% Plotting Points % frustration, data, #pinned % 2014-07-31
Yesterday, I captured some data from pidstat during a failure simulation. I had hypothesized that the components memory usage would be pretty stable regardless of failure, but wanted to confirm it just to be sure. At the end of the simulation, I had a bunch of pidstat records that looked like this (with the params I gave it):
06:53:23 PM PID %usr %system %guest %CPU CPU Command
06:53:28 PM 19437 76.05 7.78 0.00 83.83 1 lumbermill
06:53:23 PM PID minflt/s majflt/s VSZ RSS %MEM Command
06:53:28 PM 19437 31.74 0.00 1669088 1142476 7.25 lumbermill
06:53:23 PM PID kB_rd/s kB_wr/s kB_ccwr/s Command
06:53:28 PM 19437 0.00 0.00 0.00 lumbermill
I wanted only column 8 of the memory related record, so I did something pretty ugly:
$ grep -C 1 '%MEM' pidstat.testing | grep -v '%MEM' \
| grep -v -E "^$" | grep -E -v "^--" | awk '{print $8}'
This produced the right numbers in a single column, but there wasn’t a way for me to easily see them. I could have prepped a simple gnuplot script of course, but that was more trouble than I was looking for, even if it would have been most trivial to do:
$ grep -C 1 '%MEM' pidstat.testing | grep -v '%MEM' \
| grep -v -E "^$" | grep -E -v "^--" \
| awk 'BEGIN { i = 0} {print i"\t"$8; i += 1}' > /tmp/foo.dat
$ gnuplot
gnuplot> set terminal x11
Terminal type set to 'x11'
Options are ' nopersist'
gnuplot> plot '/tmp/foo.dat' using 1:2
So, I did actually do that, but I wanted to investigate other ways to go about this:
I had an obsession with [emacs calculator][calc] the other week, so I switched buffers to \*shell\* ran the above pipeline, and got a single column of output. Selected it all, C-x \* R to load it into \*Calculator\*. Then did V a -1, to flatten the data into a 1d vector. Then, g f for (graph fast). I quickly realized that I needed a second vector for the x coordinates, so I ran the 2 column variant, and inserted the xs into the calculator, and had a graph to display. Since emacs just uses gnuplot for this, the graph was no different, but the method didn’t involve actual gnuplot interaction to use it.
Surely, I thought, there must be something better. Well, this was timeseries data. I figured, maybe the venerable rrdtool can easily do this. Oh, how I was wrong about easy.
I created an rrd database, with something along the lines of:
$ rrdtool create /tmp/pidstat.rrd --step 5 \
--start 1406802180 DS:mem:GAUGE:5:0:100 \
DS:cpu:GAUGE:5:0:100 RRA:AVERAGE:0.5:1:2000
And then, I needed TS:data pairs. Back I went to reextract that data out. This time with and even uglier pipeline and some more advanced awking:
$ grep -C 1 '%MEM' pidstat.testing | grep -v '%MEM' \
| grep -v -E "^$" | grep -E -v "^--" \
| awk '{while ( ("date +%s --date=\"07/31/2013 "$1"PM""\"" \
| getline result) > 0 ) { print result":"$8 } }'
That got me epoch time from the times in the file, and when I tacked on:
| xargs rrdupdate /tmp/pidstat.rrd -t mem
It all was loaded into the database. Surely, I can generate a nice looking plot now?
No! Why? Because I need to invoke rrdtool graph with some crazy array of options to spit out anything, and I didn’t even bother!
I could have writtten a quick HTML page with a flot chart faster. Why isn’t there a simple DWIM tool that allows me to tack on | linegraph?