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