uber-archive/cpustat

cpustat - high(er) frequency stats sampling

(This project is deprecated and not maintained.)

cpustat is a tool for Linux systems to measure performance. You can think of it like a fancy sort of top that does different things. This project is motivated by Brendan Gregg's USE Method and tries to expose CPU utilization and saturation in a helpful way.

Most performance tools average CPU usage over a few seconds or even a minute. This can create the illusion of excess capacity because brief spikes in resource usage are blended in with less busy periods. cpustat takes higher frequency samples of every process running on the machine and then summarizes these samples at a lower frequency. For example, it can measure every process every 200ms and summarize these samples every 5 seconds, including min/average/max values for some metrics.

There are two ways of displaying this data: a pure text list of the summary interval and a colorful scrolling dashboard of each sample.

Here are examples of both modes observing the same workload:

Installation

To install the latest version of cpustat and all of its dependencies, try this:

go get github.com/uber-common/cpustat

If you have glide, you can use glide install to get consistent dependencies.

Usage

This program uses Linux taskstats, which requires root.

Here are the command line flags most users will want:

There are also a few less common options:

Examples:

sudo cpustat -s 500 -s 10 -n 20

This will take a sample of all processes every 500ms and summarize this data after 10 samples, which is every 5 seconds.

sudo cpustat -u mjr,mranney

Only measure processes that are owned by either user mjr or user mranney. The overall system stats will still be measured.

sudo ./cpustat -p $(pgrep -d, vim\|emacs)

Only measure processes that pgrep thinks are called "vim" or "emacs". The -p option to cpustat takes a list of process ids to measure, and pgrep is a handy way to get this list. The -d, option to pgrep prints the list of matching pids with a comma separator.

Displayed Values

In pure text mode, there are some system-wide summary metrics that come from /proc/stat:

In fancy scrolling dashboard mode, the unique panes are as follows:

In the top right, labeled "total usr/sys time", the system-wide measurements for user time and system time are displayed. User time is cyan, and system time is red. The X axis is labeled in seconds, the Y axis is percentage of a CPU. Each dot represents an individual sample. The data is not summarized.

In the top left, labeled "top procs", the combined user+system time of the topN processes is displayed. Each process gets a separate line with a color that matches the list below. The X and Y axes are the same as the top right graph. Due to limitations of a terminal UI, overlapping lines are drawn in the same cell, potentially obscuring each other.

Both modes display the same per-process summary data. The fields are:

Understanding the Output

Here are a few examples of running cpustat on a 4 processor vm on my laptop.

The first is a mostly idle system where the only thing really running is cpustat itself:

The idle min/avg/max shows that for most of the 20 samples, the system was almost completely idle. We can tell that because we know this is a 4 processor system, so the maximum value for idle is 400. It's perhaps a little surprising then that prun would show 1.0/1.1/3.0, meaning that we never woke up to find fewer than 1 process running. On an idle system, surely that number should be lower. What's happening is that cpustat ends up measuring itself as the single running process.

So let's put this computer to work:


I'm running a single instance of "CPU Burn-In" that shows up as burnP6. This process uses a single CPU. We can see that the overall system now reports only about 300% idle, and that burnP6 is only using a single thread from the thrd column.

We can also see that the runq column for burnP6 shows 0.4 on the first summary interval. This means that for the duration of the summary interval, 0.4% of an effective CPU's time was asked for by a process, but that process wasn't scheduled for whatever reason. During the same interval, icx is a lot higher than normal. icx is "involuntary context switches". It's hard to say exactly what caused this, but we can also see that several other processes have non-0 iow or "IO wait". During that interval, prun had a max value of 5.0, so a few things probably woke up at the same time to do some brief work, caused a bit of interference, then went back to sleep.

During the second summary interval, the runq is back to 0, so burnP6 is getting all of the time it wants. It's also curious to note that the burnP6 pretty clearly reports its CPU usage as 100% usr time, the overall system usr doesn't line up with this at 60.0/77.2/95.0, but the system idle does. I'm not exactly sure what causes this, but it's something about how burnP6 works and Linux accounts for it. Many other single threaded programs in a tight loop do not exhibit this behavior.

I have another program to generate a more irregular an in my experience more realistic workload called wastetime. This program uses many threads, tries to wake them up at the same time to do some work, then sleeps for a bit.


For some of the samples, we wake up and find wastetime using 0 CPU, and sometimes we find it 300% because burnP6 is using the other 100%. wastetime reports a runq time of around 230%. This means that we'd need approximately 2.3 more CPUs on average to do all of the work and avoid delay. This is the average delay, but we can see from the prun max that we pretty regularly need 35 CPUs to avoid absolutely all delays.

We also see that wastetime is causing a little bit of runq interference for burnP6.

This is what it looks like to run 10 parallel instances of go build in this vm on the cpustat source with:

for C in {0..10}; do echo $C ; (go build &) ; done


This is obviously way more work than my little vm can manage, so we've triggered some major CPU saturation.

Nearly every process running on the machine is spending more time in the runq than we'd like. Even so, there is still some idle time left in the system. That can be explained by the swap time and iowait time. Also note that the sam column is all less than 20. That means these are all short lived processes, which are often hard to account for.

Data Sources

Every sample interval, the following data sources are checked:

• The directory /proc is scanned to get the current list of process ids
• For each pid, read /proc/pid/stat, compute difference from previous sample
• If this is a new pid, read /proc/pid/cmdline
• For each pid, send a netlink message to fetch the taskstats, compute difference from previous sample.
• Fetch /proc/stat to get the overall system stats

Each sleep interval is adjusted to account for the amount of time spent fetching all of these stats. Each sample also records the time it was taken to scale each measurement by the actual elapsed time between samples. This attempts to account for delays in cpustat itself.

Run within a Docker container

 docker run --rm -ti --privileged --pid=host --net=host user/cpustat -s=200 -n=20

Limitations

There are many important limitations to understand before drawing conclusions from cpustat's output.

The first is that all sampling systems are subject to Nyquist, which is both a practical reality and a fun concept that will lend gravitas to many conversations. Even if we sample at 200ms, which is pretty fast compared to most tools, we'll not be able to accurately measure many common performance problems.

Linux CPU time accounting is done in terms of whole "clock ticks", which are often 100ms. This can cause some strange values when sampling every 200ms. Common problems are that CPU utilization will appear to slightly exceed the maximum number of CPUs on the system. On a very lightly loaded system, sometimes processes will appear to oscillate between 0 percent and some higher number.

There is no way to get a consistent snapshot of all processes on a Linux system. After we scan /proc to get the process list, new processes could come and go. Processes could exit at any time along the process of scanning /proc, reading /proc/cmdline, /proc/pid/stat, or netlink taskstats.

cpustat itself can cause the very problems it was written to expose by doing a burst of work on a regular interval. It would be nicer to the underlying system to spread the work out evenly over the sampling interval instead of trying to do it all at once.

The Linux netlink taskstats interface can only be used by root, which means this program must be run as root.

In spite of these limitations, this tool has already been useful in understanding performance problems on production systems. I hope it's useful to you as well.

Agent

In addition to the interactive version of cpustat, a long running measurement server is available in the agent directory. This program uses only the measurement logic to record the raw samples in a circular buffer. The data in the agent can be collected by the client program. This can be used to summarize the system state for publishing metrics to other systems. For example, instead of recording the average CPU utilization over a minute like many metrics systems do, you could report the min/avg/max CPU utilization over a minute or any other interval.

Future Work

There is an almost an endless set of UI-type features that would be nice.