plugeth/metrics/metrics.go
turboboost55 544e4a700b
metrics: improve accuracy of CPU gauges (#26793)
This PR changes metrics collection to actually measure the time interval between collections, rather
than assume 3 seconds. I did some ad hoc profiling, and on slower hardware (eg, my Raspberry Pi 4)
I routinely saw intervals between 3.3 - 3.5 seconds, with some being as high as 4.5 seconds. This
will generally cause the CPU gauge readings to be too high, and in some cases can cause impossibly
large values for the CPU load metrics (eg. greater than 400 for a 4 core CPU).

---------

Co-authored-by: Felix Lange <fjl@twurst.com>
2023-03-07 00:29:48 +01:00

214 lines
7.8 KiB
Go

// Go port of Coda Hale's Metrics library
//
// <https://github.com/rcrowley/go-metrics>
//
// Coda Hale's original work: <https://github.com/codahale/metrics>
package metrics
import (
"os"
"runtime/metrics"
"runtime/pprof"
"strings"
"time"
"github.com/ethereum/go-ethereum/log"
)
// Enabled is checked by the constructor functions for all of the
// standard metrics. If it is true, the metric returned is a stub.
//
// This global kill-switch helps quantify the observer effect and makes
// for less cluttered pprof profiles.
var Enabled = false
// EnabledExpensive is a soft-flag meant for external packages to check if costly
// metrics gathering is allowed or not. The goal is to separate standard metrics
// for health monitoring and debug metrics that might impact runtime performance.
var EnabledExpensive = false
// enablerFlags is the CLI flag names to use to enable metrics collections.
var enablerFlags = []string{"metrics"}
// expensiveEnablerFlags is the CLI flag names to use to enable metrics collections.
var expensiveEnablerFlags = []string{"metrics.expensive"}
// Init enables or disables the metrics system. Since we need this to run before
// any other code gets to create meters and timers, we'll actually do an ugly hack
// and peek into the command line args for the metrics flag.
func init() {
for _, arg := range os.Args {
flag := strings.TrimLeft(arg, "-")
for _, enabler := range enablerFlags {
if !Enabled && flag == enabler {
log.Info("Enabling metrics collection")
Enabled = true
}
}
for _, enabler := range expensiveEnablerFlags {
if !EnabledExpensive && flag == enabler {
log.Info("Enabling expensive metrics collection")
EnabledExpensive = true
}
}
}
}
var threadCreateProfile = pprof.Lookup("threadcreate")
type runtimeStats struct {
GCPauses *metrics.Float64Histogram
GCAllocBytes uint64
GCFreedBytes uint64
MemTotal uint64
HeapObjects uint64
HeapFree uint64
HeapReleased uint64
HeapUnused uint64
Goroutines uint64
SchedLatency *metrics.Float64Histogram
}
var runtimeSamples = []metrics.Sample{
{Name: "/gc/pauses:seconds"}, // histogram
{Name: "/gc/heap/allocs:bytes"},
{Name: "/gc/heap/frees:bytes"},
{Name: "/memory/classes/total:bytes"},
{Name: "/memory/classes/heap/objects:bytes"},
{Name: "/memory/classes/heap/free:bytes"},
{Name: "/memory/classes/heap/released:bytes"},
{Name: "/memory/classes/heap/unused:bytes"},
{Name: "/sched/goroutines:goroutines"},
{Name: "/sched/latencies:seconds"}, // histogram
}
func readRuntimeStats(v *runtimeStats) {
metrics.Read(runtimeSamples)
for _, s := range runtimeSamples {
// Skip invalid/unknown metrics. This is needed because some metrics
// are unavailable in older Go versions, and attempting to read a 'bad'
// metric panics.
if s.Value.Kind() == metrics.KindBad {
continue
}
switch s.Name {
case "/gc/pauses:seconds":
v.GCPauses = s.Value.Float64Histogram()
case "/gc/heap/allocs:bytes":
v.GCAllocBytes = s.Value.Uint64()
case "/gc/heap/frees:bytes":
v.GCFreedBytes = s.Value.Uint64()
case "/memory/classes/total:bytes":
v.MemTotal = s.Value.Uint64()
case "/memory/classes/heap/objects:bytes":
v.HeapObjects = s.Value.Uint64()
case "/memory/classes/heap/free:bytes":
v.HeapFree = s.Value.Uint64()
case "/memory/classes/heap/released:bytes":
v.HeapReleased = s.Value.Uint64()
case "/memory/classes/heap/unused:bytes":
v.HeapUnused = s.Value.Uint64()
case "/sched/goroutines:goroutines":
v.Goroutines = s.Value.Uint64()
case "/sched/latencies:seconds":
v.SchedLatency = s.Value.Float64Histogram()
}
}
}
// CollectProcessMetrics periodically collects various metrics about the running process.
func CollectProcessMetrics(refresh time.Duration) {
// Short circuit if the metrics system is disabled
if !Enabled {
return
}
// Create the various data collectors
var (
cpustats = make([]CPUStats, 2)
diskstats = make([]DiskStats, 2)
rstats = make([]runtimeStats, 2)
)
// This scale factor is used for the runtime's time metrics. It's useful to convert to
// ns here because the runtime gives times in float seconds, but runtimeHistogram can
// only provide integers for the minimum and maximum values.
const secondsToNs = float64(time.Second)
// Define the various metrics to collect
var (
cpuSysLoad = GetOrRegisterGauge("system/cpu/sysload", DefaultRegistry)
cpuSysWait = GetOrRegisterGauge("system/cpu/syswait", DefaultRegistry)
cpuProcLoad = GetOrRegisterGauge("system/cpu/procload", DefaultRegistry)
cpuThreads = GetOrRegisterGauge("system/cpu/threads", DefaultRegistry)
cpuGoroutines = GetOrRegisterGauge("system/cpu/goroutines", DefaultRegistry)
cpuSchedLatency = getOrRegisterRuntimeHistogram("system/cpu/schedlatency", secondsToNs, nil)
memPauses = getOrRegisterRuntimeHistogram("system/memory/pauses", secondsToNs, nil)
memAllocs = GetOrRegisterMeter("system/memory/allocs", DefaultRegistry)
memFrees = GetOrRegisterMeter("system/memory/frees", DefaultRegistry)
memTotal = GetOrRegisterGauge("system/memory/held", DefaultRegistry)
heapUsed = GetOrRegisterGauge("system/memory/used", DefaultRegistry)
heapObjects = GetOrRegisterGauge("system/memory/objects", DefaultRegistry)
diskReads = GetOrRegisterMeter("system/disk/readcount", DefaultRegistry)
diskReadBytes = GetOrRegisterMeter("system/disk/readdata", DefaultRegistry)
diskReadBytesCounter = GetOrRegisterCounter("system/disk/readbytes", DefaultRegistry)
diskWrites = GetOrRegisterMeter("system/disk/writecount", DefaultRegistry)
diskWriteBytes = GetOrRegisterMeter("system/disk/writedata", DefaultRegistry)
diskWriteBytesCounter = GetOrRegisterCounter("system/disk/writebytes", DefaultRegistry)
)
var lastCollectTime time.Time
// Iterate loading the different stats and updating the meters.
now, prev := 0, 1
for ; ; now, prev = prev, now {
// Gather CPU times.
ReadCPUStats(&cpustats[now])
collectTime := time.Now()
secondsSinceLastCollect := collectTime.Sub(lastCollectTime).Seconds()
lastCollectTime = collectTime
if secondsSinceLastCollect > 0 {
sysLoad := (cpustats[now].GlobalTime - cpustats[prev].GlobalTime) / secondsSinceLastCollect
sysWait := (cpustats[now].GlobalWait - cpustats[prev].GlobalWait) / secondsSinceLastCollect
procLoad := (cpustats[now].LocalTime - cpustats[prev].LocalTime) / secondsSinceLastCollect
// Convert to integer percentage.
cpuSysLoad.Update(int64(sysLoad * 100))
cpuSysWait.Update(int64(sysWait * 100))
cpuProcLoad.Update(int64(procLoad * 100))
}
// Threads
cpuThreads.Update(int64(threadCreateProfile.Count()))
// Go runtime metrics
readRuntimeStats(&rstats[now])
cpuGoroutines.Update(int64(rstats[now].Goroutines))
cpuSchedLatency.update(rstats[now].SchedLatency)
memPauses.update(rstats[now].GCPauses)
memAllocs.Mark(int64(rstats[now].GCAllocBytes - rstats[prev].GCAllocBytes))
memFrees.Mark(int64(rstats[now].GCFreedBytes - rstats[prev].GCFreedBytes))
memTotal.Update(int64(rstats[now].MemTotal))
heapUsed.Update(int64(rstats[now].MemTotal - rstats[now].HeapUnused - rstats[now].HeapFree - rstats[now].HeapReleased))
heapObjects.Update(int64(rstats[now].HeapObjects))
// Disk
if ReadDiskStats(&diskstats[now]) == nil {
diskReads.Mark(diskstats[now].ReadCount - diskstats[prev].ReadCount)
diskReadBytes.Mark(diskstats[now].ReadBytes - diskstats[prev].ReadBytes)
diskWrites.Mark(diskstats[now].WriteCount - diskstats[prev].WriteCount)
diskWriteBytes.Mark(diskstats[now].WriteBytes - diskstats[prev].WriteBytes)
diskReadBytesCounter.Inc(diskstats[now].ReadBytes - diskstats[prev].ReadBytes)
diskWriteBytesCounter.Inc(diskstats[now].WriteBytes - diskstats[prev].WriteBytes)
}
time.Sleep(refresh)
}
}