psi-notify |

tl;dr: psi-notify can alert you when resources on your machine are becoming oversaturated, and allow you to take action before your system slows to a crawl.

psi-notify is a minimal unprivileged notifier for system-wide resource pressure using PSI. This can help you to identify misbehaving applications on your machine before they start to severely impact system responsiveness, in a way which MemAvailable, CPU graphs, I/O utilisation graphs and other metrics cannot.

Features

• Runs unprivileged
• Minimal resource usage
• Works with any notifier using Desktop Notifications

Requirements

• Linux 4.20+ with CONFIG_PSI (enabled by default in most distributions)
• libnotify

Installation

On Arch, the psi-notify AUR package is available.

On Fedora and RHEL/CentOS 8, the psi-notify package is available in Fedora/EPEL.

Otherwise, manual installation is as simple as running make and putting the resulting psi-notify binary in your PATH. You will need libnotify installed.

After that, you just start psi-notify. A systemd user service is packaged and can be used like so:

systemctl --user start psi-notify

Config

Put your configuration in ~/.config/psi-notify. Here's an example that will check roughly every 5 seconds⁺, and pop up a notification when the values are exceeded:

update 5
log_pressures false

threshold cpu some avg10 50.00
threshold memory some avg10 10.00
threshold io full avg10 15.00

The above is the default configuration if no config file exists. You may have to tweak these depending on your hardware, machine behaviour, and normal workloads.

You can reload the config without restarting by sending SIGHUP to psi-notify.

Look at the "config format" section below to find out more about what a valid config looks like.

⁺ PSI has poll() support, but it's not currently available to unprivileged users. See this discussion.

Comparison with oomd

oomd and psi-notify are two compatible and complementary projects -- they're not in opposition to each other. oomd also uses PSI metrics, but it requires a policy about "what to kill" in high-pressure scenarios. For example, on a web server we obviously don't want to kill the web server if we can avoid that, so we should prioritise other applications. On the desktop though, it's hard to say: should we kill Chrome, or some IDE, or maybe something playing a movie? It's extremely difficult (although perhaps possible) to produce a single configuration that will do the right thing in even the majority of cases, so we opt to alert early instead and have the user make the decision about what's high priority in their user session. When integrating oomd for the desktop, most distributions will likely end up having to make it less aggressive than would be ideal, so they can still interoperate.

It's hard to produce a good policy for, say, one's working day, where at one time my terminal is the most critical thing, at another my browser is, and at another it's my mail client. At other times maybe I'm ok with the slowdown and am willing to ride it out without killing anything. psi-notify sidesteps this problem by simply notifying, rather than taking action.

Config format

update

The update interval in seconds is specified with update [int]. The default is update 5 if unspecified.

log_pressures

If you'd like messages like this at every update interval, you can set log_pressures true (the default is false):

INFO: Current CPU pressures: some avg10=0.00 avg60=0.02 avg300=0.01
INFO: Current memory pressures: some avg10=0.00 avg60=0.00 avg300=0.00
INFO: Current memory pressures: full avg10=0.00 avg60=0.00 avg300=0.00
INFO: Current I/O pressures: some avg10=0.00 avg60=0.00 avg300=0.00
INFO: Current I/O pressures: full avg10=0.00 avg60=0.00 avg300=0.00

threshold

Thresholds are specified with fields in the following format:

1. The word threshold.
2. The resource name, as shown in cgroup.controllers. cpu, memory, and io are currently supported.
3. Whether to use the some or full metric. See the definition here.
4. The PSI time period. avg10, avg60, and avg300 are currently supported.
5. The threshold, as a real number between 0 and 100. Decimals are ok.

Contributing

Issues and pull requests are welcome! Please feel free to file them on GitHub.