Add Intel sysfs GPU power collector (Xe/i915 hwmon energy counters) (#2020)

This commit is contained in:
T.J. Tarazevits
2026-08-16 10:20:43 -05:00
committed by GitHub
parent b4e1f3fafe
commit 3688b2d033
5 changed files with 532 additions and 4 deletions

View File

@@ -48,6 +48,8 @@ type GPUManager struct {
// Per-cache-key tracking for delta calculations
// cacheKey -> gpuId -> snapshot of last count/usage/power values
lastSnapshots map[uint16]map[string]*gpuSnapshot
// Per-card energy snapshots for Intel sysfs power calculation.
intelSysfsEnergySnapshots map[string]intelSysfsEnergySnapshot
}
// gpuSnapshot stores the last observed incremental values for delta tracking
@@ -90,6 +92,7 @@ const (
collectorSourceNVML collectorSource = "nvml"
collectorSourceNvidiaSMI collectorSource = collectorSource(nvidiaSmiCmd)
collectorSourceIntelGpuTop collectorSource = collectorSource(intelGpuStatsCmd)
collectorSourceIntelSysfs collectorSource = "intel_sysfs"
collectorSourceAmdSysfs collectorSource = "amd_sysfs"
collectorSourceRocmSMI collectorSource = collectorSource(rocmSmiCmd)
collectorSourceMacmon collectorSource = collectorSource(macmonCmd)
@@ -106,6 +109,7 @@ func isValidCollectorSource(source collectorSource) bool {
collectorSourceNVML,
collectorSourceNvidiaSMI,
collectorSourceIntelGpuTop,
collectorSourceIntelSysfs,
collectorSourceAmdSysfs,
collectorSourceRocmSMI,
collectorSourceMacmon,
@@ -122,6 +126,7 @@ type gpuCapabilities struct {
hasAmdSysfs bool
hasTegrastats bool
hasIntelGpuTop bool
hasIntelSysfs bool
hasNvtop bool
hasMacmon bool
hasPowermetrics bool
@@ -369,12 +374,13 @@ func (gm *GPUManager) calculateGPUAverage(id string, gpu *system.GPUData, cacheK
gpuAvg.Power = utils.TwoDecimals(deltaPower / float64(deltaCount))
gpuAvg.PowerPkg = utils.TwoDecimals(deltaPowerPkg / float64(deltaCount))
if gpu.Engines != nil {
// make fresh map for averaged engine metrics to avoid mutating
// the accumulator map stored in gm.GpuDataMap
gpuAvg.Engines = make(map[string]float64, len(gpu.Engines))
gpuAvg.Usage = gm.calculateIntelGPUUsage(&gpuAvg, gpu, lastSnapshot, deltaCount)
gpuAvg.PowerPkg = utils.TwoDecimals(deltaPowerPkg / float64(deltaCount))
} else {
gpuAvg.Usage = utils.TwoDecimals(deltaUsage / float64(deltaCount))
}
@@ -443,7 +449,8 @@ func (gm *GPUManager) storeSnapshot(id string, gpu *system.GPUData, cacheKey uin
// It only reports capability presence and does not apply policy decisions.
func (gm *GPUManager) discoverGpuCapabilities() gpuCapabilities {
caps := gpuCapabilities{
hasAmdSysfs: gm.hasAmdSysfs(),
hasAmdSysfs: gm.hasAmdSysfs(),
hasIntelSysfs: gm.hasIntelSysfs(),
}
if _, err := exec.LookPath(nvidiaSmiCmd); err == nil {
caps.hasNvidiaSmi = true
@@ -472,7 +479,7 @@ func (gm *GPUManager) discoverGpuCapabilities() gpuCapabilities {
}
func hasAnyGpuCollector(caps gpuCapabilities) bool {
return caps.hasNvidiaSmi || caps.hasRocmSmi || caps.hasAmdSysfs || caps.hasTegrastats || caps.hasIntelGpuTop || caps.hasNvtop || caps.hasMacmon || caps.hasPowermetrics
return caps.hasNvidiaSmi || caps.hasRocmSmi || caps.hasAmdSysfs || caps.hasTegrastats || caps.hasIntelGpuTop || caps.hasIntelSysfs || caps.hasNvtop || caps.hasMacmon || caps.hasPowermetrics
}
func (gm *GPUManager) startIntelCollector() {
@@ -563,6 +570,13 @@ func (gm *GPUManager) collectorDefinitions(caps gpuCapabilities) map[collectorSo
return true
},
},
collectorSourceIntelSysfs: {
group: collectorGroupIntel,
available: caps.hasIntelSysfs,
start: func(_ func()) bool {
return gm.startIntelSysfsCollector()
},
},
collectorSourceAmdSysfs: {
group: collectorGroupAmd,
available: caps.hasAmdSysfs,
@@ -708,6 +722,9 @@ func (gm *GPUManager) resolveLegacyCollectorPriority(caps gpuCapabilities) []col
if caps.hasIntelGpuTop {
priorities = append(priorities, collectorSourceIntelGpuTop)
}
if caps.hasIntelSysfs {
priorities = append(priorities, collectorSourceIntelSysfs)
}
// Apple collectors are currently opt-in only for testing.
// Enable them with GPU_COLLECTOR=macmon or GPU_COLLECTOR=powermetrics.