summaryrefslogtreecommitdiffstats
path: root/Documentation/accounting/psi.rst
blob: f2b3439edcc2cc772c7f92cd3510060df6b5e8b8 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
.. _psi:

================================
PSI - Pressure Stall Information
================================

:Date: April, 2018
:Author: Johannes Weiner <hannes@cmpxchg.org>

When CPU, memory or IO devices are contended, workloads experience
latency spikes, throughput losses, and run the risk of OOM kills.

Without an accurate measure of such contention, users are forced to
either play it safe and under-utilize their hardware resources, or
roll the dice and frequently suffer the disruptions resulting from
excessive overcommit.

The psi feature identifies and quantifies the disruptions caused by
such resource crunches and the time impact it has on complex workloads
or even entire systems.

Having an accurate measure of productivity losses caused by resource
scarcity aids users in sizing workloads to hardware--or provisioning
hardware according to workload demand.

As psi aggregates this information in realtime, systems can be managed
dynamically using techniques such as load shedding, migrating jobs to
other systems or data centers, or strategically pausing or killing low
priority or restartable batch jobs.

This allows maximizing hardware utilization without sacrificing
workload health or risking major disruptions such as OOM kills.

Pressure interface
==================

Pressure information for each resource is exported through the
respective file in /proc/pressure/ -- cpu, memory, and io.

The format for CPU is as such::

	some avg10=0.00 avg60=0.00 avg300=0.00 total=0

and for memory and IO::

	some avg10=0.00 avg60=0.00 avg300=0.00 total=0
	full avg10=0.00 avg60=0.00 avg300=0.00 total=0

The "some" line indicates the share of time in which at least some
tasks are stalled on a given resource.

The "full" line indicates the share of time in which all non-idle
tasks are stalled on a given resource simultaneously. In this state
actual CPU cycles are going to waste, and a workload that spends
extended time in this state is considered to be thrashing. This has
severe impact on performance, and it's useful to distinguish this
situation from a state where some tasks are stalled but the CPU is
still doing productive work. As such, time spent in this subset of the
stall state is tracked separately and exported in the "full" averages.

The ratios (in %) are tracked as recent trends over ten, sixty, and
three hundred second windows, which gives insight into short term events
as well as medium and long term trends. The total absolute stall time
(in us) is tracked and exported as well, to allow detection of latency
spikes which wouldn't necessarily make a dent in the time averages,
or to average trends over custom time frames.

Monitoring for pressure thresholds
==================================

Users can register triggers and use poll() to be woken up when resource
pressure exceeds certain thresholds.

A trigger describes the maximum cumulative stall time over a specific
time window, e.g. 100ms of total stall time within any 500ms window to
generate a wakeup event.

To register a trigger user has to open psi interface file under
/proc/pressure/ representing the resource to be monitored and write the
desired threshold and time window. The open file descriptor should be
used to wait for trigger events using select(), poll() or epoll().
The following format is used::

	<some|full> <stall amount in us> <time window in us>

For example writing "some 150000 1000000" into /proc/pressure/memory
would add 150ms threshold for partial memory stall measured within
1sec time window. Writing "full 50000 1000000" into /proc/pressure/io
would add 50ms threshold for full io stall measured within 1sec time window.

Triggers can be set on more than one psi metric and more than one trigger
for the same psi metric can be specified. However for each trigger a separate
file descriptor is required to be able to poll it separately from others,
therefore for each trigger a separate open() syscall should be made even
when opening the same psi interface file.

Monitors activate only when system enters stall state for the monitored
psi metric and deactivates upon exit from the stall state. While system is
in the stall state psi signal growth is monitored at a rate of 10 times per
tracking window.

The kernel accepts window sizes ranging from 500ms to 10s, therefore min
monitoring update interval is 50ms and max is 1s. Min limit is set to
prevent overly frequent polling. Max limit is chosen as a high enough number
after which monitors are most likely not needed and psi averages can be used
instead.

When activated, psi monitor stays active for at least the duration of one
tracking window to avoid repeated activations/deactivations when system is
bouncing in and out of the stall state.

Notifications to the userspace are rate-limited to one per tracking window.

The trigger will de-register when the file descriptor used to define the
trigger  is closed.

Userspace monitor usage example
===============================

::

  #include <errno.h>
  #include <fcntl.h>
  #include <stdio.h>
  #include <poll.h>
  #include <string.h>
  #include <unistd.h>

  /*
   * Monitor memory partial stall with 1s tracking window size
   * and 150ms threshold.
   */
  int main() {
	const char trig[] = "some 150000 1000000";
	struct pollfd fds;
	int n;

	fds.fd = open("/proc/pressure/memory", O_RDWR | O_NONBLOCK);
	if (fds.fd < 0) {
		printf("/proc/pressure/memory open error: %s\n",
			strerror(errno));
		return 1;
	}
	fds.events = POLLPRI;

	if (write(fds.fd, trig, strlen(trig) + 1) < 0) {
		printf("/proc/pressure/memory write error: %s\n",
			strerror(errno));
		return 1;
	}

	printf("waiting for events...\n");
	while (1) {
		n = poll(&fds, 1, -1);
		if (n < 0) {
			printf("poll error: %s\n", strerror(errno));
			return 1;
		}
		if (fds.revents & POLLERR) {
			printf("got POLLERR, event source is gone\n");
			return 0;
		}
		if (fds.revents & POLLPRI) {
			printf("event triggered!\n");
		} else {
			printf("unknown event received: 0x%x\n", fds.revents);
			return 1;
		}
	}

	return 0;
  }

Cgroup2 interface
=================

In a system with a CONFIG_CGROUP=y kernel and the cgroup2 filesystem
mounted, pressure stall information is also tracked for tasks grouped
into cgroups. Each subdirectory in the cgroupfs mountpoint contains
cpu.pressure, memory.pressure, and io.pressure files; the format is
the same as the /proc/pressure/ files.

Per-cgroup psi monitors can be specified and used the same way as
system-wide ones.