Pressure Stall Information

Pressure stall information (PSI) is a Linux kernel metric for the share of time tasks spent stalled waiting on a resource: CPU, memory, or I/O. It appears in /proc/pressure/{cpu,memory,io} and per cgroup, as running averages over 10, 60, and 300 seconds.

Each resource reports two lines:

• some: the share of time at least one task was stalled on that resource. Some work was being delayed.
• full: the share of time all non-idle tasks were stalled at once. No useful work was happening at all.

PSI exists because the traditional figures answer the wrong question. CPU utilization says how busy a resource is, not whether anyone is waiting for it, and 100% busy with nobody waiting is a well-used machine, not a problem. Load average mixes CPU and I/O waiters into one number, lags by minutes, and never says which resource is short. Free memory says almost nothing on a system where unused RAM holds the Page Cache by design. PSI measures the thing that actually hurts, lost time, per resource. A memory "some" figure climbing while free memory looks fine is how Swap thrashing shows itself before anything is killed.

That makes PSI a good trigger. Userspace OOM daemons act on sustained memory pressure instead of waiting for the kernel's last resort, and services can shed load or apply Backpressure when their own cgroup's pressure rises. It is a signal of saturation rather than activity, which is what alerting should be built on.

Like any single metric, it needs cross-checking. The I/O figure counts time tasks spend in certain kinds of wait, which does not always correspond to device work: a reading can be high on a machine whose disks show no throughput at all. Before acting on a surprising number, confirm it against something independent, such as per-device throughput, per-process stall states, or whether anything actually feels slow. A metric that disagrees with every other piece of evidence is evidence about the metric; see Ground Truth.