add per-container Pressure Stall Information (PSI) metrics

cgroup/cpu.go

@@ -16,7 +16,7 @@ type CPUStat struct {
 	LimitCores           float64
 }
 
-func (cg Cgroup) CpuStat() *CPUStat {
+func (cg *Cgroup) CpuStat() *CPUStat {
 	cpu, cpuacct := cg.subsystems["cpu"], cg.subsystems["cpuacct"]
 	if cpu == "" || cpuacct == "" {
 		st, _ := cg.cpuStatV2()
@@ -26,7 +26,7 @@ func (cg Cgroup) CpuStat() *CPUStat {
 	return st
 }
 
-func (cg Cgroup) cpuStatV1() (*CPUStat, error) {
+func (cg *Cgroup) cpuStatV1() (*CPUStat, error) {
 	if cg.subsystems["cpu"] == "" || cg.subsystems["cpuacct"] == "" {
 		return nil, nil
 	}
@@ -56,7 +56,7 @@ func (cg Cgroup) cpuStatV1() (*CPUStat, error) {
 	return res, nil
 }
 
-func (cg Cgroup) cpuStatV2() (*CPUStat, error) {
+func (cg *Cgroup) cpuStatV2() (*CPUStat, error) {
 	if cg.subsystems[""] == "" {
 		return nil, nil
 	}

cgroup/fixtures/cgroup/kubepods.slice/kubepods-besteffort.slice/kubepods-besteffort-pod8712f785_1a3e_41ec_a00b_e2dcc77431cb.slice/docker-73051af271105c07e1f493b34856a77e665e3b0b4fc72f76c807dfbffeb881bd.scope/cpu.pressure

@@ -0,0 +1,2 @@
+some avg10=0.00 avg60=0.00 avg300=0.00 total=465907442
+full avg10=0.00 avg60=0.00 avg300=0.00 total=463529433

cgroup/fixtures/cgroup/kubepods.slice/kubepods-besteffort.slice/kubepods-besteffort-pod8712f785_1a3e_41ec_a00b_e2dcc77431cb.slice/docker-73051af271105c07e1f493b34856a77e665e3b0b4fc72f76c807dfbffeb881bd.scope/io.pressure

@@ -0,0 +1,2 @@
+some avg10=0.00 avg60=0.00 avg300=0.05 total=17657662684
+full avg10=0.00 avg60=0.00 avg300=0.05 total=17636951020

cgroup/fixtures/cgroup/kubepods.slice/kubepods-besteffort.slice/kubepods-besteffort-pod8712f785_1a3e_41ec_a00b_e2dcc77431cb.slice/docker-73051af271105c07e1f493b34856a77e665e3b0b4fc72f76c807dfbffeb881bd.scope/memory.pressure

@@ -0,0 +1,2 @@
+some avg10=0.00 avg60=0.00 avg300=0.00 total=6937313991
+full avg10=0.00 avg60=0.00 avg300=0.00 total=6934649214

cgroup/psi.go

@@ -0,0 +1,85 @@
+package cgroup
+
+import (
+	"os"
+	"path"
+	"strconv"
+	"strings"
+
+	"github.com/coroot/coroot-node-agent/common"
+	"k8s.io/klog/v2"
+)
+
+type PSIStats struct {
+	CPUSecondsSome    float64
+	CPUSecondsFull    float64
+	MemorySecondsSome float64
+	MemorySecondsFull float64
+	IOSecondsSome     float64
+	IOSecondsFull     float64
+}
+
+type PressureTotals struct {
+	SomeSecondsTotal float64
+	FullSecondsTotal float64
+}
+
+func (cg *Cgroup) PSI() *PSIStats {
+	if cg.subsystems[""] == "" {
+		return nil
+	}
+	stats := &PSIStats{}
+	for _, controller := range []string{"cpu", "memory", "io"} {
+		p, err := cg.readPressure(controller)
+		if err != nil {
+			if !common.IsNotExist(err) {
+				klog.Warningln(err)
+			}
+			return nil
+		}
+		switch controller {
+		case "cpu":
+			stats.CPUSecondsSome = p.SomeSecondsTotal
+			stats.CPUSecondsFull = p.FullSecondsTotal
+		case "memory":
+			stats.MemorySecondsSome = p.SomeSecondsTotal
+			stats.MemorySecondsFull = p.FullSecondsTotal
+		case "io":
+			stats.IOSecondsSome = p.SomeSecondsTotal
+			stats.IOSecondsFull = p.FullSecondsTotal
+		}
+	}
+	return stats
+}
+
+func (cg *Cgroup) readPressure(controller string) (*PressureTotals, error) {
+	data, err := os.ReadFile(path.Join(cg2Root, cg.subsystems[""], controller+".pressure"))
+	if err != nil {
+		return nil, err
+	}
+	pressure := &PressureTotals{}
+	for _, line := range strings.Split(strings.TrimSpace(string(data)), "\n") {
+		parts := strings.Fields(line)
+		if len(parts) == 0 {
+			continue
+		}
+		kind := parts[0]
+		for _, p := range parts[1:] {
+			if strings.HasPrefix(p, "total=") {
+				vStr := strings.TrimPrefix(p, "total=")
+				v, err := strconv.ParseUint(vStr, 10, 64)
+				if err != nil {
+					return nil, err
+				}
+				switch kind {
+				case "some":
+					pressure.SomeSecondsTotal = float64(v) / 1e6 // microseconds to seconds
+				case "full":
+					pressure.FullSecondsTotal = float64(v) / 1e6
+				}
+				break
+			}
+		}
+	}
+	return pressure, nil
+}

cgroup/psi_test.go

@@ -0,0 +1,27 @@
+package cgroup
+
+import (
+	"path"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestCgroupPSI(t *testing.T) {
+	cgRoot = "fixtures/cgroup"
+	cg2Root = "fixtures/cgroup"
+
+	cg, _ := NewFromProcessCgroupFile(path.Join("fixtures/proc/400/cgroup"))
+	stat := cg.PSI()
+	require.NotNil(t, stat)
+	assert.Equal(t, float64(465907442)/1e6, stat.CPUSecondsSome)
+	assert.Equal(t, float64(463529433)/1e6, stat.CPUSecondsFull)
+	assert.Equal(t, float64(6937313991)/1e6, stat.MemorySecondsSome)
+	assert.Equal(t, float64(6934649214)/1e6, stat.MemorySecondsFull)
+	assert.Equal(t, float64(17657662684)/1e6, stat.IOSecondsSome)
+	assert.Equal(t, float64(17636951020)/1e6, stat.IOSecondsFull)
+
+	cg, _ = NewFromProcessCgroupFile(path.Join("fixtures/proc/100/cgroup"))
+	assert.Nil(t, cg.PSI())
+}

containers/container.go

@@ -265,6 +265,15 @@ func (c *Container) Collect(ch chan<- prometheus.Metric) {
 		}
 	}
 
+	if psi := c.cgroup.PSI(); psi != nil {
+		ch <- counter(metrics.PsiCPU, psi.CPUSecondsSome, "some")
+		ch <- counter(metrics.PsiCPU, psi.CPUSecondsFull, "full")
+		ch <- counter(metrics.PsiMemory, psi.MemorySecondsSome, "some")
+		ch <- counter(metrics.PsiMemory, psi.MemorySecondsFull, "full")
+		ch <- counter(metrics.PsiIO, psi.IOSecondsSome, "some")
+		ch <- counter(metrics.PsiIO, psi.IOSecondsFull, "full")
+	}
+
 	if c.oomKills > 0 {
 		ch <- counter(metrics.OOMKills, float64(c.oomKills))
 	}

containers/metrics.go

@@ -19,6 +19,10 @@ var metrics = struct {
 	MemoryCache *prometheus.Desc
 	OOMKills    *prometheus.Desc
 
+	PsiCPU    *prometheus.Desc
+	PsiMemory *prometheus.Desc
+	PsiIO     *prometheus.Desc
+
 	DiskDelay      *prometheus.Desc
 	DiskSize       *prometheus.Desc
 	DiskUsed       *prometheus.Desc
@@ -71,6 +75,10 @@ var metrics = struct {
 	MemoryCache: metric("container_resources_memory_cache_bytes", "Amount of page cache memory allocated by the container"),
 	OOMKills:    metric("container_oom_kills_total", "Total number of times the container was terminated by the OOM killer"),
 
+	PsiCPU:    metric("container_resources_cpu_pressure_waiting_seconds_total", "Total time in seconds tha the container were delayed due to CPU pressure", "kind"),
+	PsiMemory: metric("container_resources_memory_pressure_waiting_seconds_total", "Total time in seconds that the container were delayed due to memory pressure", "kind"),
+	PsiIO:     metric("container_resources_io_pressure_waiting_seconds_total", "Total time in seconds that the container were delayed due to I/O pressure", "kind"),
+
 	DiskDelay:      metric("container_resources_disk_delay_seconds_total", "Total time duration processes of the container have been waiting fot I/Os to complete"),
 	DiskSize:       metric("container_resources_disk_size_bytes", "Total capacity of the volume", "mount_point", "device", "volume"),
 	DiskUsed:       metric("container_resources_disk_used_bytes", "Used capacity of the volume", "mount_point", "device", "volume"),