add more dcgm metrics

Author: atelpis

cmd/do-agent/aggregation.go

@@ -115,6 +115,15 @@ var gpuAggregationSpec = map[string][]string{
 	"dcgm_fi_dev_fb_free":                nvidiaAggregatedLabels,
 	"dcgm_fi_dev_fb_used":                nvidiaAggregatedLabels,
 	"dcgm_fi_dev_nvlink_bandwidth_total": nvidiaAggregatedLabels,
+	"dcgm_fi_prof_sm_occupancy":          nvidiaAggregatedLabels,
+	"dcgm_fi_prof_pipe_tensor_active":    nvidiaAggregatedLabels,
+	"dcgm_fi_prof_pcie_rx_bytes":         nvidiaAggregatedLabels,
+	"dcgm_fi_prof_pcie_tx_bytes":         nvidiaAggregatedLabels,
+	"dcgm_fi_dev_gpu_temp":               nvidiaAggregatedLabels,
+	"dcgm_fi_dev_memory_temp":            nvidiaAggregatedLabels,
+	"dcgm_fi_dev_power_usage":            nvidiaAggregatedLabels,
+	"dcgm_fi_dev_power_violation":        nvidiaAggregatedLabels,
+	"dcgm_fi_dev_thermal_violation":      nvidiaAggregatedLabels,
 
 	// GPU Utilization metrics
 	"amd_gpu_prof_gui_util_percent":      amdAggregatedLabels,

cmd/do-agent/whitelist.go

@@ -170,6 +170,15 @@ var gpuWhitelist = map[string]bool{
 	"DCGM_FI_DEV_FB_FREE":                true,
 	"DCGM_FI_DEV_FB_USED":                true,
 	"DCGM_FI_DEV_NVLINK_BANDWIDTH_TOTAL": true,
+	"DCGM_FI_PROF_SM_OCCUPANCY":          true,
+	"DCGM_FI_PROF_PIPE_TENSOR_ACTIVE":    true,
+	"DCGM_FI_PROF_PCIE_RX_BYTES":         true,
+	"DCGM_FI_PROF_PCIE_TX_BYTES":         true,
+	"DCGM_FI_DEV_GPU_TEMP":               true,
+	"DCGM_FI_DEV_MEMORY_TEMP":            true,
+	"DCGM_FI_DEV_POWER_USAGE":            true,
+	"DCGM_FI_DEV_POWER_VIOLATION":        true,
+	"DCGM_FI_DEV_THERMAL_VIOLATION":      true,
 
 	// GPU Utilization
 	"amd_gpu_prof_gui_util_percent":      true,