Monitor resource contention with Pressure Stall Information

Pressure Gauge

Article from Issue 238/2020

Pressure Stall Information (PSI) is a new feature that gives users a better view of resource contention.

CPU, RAM, and I/O are the three most important computing resources. If these resources are depleted, processes start to fight for them and resource contention scenarios occur. To avoid these problems, it makes sense to monitor precisely what resources applications use in order to correctly dimension the hardware requirements and make optimal use of the existing hardware.

Smartphones can also benefit from this type of monitoring: when resources become scarce, phones often terminate programs in the background in order to offer the foreground application sufficient resources.

Load Average

Most administrators keep an eye on the load average to determine the extent to which a system is exposed to load. The uptime or top (Listing 1) commands both display the load average, reading the values from the /proc/loadavg file. The file contains five values (Line 6). In addition to the three load average values, the fourth value counts the current executable kernel scheduling entities and the fifth shows the process ID of the latest process created.

Listing 1

Load Average

01 $ strace -e file uptime
02 [...]
03 openat(AT_FDCWD, "/proc/loadavg", O_RDONLY) = 4
04 16:17:55 up 3 days, 3:27, 1 user,  load average: 0,44, 0,78, 0,64
05 $ more /proc/loadavg
06 0.12 0.26 0.36 2/909 130726

The first three values state the load average for the executable processes, that is, the number of processes with a status of R (for runnable) that are waiting for CPU cycles and those with a status of D (for disk sleep) that are waiting for I/O. The three values show the load average for a period of 1, 5, and 15 minutes.

Few admins actually understand these values, and even the kernel developers don't actually think too much of them (Figure 1) [1].

Figure 1: Comment by Linux scheduler maintainer Peter Zijlstra on loadavg.c.

The brief summary of their findings is: the load average is not useful when it comes to assessing system utilization. The admin can, at most, use the three values to estimate whether a reported performance problem has already been solved because the values become smaller again. However, computing the load average has other disadvantages:

  • The admin has to interpret the values relative to the number of threads and CPUs. A value of 128 can be completely acceptable on a 128-thread system. On a 4-threaded system, however, probably nothing will work.
  • The value does not tell you how long a process had to wait for resources.
  • CPU and I/O are related, which makes it impossible to examine CPU and I/O requirements separately.
  • Because the minimum resolution is one minute, the load cannot be observed in real time.

These problems make the load average unsuitable as a metric for admins who need to react quickly to performance events.


To better identify resource conflicts, and to make it easier to manage overloaded systems at Facebook's data centers, Facebook's kernel team developed Pressure Stall Information (PSI). According to the project's website [2], PSI is a canonical new way to obtain utilization metrics for memory, CPUs, and I/O via the Linux kernel.

Linus Torvalds published the 2018 implementation by Johannes Weiner in Linux 4.20. Weiner is also listed as the PSI maintainer [3]. In Debian 10 "Buster," this functionality is not yet included, but it can be found in Ubuntu from version 19.04 with a Linux kernel as of version 5.0.

Like other common kernel parameters, PSI values are accessed through the /proc pseudo-filesystem, which acts as an interface to kernel data structures. PSI values appear in three different files within the /proc/pressure directory:

  • cpu – information related to CPU utilization
  • memory – information on the time processes spend waiting due to memory issues
  • io – information on the time spent waiting for I/O

Though the three files offer slightly different data, they are organized in a similar way.


The /proc/pressure/cpu file contains four values: avg10, avg60, avg300, and total (Listing 2, Line 2). The values starting with avg* represent the percentage of processes in the last 10, 60, and 300 seconds that had to wait for CPU resources.

Listing 2

Measuring with cpu

01 $ more /proc/pressure/cpu
02 some avg10=0.02 avg60=0.05 avg300=0.02 total=3990454056
03 [...]
04 $ watch -n ,1 grep -R . /proc/pressure/cpu
05 Every 0.1s: grep -R . /proc/pressure/cpu   machine: Mon Jun  8 17:54:34 2020
06 some avg10=0.00 avg60=0.00 avg300=1.07 total=2684388193

Suppose two processes each occupy 100 percent of one thread and run for five minutes on a two-threaded system where nothing else is running. In this case, all three averages are zero, of course, because the processes didn't have to wait. The fact that PSI delivers accuracy to within 10 seconds is an important benefit, because you can react 50 seconds earlier than with the load average.

The last value, total, also makes a big difference: the total value specifies the total time in microseconds for which processes had no CPU resources available. You can read this value at any time – about every 500 milliseconds. The difference between two successive values is the number of microseconds in this time interval in which processes had to wait for CPU resources.

This kind of granular resolution was not available previously in the Linux kernel. Thanks to watch, you can see how the value increases in realtime (Listing 2, Lines 4 to 6), which means you can monitor resource bottlenecks in realtime and, if necessary, kill unimportant processes or migrate them to other systems.

Buy this article as PDF

Express-Checkout as PDF
Price $2.95
(incl. VAT)

Buy Linux Magazine

Get it on Google Play

US / Canada

Get it on Google Play

UK / Australia

Related content

  • System Diagnosis Tools

    To check on the health of a Linux system, administrators can turn to vmstat, iostat, netstat, and ifstat. Or, you can just use the versatile dstat, which combines the features of several tools in a single package.

  • OOM Killer

    When a Linux system runs out of memory, a special agent, the out-of-memory killer, rushes to its aid. Facebook has now introduced its own OOM killer. What makes it different from its kernel-based counterpart? And what is an OOM killer really?

  • Command Line – Probing /proc

    The mysterious /proc virtual filesystem is a rich mine of information about everything in your system.

  • Load Average

    What is the real meaning of those little “load average” values in the output of shell commands like procinfo and uptime, and what can you do with these numbers?

  • System Status Tools

    A system monitor lets you query the system’s current health state. If you are unhappy with the spartan Top tool, try one of these easy alternatives.

comments powered by Disqus

Direct Download

Read full article as PDF:

Price $2.95