partitioning sets of tasks, and their children, into hierarchical
groups with specialized behaviour.
This is the underlying technology used for controlling the set of
processes within each container.
</P>
<P>Recall that the concept of a &quot;container&quot; is a user-space
construct only, and it is the role of the container engine to ensure
the kernel cgroup hierarchies are constructed and managed appropriately
for the containers it provides.
</P>
<P>A cgroup <B>subsystem</B> is kernel code that makes use of the task
grouping facilities provided by cgroups to treat groups of tasks in
particular ways.  A subsystem is typically a "resource controller" that
schedules a resource or applies per-cgroup limits.  Examples of cgroup
subsystems used by container engines include the virtual memory subsystem
(<I>memory</I>), the processor accounting subsystem (<I>cpuacct</I>),
the block accounting cgroup (<I>blkio</I>), and several others.
</P>
<P>Within the scope of individual cgroup subsystems, hierarchies
can be created, managed and shaped in terms of the tasks within them.
A <B>hierarchy</B> is a set of cgroups arranged in a tree, such that
every task in the system is in exactly one of the cgroups in the
hierarchy, and a set of subsystems; each <I>subsystem</I> has
system-specific state attached to each cgroup in the hierarchy.
</P>
<P>Each hierarchy has an instance of the cgroup virtual filesystem
associated with it.
</P>

<TABLE WIDTH="100%" BORDER=0 CELLPADDING=10 CELLSPACING=20>
	<TR><TD BGCOLOR="#e2e2e2" WIDTH="70%"><BR><IMG SRC="images/stepfwd_on.png" ALT="" WIDTH=16 HEIGHT=16 BORDER=0>&nbsp;&nbsp;&nbsp;These can be interrogated by querying the PCP <B>cgroup.subsys</B> and <B>cgroup.mounts</B> metrics.<BR>
<PRE><B>$ pminfo --fetch cgroup.subsys.num_cgroups

cgroup.mounts.subsys
    inst [0 or "/sys/fs/cgroup/systemd"] value "?"
    inst [1 or "/sys/fs/cgroup/cpuset"] value "cpuset"
    inst [2 or "/sys/fs/cgroup/cpu,cpuacct"] value "cpu,cpuacct"
    inst [3 or "/sys/fs/cgroup/memory"] value "memory"
    inst [4 or "/sys/fs/cgroup/devices"] value "devices"
    inst [5 or "/sys/fs/cgroup/freezer"] value "freezer"
    inst [6 or "/sys/fs/cgroup/net_cls,net_prio"] value "net_cls,net_prio"
    inst [7 or "/sys/fs/cgroup/blkio"] value "blkio"
    inst [8 or "/sys/fs/cgroup/perf_event"] value "perf_event"
    inst [9 or "/sys/fs/cgroup/hugetlb"] value "hugetlb"

cgroup.subsys.num_cgroups
    inst [0 or "cpuset"] value 1
    inst [1 or "cpu"] value 77
    inst [2 or "cpuacct"] value 77
    inst [3 or "memory"] value 3
    inst [4 or "devices"] value 3
    inst [5 or "freezer"] value 3
    inst [6 or "net_cls"] value 1
    inst [7 or "blkio"] value 77
    inst [8 or "perf_event"] value 1
    inst [9 or "net_prio"] value 1
    inst [10 or "hugetlb"] value 1</B></PRE>
</TD></TR>
</TABLE>

<P>Userspace code (i.e. container engines like Docker and LXC) can
create and destroy cgroups by name in an instance of the cgroup
virtual file system, specify and query to which cgroup a task is
assigned, and list the task PIDs assigned to a cgroup.
Those creations and assignments only affect the hierarchy associated
with that instance of the cgroup file system.
</P>

<P><BR></P>
<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0 BGCOLOR="#e2e2e2">
        <TR><TD WIDTH="100%" BGCOLOR="#081c59"><P ALIGN=LEFT><FONT SIZE=5 COLOR="#ffffff"><B><A NAME="namespaces">Namespaces</A></B></FONT></P></TD></TR>
</TABLE>
<P>Completely distinct to cgroups, at least within the kernel, is the
concept of <I>namespaces</I>.
Namespaces allow different processes to have differing views of several
aspects of the kernel, such as the hostname (UTS namespace), network
interfaces (NET namespace), process identifiers (PID namespace), mounted
filesystems (MNT namespace) and so on.
</P>
<P>When processes share a namespace, they share the same view of a resource.
For example, objects created in one IPC namespace are visible to all other
processes that are members of that namespace, but are not visible to
processes having another IPC namespace.
So for our purposes, all processes running in one container can thus have
a different view to the IPC resources visible to a process running on the
host or a different container.
</P>
<P>Returning to the first networking example in this tutorial, we can see
how namespaces become important from a performance tooling point of view.
For network metrics within a container, we want to be able to report values
for the set of network interfaces visible within that container, instead of
the set from the host itself.
</P>
<P>Finally, it is important to note that namespaces are not a complete
abstraction, in that many aspects of the underlying host remain visible
from within the container.  This affects performance tools in that the
values exported for some metrics can be adjusted and fine-tuned relative
to the container, while others cannot.
</P>

<P><BR></P>
<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0 BGCOLOR="#e2e2e2">
        <TR><TD WIDTH="100%" BGCOLOR="#081c59"><P ALIGN=LEFT><FONT SIZE=5 COLOR="#ffffff"><B><A NAME="details">Containers and PCP</A></B></FONT></P></TD></TR>
</TABLE>
<P ALIGN=LEFT><FONT SIZE=4><B>Core Extensions</B></FONT>
<P>All connections made to the PCP metrics collector daemon (<I>pmcd</I>)
are made using the PCP protocol, which is TCP/IP based and thus
(importantly for containers) connection-oriented.
Each individual monitoring tool has a unique connection to <I>pmcd</I>
and can request values for a specific, custom set of metrics. This includes
being able to request metric values related to a specific, named container.
Note that PCP differs to the design of several other monitoring systems
in this regard, which write or send out a specified set of system-wide
values, on a set interval.
</P>
<P>From a user point of view, this boils down to being able to specify a
container via the interface (whether command line or graphical) of the
PCP monitoring tools and to have that container name transfered to the
PCP collector system.  This allows for filtering and fine-tuning of the
metric values it returns, such that the values are specific to the named
container.
</P>

<P ALIGN=LEFT><FONT SIZE=4><B>Elevated Privileges</B></FONT>
<P>Support for containers was first added in the 3.10.2 version of PCP
(released in January 2015).
This version includes the <I>pmdaroot</I> daemon - a critical component
of the container support, it must be enabled in order to monitor containers.
</P>
<P>It performs privileged operations on behalf of other PCP agents and plays
a pivotal role in informing the other agents about various attributes of the
active containers that it discovers on the PCP collector system.
</P>
<P>Verify that there is a <I>pmdaroot</I> line in <I>/etc/pcp/pmcd/pmcd.conf</I>
and that the <I>pcp</I> command reports that it is running.
</P>

<P ALIGN=LEFT><FONT SIZE=4><B>Container-specific Metric Values</B></FONT>
<P>With that core functionality in place, several kernel agents have been
taught to customize the metric values they report when the monitoring of a
named container has been requested.
These include the network, filesys, ipc, and other metrics in
<I>pmdalinux</I>, as well as the per-process and cgroup metrics in
<I>pmdaproc</I>.
</P>
<TABLE WIDTH="100%" BORDER=0 CELLPADDING=10 CELLSPACING=20>
	<TR><TD BGCOLOR="#e2e2e2" WIDTH="70%"><BR><IMG SRC="images/stepfwd_on.png" ALT="" WIDTH=16 HEIGHT=16 BORDER=0>&nbsp;&nbsp;&nbsp;Requesting container-specific process and control group metrics:
<BR><PRE><B>$ pminfo -t --fetch --container <SPAN>sharp_feynman</SPAN> cgroup.memory.stat.cache proc.psinfo.rss pmcd.hostname

cgroup.memory.stat.cache [Number of bytes of page cache memory]
    inst [2 or "/system.slice/docker-f4d3b90bea15..."] value 9695232

proc.psinfo.rss [resident set size (i.e. physical memory) of the process]
    inst [21967 or "021967 dd if=/dev/random of=/tmp/bits count=200k"] value 676
    inst [27996 or "027996 /bin/bash"] value 2964

pmcd.hostname [local hostname]
    value "<SPAN>f4d3b90bea15</SPAN>"</B></PRE>
</TD></TR>
</TABLE>

<P><BR></P>
<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0 BGCOLOR="#e2e2e2">
        <TR><TD WIDTH="100%" BGCOLOR="#081c59"><P ALIGN=LEFT><FONT SIZE=5 COLOR="#ffffff"><B><A NAME="nextstep">Next Steps</A></B></FONT></P></TD></TR>
</TABLE>

<P>Upcoming versions of PCP will extend the initial work done for container
analysis in several directions, particularly on the monitoring tool side.</P>

<P ALIGN=LEFT><FONT SIZE=4><B>Console Tools</B></FONT>
<P>In addition to adding the container command line option to more of the
existing monitoring tools, a light-weight console-based monitoring tool
with deep container knowledge is planned. This tool will sample and report
cgroup and namespaced metrics for a specified container, in an easily
consumed form - a little like <I>top</I> or <I>vmstat</I> perhaps.</P>

<P ALIGN=LEFT><FONT SIZE=4><B>Graphical Tools</B></FONT>
<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0 STYLE="page-break-before: always">
        <TR><TD WIDTH=500><P ALIGN=CENTER><BR><IMG ALIGN=MIDDLE SRC="images/pmchart_add_host_container.png" ALT="" BORDER=0></P></TD>
        <TD><P>In the PCP strip chart utility <I>pmchart</I> from version 3.10.3 onward, connections to containers can be established using the "Add Host" dialog, as shown to the left.</P>
	<P>This can be accessed via the "New Chart" or "Open View" menu entries.</P>
	<UL>
        <LI>Specify the name of the PCP Collector system where <I>pmcd</I> is running.
        <LI>Press the "Advanced" push button to enable additional connection attributes to be specified.
        <LI>Select the "Container" check box, and enter a container name.
        <LI>Press "OK" to establish a connection to the container on host -
            this functions in much the same was as the <I>pminfo</I> examples
            from earlier in this tutorial.
	</UL>
	</TD>
        </TR>
</TABLE>

<P ALIGN=LEFT><FONT SIZE=4><B>Domain Agents</B></FONT>
<P>As the underlying container technologies are maturing, they are becoming
instrumented for analysis. For example, <B>etcd</B> and <B>docker</B> now
have APIs to extract the instrumentation about their behaviour, which will
be exported as new PCP metrics.</P>

<P><BR></P>
<HR>
<CENTER>
<TABLE WIDTH="100%" BORDER=0 CELLPADDING=0 CELLSPACING=0>
	<TR> <TD WIDTH="50%"><P>Copyright &copy; 2007-2010 <A HREF="http://www.aconex.com/"><FONT COLOR="#000060">Aconex</FONT></A><BR>Copyright &copy; 2000-2004 <A HREF="http://www.sgi.com/"><FONT COLOR="#000060">Silicon Graphics Inc</FONT></A></P></TD>
	<TD WIDTH="50%"><P ALIGN=RIGHT><A HREF="http://pcp.io/"><FONT COLOR="#000060">PCP Site</FONT></A><BR>Copyright &copy; 2012-2015 <A HREF="http://www.redhat.com/"><FONT COLOR="#000060">Red Hat</FONT></A></P></TD></TR>
</TABLE>
</CENTER>
</BODY>
</HTML>
pcp 3.10.8build1 / usr / share / doc / pcp-doc / html / lab.containers.html
