xref: /linux/Documentation/accounting/psi.rst (revision d87c25e8f4051f813762da6a182c57f246b17441)
1.. _psi:
2
3================================
4PSI - Pressure Stall Information
5================================
6
7:Date: April, 2018
8:Author: Johannes Weiner <hannes@cmpxchg.org>
9
10When CPU, memory or IO devices are contended, workloads experience
11latency spikes, throughput losses, and run the risk of OOM kills.
12
13Without an accurate measure of such contention, users are forced to
14either play it safe and under-utilize their hardware resources, or
15roll the dice and frequently suffer the disruptions resulting from
16excessive overcommit.
17
18The psi feature identifies and quantifies the disruptions caused by
19such resource crunches and the time impact it has on complex workloads
20or even entire systems.
21
22Having an accurate measure of productivity losses caused by resource
23scarcity aids users in sizing workloads to hardware--or provisioning
24hardware according to workload demand.
25
26As psi aggregates this information in realtime, systems can be managed
27dynamically using techniques such as load shedding, migrating jobs to
28other systems or data centers, or strategically pausing or killing low
29priority or restartable batch jobs.
30
31This allows maximizing hardware utilization without sacrificing
32workload health or risking major disruptions such as OOM kills.
33
34Pressure interface
35==================
36
37Pressure information for each resource is exported through the
38respective file in /proc/pressure/ -- cpu, memory, and io.
39
40The format for CPU is as such::
41
42	some avg10=0.00 avg60=0.00 avg300=0.00 total=0
43
44and for memory and IO::
45
46	some avg10=0.00 avg60=0.00 avg300=0.00 total=0
47	full avg10=0.00 avg60=0.00 avg300=0.00 total=0
48
49The "some" line indicates the share of time in which at least some
50tasks are stalled on a given resource.
51
52The "full" line indicates the share of time in which all non-idle
53tasks are stalled on a given resource simultaneously. In this state
54actual CPU cycles are going to waste, and a workload that spends
55extended time in this state is considered to be thrashing. This has
56severe impact on performance, and it's useful to distinguish this
57situation from a state where some tasks are stalled but the CPU is
58still doing productive work. As such, time spent in this subset of the
59stall state is tracked separately and exported in the "full" averages.
60
61The ratios (in %) are tracked as recent trends over ten, sixty, and
62three hundred second windows, which gives insight into short term events
63as well as medium and long term trends. The total absolute stall time
64(in us) is tracked and exported as well, to allow detection of latency
65spikes which wouldn't necessarily make a dent in the time averages,
66or to average trends over custom time frames.
67
68Monitoring for pressure thresholds
69==================================
70
71Users can register triggers and use poll() to be woken up when resource
72pressure exceeds certain thresholds.
73
74A trigger describes the maximum cumulative stall time over a specific
75time window, e.g. 100ms of total stall time within any 500ms window to
76generate a wakeup event.
77
78To register a trigger user has to open psi interface file under
79/proc/pressure/ representing the resource to be monitored and write the
80desired threshold and time window. The open file descriptor should be
81used to wait for trigger events using select(), poll() or epoll().
82The following format is used::
83
84	<some|full> <stall amount in us> <time window in us>
85
86For example writing "some 150000 1000000" into /proc/pressure/memory
87would add 150ms threshold for partial memory stall measured within
881sec time window. Writing "full 50000 1000000" into /proc/pressure/io
89would add 50ms threshold for full io stall measured within 1sec time window.
90
91Triggers can be set on more than one psi metric and more than one trigger
92for the same psi metric can be specified. However for each trigger a separate
93file descriptor is required to be able to poll it separately from others,
94therefore for each trigger a separate open() syscall should be made even
95when opening the same psi interface file. Write operations to a file descriptor
96with an already existing psi trigger will fail with EBUSY.
97
98Monitors activate only when system enters stall state for the monitored
99psi metric and deactivates upon exit from the stall state. While system is
100in the stall state psi signal growth is monitored at a rate of 10 times per
101tracking window.
102
103The kernel accepts window sizes ranging from 500ms to 10s, therefore min
104monitoring update interval is 50ms and max is 1s. Min limit is set to
105prevent overly frequent polling. Max limit is chosen as a high enough number
106after which monitors are most likely not needed and psi averages can be used
107instead.
108
109When activated, psi monitor stays active for at least the duration of one
110tracking window to avoid repeated activations/deactivations when system is
111bouncing in and out of the stall state.
112
113Notifications to the userspace are rate-limited to one per tracking window.
114
115The trigger will de-register when the file descriptor used to define the
116trigger  is closed.
117
118Userspace monitor usage example
119===============================
120
121::
122
123  #include <errno.h>
124  #include <fcntl.h>
125  #include <stdio.h>
126  #include <poll.h>
127  #include <string.h>
128  #include <unistd.h>
129
130  /*
131   * Monitor memory partial stall with 1s tracking window size
132   * and 150ms threshold.
133   */
134  int main() {
135	const char trig[] = "some 150000 1000000";
136	struct pollfd fds;
137	int n;
138
139	fds.fd = open("/proc/pressure/memory", O_RDWR | O_NONBLOCK);
140	if (fds.fd < 0) {
141		printf("/proc/pressure/memory open error: %s\n",
142			strerror(errno));
143		return 1;
144	}
145	fds.events = POLLPRI;
146
147	if (write(fds.fd, trig, strlen(trig) + 1) < 0) {
148		printf("/proc/pressure/memory write error: %s\n",
149			strerror(errno));
150		return 1;
151	}
152
153	printf("waiting for events...\n");
154	while (1) {
155		n = poll(&fds, 1, -1);
156		if (n < 0) {
157			printf("poll error: %s\n", strerror(errno));
158			return 1;
159		}
160		if (fds.revents & POLLERR) {
161			printf("got POLLERR, event source is gone\n");
162			return 0;
163		}
164		if (fds.revents & POLLPRI) {
165			printf("event triggered!\n");
166		} else {
167			printf("unknown event received: 0x%x\n", fds.revents);
168			return 1;
169		}
170	}
171
172	return 0;
173  }
174
175Cgroup2 interface
176=================
177
178In a system with a CONFIG_CGROUP=y kernel and the cgroup2 filesystem
179mounted, pressure stall information is also tracked for tasks grouped
180into cgroups. Each subdirectory in the cgroupfs mountpoint contains
181cpu.pressure, memory.pressure, and io.pressure files; the format is
182the same as the /proc/pressure/ files.
183
184Per-cgroup psi monitors can be specified and used the same way as
185system-wide ones.
186