Difference between revisions of "Sar/Visualize CPU data"

From braindump
Jump to navigation Jump to search
 
(11 intermediate revisions by the same user not shown)
Line 1: Line 1:
{{DISPLAYTITLE: Visualize sar CPU data with R}}
This is a five minute guide how to visualize Linux's <tt>sar</tt> data provided by the <tt>sysstat</tt> utility without a lot of mangeling the data.
This is a five minute guide how to visualize Linux's <tt>sar</tt> data provided by the <tt>sysstat</tt> utility without a lot of mangeling the data. The examples outlined below were done in CentOS 6.5. For other distros or Unix flavors your milage may vary.


== Goal ==
== Goal ==
Line 12: Line 13:
=== Dumping the <tt>sar</tt> data with <tt>sadf</tt> ===
=== Dumping the <tt>sar</tt> data with <tt>sadf</tt> ===
The data <tt>sar</tt> collects is in binary format and needs to be converted first to a format that can be imported into <tt>R</tt>. This is done with the <tt>sadf</tt> command which converts the collected data into tabular data delimited by semicolon. <br />
The data <tt>sar</tt> collects is in binary format and needs to be converted first to a format that can be imported into <tt>R</tt>. This is done with the <tt>sadf</tt> command which converts the collected data into tabular data delimited by semicolon. <br />
'''Note:''' On CentOS 6 and higher the <tt>sadf</tt> command also prints a header file to make most use of it we need to slightly changes it like remove the leading <tt>#</tt>, plus remove the <tt>%</tt> from the cpu data but only in the first. Other lines starting with <tt>#</tt> or containing a <tt>LINUX-RESTART</tt> should also be removed.
'''Note:''' On CentOS 6 and higher the <tt>sadf</tt> command also prints a header file to make most use of it we need to slightly changes it like remove the leading <tt>#</tt>, plus remove the <tt>%</tt> from the cpu data but only in the first. Other lines starting with <tt>#</tt> or containing a <tt>LINUX-RESTART</tt> should also be removed. Your milage may vary!

sadf -t -d -P ALL <span class="input"><SAR-FILE></span> | \
sadf -t -d -P ALL <span class="input"><SAR-FILE></span> | \
sed -e '1,1s/\(^#\|%\)//g' \
sed -e '1,1s/\(^#\|%\)//g' \
Line 22: Line 24:


library( ggplot2 )
library( ggplot2 )
cpu.data <- read.csv( file="<span class="input"><sadf-output></span>", sep=";" )
cpu.data <- read.csv( file="<span class="input"><SADF-OUTPUT></span>", sep=";" )
cpu.data$timestamp <- as.POSIXct( cpu.data$timestamp )
cpu.data$timestamp <- as.POSIXct( cpu.data$timestamp )
cpu.data$CPU[ cpu.data$CPU == "-1" ] <- "all"
cpu.data$CPU[ cpu.data$CPU == "-1" ] <- "all"

With very little effort data is read into <tt>data.table</tt> format. Then we have to change the format of two fields. Namely <tt>timestamp</tt> needs conversion from a string to a proper time format like <tt>POSIXct</tt>. Plus the value for all CPUs is "-1" and to make it clear to the viewer we want it to be "all".

After the changes the structure of the cpu.data should look like the below data.
str( cpu.data )
'data.frame': 429 obs. of 10 variables:
$ hostname : Factor w/ 1 level "hostname.local": 1 1 1 1 1 1 1 1 1 1 ...
$ interval : int 588 588 588 590 590 590 589 589 589 588 ...
$ timestamp: <span class="highlight">POSIXct</span>, format: "2014-05-04 00:10:01" "2014-05-04 00:10:01" ...
$ CPU : <span class="highlight">chr</span> "all" "0" "1" "all" ...
$ user : <span class="highlight">num</span> 8.05 9.39 6.73 5.28 5.02 ...
$ nice : num 0 0 0 0 0 0 0 0 0 0 ...
$ system : num 3.83 4.32 3.34 2.75 3.21 2.3 4.59 4.99 4.19 5.1 ...
$ iowait : num 0.71 1.1 0.33 0.7 1.1 0.3 0.7 1.12 0.29 0.81 ...
$ steal : num 0 0 0 0 0 0 0 0 0 0 ...
$ idle : num 87.4 85.2 89.6 91.3 90.7 ...

=== Setting up the graph ===
The next step is to store the ggplot data into a variable for further processing. It's important to group the data by the <tt>CPU</tt> field and assign a different color to each CPU with the <tt>colour=</tt> assignment.

cpu.graph <- ggplot( data=cpu.data, aes( x=timestamp, y=user, group=CPU, colour=CPU ) )
cpu.graph <- ggplot( data=cpu.data, aes( x=timestamp, y=user, group=CPU, colour=CPU ) )
cpu.graph + geom_line()


Will result in a graph like this:
=== Display the result as a line graph ===
To display the graph a simply calling the stored graph data and assigning a layer with <tt>geom_line()</tt>.
cpu.graph + geom_line()
Will result in a graph like this:

[[Image:Cpu-data-consolidated.png]]
[[Image:Cpu-data-consolidated.png]]

=== Plotting each CPU separately ===
=== Plotting each CPU separately ===
To show which CPU or core is used most it is probably better to separately print the CPUs. The <tt>ggplot2</tt> library comes with a nifty command called <tt>facet_grid()</tt>. To print each separately simply add it to the end of the previous command.
To show which CPU or core is used most it is probably better to separately print the CPUs. The <tt>ggplot2</tt> library comes with a nifty command called <tt>facet_grid()</tt>. To print each separately simply add it to the end of the previous command.
Line 38: Line 64:


[[Image:Cpu-data-grid.png]]
[[Image:Cpu-data-grid.png]]

=== Combining it with more data ===
To really make this graph useful a few more data points need to be added in this case iowait and system. To make sure they are not overlapping the data needs to be stacked. Firstly a new graph variable needs to be defined. And then for each y axis a new <tt>geom_line</tt> is defined. There is probably better ways but the of each is calculated and then plotted. The <tt>facet_grid</tt> got a bit of an upgrade and has better labeling. The last three lines are sugar coating to make the graph fit for management :).

cpu.stacked <- ggplot( data=cpu.data, aes( x=timestamp, group=CPU ) )
cpu.stacked +
geom_line( aes( y=iowait+user+system, colour="system" ) ) +
geom_line( aes( y=iowait+user, colour="user" ) ) +
geom_line( aes( y=iowait, colour="iowait" ) ) +
facet_grid( CPU~., labeller=label_both ) +
xlab( "Datetime" ) +
ylab( "CPU usage" ) +
theme( legend.title=element_blank() )

The result is:

[[Image:Cpu-data-full.png]]

=== Showing more detail by reducing the window ===
When using SAR data with a high frequency probe over a long time span it is good to show an overall trend. When locating a hotspot it makes sense to zoom in by reducing the window. This is done by subsetting the data.

cpu.data.sub <- subset( cpu.data, timestamp %in% as.POSIXct( "2014-01-29 02:00:00" ):as.POSIXct( "2014-01-29 04:00:00" ) )

Then run the graph code above again with the newly created data frame.

=== Saving the graph to a file ===
Especially when automation comes into play saving to a file is a must this example shows how to save to PNG.

setwd( "/home/r-user/cpu-data" )
png( "cpu-graph-grid.png", width=600, height=360, res=72 )
cpu.graph + geom_line() + facet_grid( CPU~. )
dev.off()


[[Category: R]]
[[Category: R]]

Latest revision as of 22:50, 31 January 2016

This is a five minute guide how to visualize Linux's sar data provided by the sysstat utility without a lot of mangeling the data. The examples outlined below were done in CentOS 6.5. For other distros or Unix flavors your milage may vary.

Goal

Create CPU graphs in R from the sar utility without massaging the output data too much.

Prerequisites

  • The Linux sysstat package installed and configured to report performance data.
  • R
  • ggplot2 R library

Howto

Dumping the sar data with sadf

The data sar collects is in binary format and needs to be converted first to a format that can be imported into R. This is done with the sadf command which converts the collected data into tabular data delimited by semicolon.
Note: On CentOS 6 and higher the sadf command also prints a header file to make most use of it we need to slightly changes it like remove the leading #, plus remove the % from the cpu data but only in the first. Other lines starting with # or containing a LINUX-RESTART should also be removed. Your milage may vary!

sadf -t -d -P ALL <SAR-FILE> | \
  sed -e '1,1s/\(^#\|%\)//g' \
      -e '/\(^#\|LINUX-RESTART\)/d' \
  > <SADF-OUTPUT>

Importing the data into R

The next step is to read the tabular data into R and print the graphs there are just a handful of commands to do this. In R type the following commands.

library( ggplot2 )
cpu.data <- read.csv( file="<SADF-OUTPUT>", sep=";" )
cpu.data$timestamp <- as.POSIXct( cpu.data$timestamp )
cpu.data$CPU[ cpu.data$CPU == "-1" ] <- "all"

With very little effort data is read into data.table format. Then we have to change the format of two fields. Namely timestamp needs conversion from a string to a proper time format like POSIXct. Plus the value for all CPUs is "-1" and to make it clear to the viewer we want it to be "all".

After the changes the structure of the cpu.data should look like the below data.

str( cpu.data ) 
'data.frame':	429 obs. of  10 variables:
 $ hostname : Factor w/ 1 level "hostname.local": 1 1 1 1 1 1 1 1 1 1 ...
 $ interval : int  588 588 588 590 590 590 589 589 589 588 ...
 $ timestamp: POSIXct, format: "2014-05-04 00:10:01" "2014-05-04 00:10:01" ...
 $ CPU      : chr  "all" "0" "1" "all" ...
 $ user     : num  8.05 9.39 6.73 5.28 5.02 ...
 $ nice     : num  0 0 0 0 0 0 0 0 0 0 ...
 $ system   : num  3.83 4.32 3.34 2.75 3.21 2.3 4.59 4.99 4.19 5.1 ...
 $ iowait   : num  0.71 1.1 0.33 0.7 1.1 0.3 0.7 1.12 0.29 0.81 ...
 $ steal    : num  0 0 0 0 0 0 0 0 0 0 ...
 $ idle     : num  87.4 85.2 89.6 91.3 90.7 ...

Setting up the graph

The next step is to store the ggplot data into a variable for further processing. It's important to group the data by the CPU field and assign a different color to each CPU with the colour= assignment.

cpu.graph <- ggplot( data=cpu.data, aes( x=timestamp, y=user, group=CPU, colour=CPU ) )

Display the result as a line graph

To display the graph a simply calling the stored graph data and assigning a layer with geom_line().

cpu.graph + geom_line()

Will result in a graph like this:

Cpu-data-consolidated.png

Plotting each CPU separately

To show which CPU or core is used most it is probably better to separately print the CPUs. The ggplot2 library comes with a nifty command called facet_grid(). To print each separately simply add it to the end of the previous command.

cpu.graph + geom_line() + facet_grid( CPU~. ) 

Which will result in a graph like this.

Cpu-data-grid.png

Combining it with more data

To really make this graph useful a few more data points need to be added in this case iowait and system. To make sure they are not overlapping the data needs to be stacked. Firstly a new graph variable needs to be defined. And then for each y axis a new geom_line is defined. There is probably better ways but the of each is calculated and then plotted. The facet_grid got a bit of an upgrade and has better labeling. The last three lines are sugar coating to make the graph fit for management :).

cpu.stacked <- ggplot( data=cpu.data, aes( x=timestamp, group=CPU ) ) 
cpu.stacked + 
  geom_line( aes( y=iowait+user+system, colour="system" ) ) + 
  geom_line( aes( y=iowait+user, colour="user" ) ) + 
  geom_line( aes( y=iowait, colour="iowait" ) ) + 
  facet_grid( CPU~., labeller=label_both ) + 
  xlab( "Datetime" ) +
  ylab( "CPU usage" ) + 
  theme( legend.title=element_blank() ) 

The result is:

Cpu-data-full.png

Showing more detail by reducing the window

When using SAR data with a high frequency probe over a long time span it is good to show an overall trend. When locating a hotspot it makes sense to zoom in by reducing the window. This is done by subsetting the data.

cpu.data.sub <- subset( cpu.data, timestamp %in% as.POSIXct( "2014-01-29 02:00:00" ):as.POSIXct( "2014-01-29 04:00:00" ) )

Then run the graph code above again with the newly created data frame.

Saving the graph to a file

Especially when automation comes into play saving to a file is a must this example shows how to save to PNG.

 setwd( "/home/r-user/cpu-data" )
 png( "cpu-graph-grid.png", width=600, height=360, res=72 )
 cpu.graph + geom_line() + facet_grid( CPU~. )  
 dev.off()