Use proper DCGM metric for uncorrectable errors (#341)

Author: atelpis

cmd/do-agent/aggregation.go

@@ -125,7 +125,7 @@ var gpuAggregationSpec = map[string][]string{
 	"dcgm_fi_dev_power_usage":            nvidiaAggregatedLabels,
 	"dcgm_fi_dev_power_violation":        nvidiaAggregatedLabels,
 	"dcgm_fi_dev_thermal_violation":      nvidiaAggregatedLabels,
-	"dcgm_fi_dev_ecc_sbe_agg_total":      nvidiaAggregatedLabels,
+	"dcgm_fi_dev_ecc_dbe_agg_total":      nvidiaAggregatedLabels,
 
 	// GPU Utilization metrics
 	"amd_gpu_prof_gui_util_percent":      amdAggregatedLabels,

cmd/do-agent/whitelist.go

@@ -179,7 +179,7 @@ var gpuWhitelist = map[string]bool{
 	"DCGM_FI_DEV_POWER_USAGE":            true,
 	"DCGM_FI_DEV_POWER_VIOLATION":        true,
 	"DCGM_FI_DEV_THERMAL_VIOLATION":      true,
-	"DCGM_FI_DEV_ECC_SBE_AGG_TOTAL":      true,
+	"DCGM_FI_DEV_ECC_DBE_AGG_TOTAL":      true,
 
 	// GPU Utilization
 	"amd_gpu_prof_gui_util_percent":      true,