Performance

Introduction

When thinking about the performance of collectl keep in mind that efficiency is one of the main design principles and it should therefore be possible to run collectl as a daemon on most systems with little concern for the overhead it generates. However, given that the another design principal it to be able to collect a very broad and deep set of data, it is worth understanding a little more about just what goes on in collectl.

During the earlier days of collectl development a third design consideration was minimizing the amount of storage used for raw files. At the time disks were smaller and the amount of data collected was small enough that being more judicious about what was saved felt like it made a difference. However, since the inclusion of process, slab and now interrupt data along with larger disks the amount of storage used has become less of a consideration. In other words don't spend extra data collection cycles trying to be more selective about what is recorded if it's going to add to the overhead.

Measuring the Overhead

If you take a closer look at the architecture you can see that the path of minimal overhead is when collectl reads from /proc and write it to a file. This is not an accident! In fact, on many system this has been observed to take less than 0.1% of the CPU and this is when collecting almost all the data collectl is capable of. However, this also raises several important questions for consideration:

What is the overhead on a specific system?
Is there a way to reduce the overhead further?
Is collection overhead for all data types the same and if not, what is it?

In fact, these are the very questions I asked during the initial development as well as every time I add support for new data types. In order to be able to answer these questions quantitatively, collectl should be run with a collection interval of 0 and told to collect 8640 samples, the number of 10-second samples in a day. By timing the execution you can then get a reasonable estimate of the daily overhead. Consider the following:

# time collectl -scdnm -i0 -c8640 -f /tmp
real    0m9.711s
user    0m7.480s
sys     0m2.140s

and you can see that collectl uses about 10 seconds out of 86400 or about 0.01% of the cpu to collect cpu, disk, network and memory data. If we repeat the test again for just cpu the time drops to under 4 seconds, so you can see if performance is really critical, you can improve things by recording less data or maybe just do it less frequently. The point is these are tradeoffs only you can make if you feel collectl is using too much resource.

So what happens if you take a different processing path and save collectl data in plot format? This means adding the additional overhead of parsing the /proc data and performing some basic math of the values. If we use the same command as above and include -P:

# time collectl -scdnm -i0 -c8640 -f /tmp
real    0m20.607s
user    0m17.970s
sys     0m2.580s

we can see that this takes a little over twice as much overhead even though it is still pretty low.

One other example worth mentioning and is process monitoring overhead, which is the highest overhead operation collectl can do and one of the reasons it has its own monitoring interval. The overheard for collection of this type of data can vary quite broadly depending on how may processes are running at the time and on a system with only 138 processes look at this:

# time ./collectl.pl -sZ -i0 -c8640 -f /tmp
real    1m8.453s
user    0m54.650s
sys     0m13.430s

noting collectl is also smart enough to only look at 1/6 as many samples of process data since that is the default relationship of process monitoring to other subsystem data. This also leads to the mention of a way to further optimize process monitoring. If you are monitoring a specify set of processes, say http daemons, collectl no longer has to look at as much data in /proc and so we now see:

# time ./collectl.pl -sZ -Zchttp -i0 -c8640 -f /tmp
real    0m5.721s
user    0m4.480s
sys     0m1.180s

In fact, if we know there are never going to be any new http process appearing (collectl looks for new processes that match selection strings by default):

# time ./collectl.pl -sZ -Zchttp -i0 -c8640 --procopts p  -f /tmp
real    0m5.080s
user    0m3.930s
sys     0m1.130s

And things get even better as you could even image monitoring these processes at the same interval as everything else with almost no additional overhead!

Another wrinkle - process I/O statistics

Since the inclusion of process I/O stats, collectl now has to read an addition set of data for each process, specifically /proc/pid/io, which adds over 25% to the total process data collection load. For most users, collectl's overhead is reasonable low enough for this extra overhead to not be a problem. Remember - we're talking about an increase of 25% to a relatively small number. But for those concerned with this extra overhead, a new --procopt value of I has been added to V3.6.1 which suppresses the reading of process I/O stats.

Remember - your milage can and will vary!

The number of different combinations of switches one can measure far exceeds the scope of this discussion - for example we haven't even talked about the pros/cons of compression. But in conclusion, the main thing to remember is collectl's overhead is already pretty low but if you're really concerned, measure it yourself and adjust accordingly.

updated Jan 15, 2012