Sar/Visualize CPU data

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:

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.

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()