Accommodate metric and label renames.

GoogleCloudPlatform · Jun 25, 2024 · da2011d · da2011d
1 parent e384d80
commit da2011d
Show file tree

Hide file tree

Showing 6 changed files with 228 additions and 27 deletions.
diff --git a/apps/dcgm.go b/apps/dcgm.go
@@ -52,6 +52,36 @@ func (r MetricsReceiverDcgm) Pipelines(_ context.Context) ([]otel.ReceiverPipeli
 				},
 			},
 			Processors: map[string][]otel.Component{"metrics": {
+				otel.MetricsTransform(
+					otel.UpdateMetric(
+						"gpu.dcgm.pipe.utilization",
+						otel.RenameLabel("gpu.pipe", "pipe"),
+					),
+					otel.UpdateMetric(
+						"gpu.dcgm.memory.bytes_used",
+						otel.RenameLabel("gpu.memory.state", "state"),
+					),
+					otel.UpdateMetric(
+						"gpu.dcgm.nvlink.io",
+						otel.RenameLabel("network.io.direction", "direction"),
+					),
+					otel.UpdateMetric(
+						"gpu.dcgm.pcie.io",
+						otel.RenameLabel("network.io.direction", "direction"),
+					),
+					otel.UpdateMetric(
+						"gpu.dcgm.clock.throttle_duration.time",
+						otel.RenameLabel("gpu.clock.violation", "violation"),
+					),
+					otel.UpdateMetric(
+						"gpu.dcgm.ecc_errors",
+						otel.RenameLabel("gpu.error.type", "error_type"),
+					),
+					otel.UpdateMetric(
+						"gpu.dcgm.xid_errors",
+						otel.RenameLabel("gpu.error.xid", "xid"),
+					),
+				),
 				otel.MetricsTransform(
 					otel.AddPrefix("workload.googleapis.com"),
 				),
@@ -72,31 +102,49 @@ func (r MetricsReceiverDcgm) Pipelines(_ context.Context) ([]otel.ReceiverPipeli
 				"collection_interval": r.CollectionIntervalString(),
 				"endpoint":            r.Endpoint,
 				"metrics": map[string]interface{}{
-					"gpu.dcgm.clock.frequency": map[string]bool{
+					"gpu.dcgm.utilization": map[string]bool{
 						"enabled": false,
 					},
-					"gpu.dcgm.clock.throttle_duration.time": map[string]bool{
+					"gpu.dcgm.sm.utilization": map[string]bool{
+						"enabled": true,
+					},
+					"gpu.dcgm.sm.occupancy": map[string]bool{
+						"enabled": true,
+					},
+					"gpu.dcgm.pipe.utilization": map[string]bool{
+						"enabled": true,
+					},
+					"gpu.dcgm.codec.encoder.utilization": map[string]bool{
 						"enabled": false,
 					},
 					"gpu.dcgm.codec.decoder.utilization": map[string]bool{
 						"enabled": false,
 					},
-					"gpu.dcgm.codec.encoder.utilization": map[string]bool{
+					"gpu.dcgm.memory.bytes_used": map[string]bool{
 						"enabled": false,
 					},
-					"gpu.dcgm.ecc_errors": map[string]bool{
-						"enabled": false,
+					"gpu.dcgm.memory.bandwidth_utilization": map[string]bool{
+						"enabled": true,
+					},
+					"gpu.dcgm.pcie.io": map[string]bool{
+						"enabled": true,
+					},
+					"gpu.dcgm.nvlink.io": map[string]bool{
+						"enabled": true,
 					},
 					"gpu.dcgm.energy_consumption": map[string]bool{
 						"enabled": false,
 					},
-					"gpu.dcgm.memory.bytes_used": map[string]bool{
+					"gpu.dcgm.temperature": map[string]bool{
 						"enabled": false,
 					},
-					"gpu.dcgm.temperature": map[string]bool{
+					"gpu.dcgm.clock.frequency": map[string]bool{
 						"enabled": false,
 					},
-					"gpu.dcgm.utilization": map[string]bool{
+					"gpu.dcgm.clock.throttle_duration.time": map[string]bool{
+						"enabled": false,
+					},
+					"gpu.dcgm.ecc_errors": map[string]bool{
 						"enabled": false,
 					},
 					"gpu.dcgm.xid_errors": map[string]bool{
@@ -112,16 +160,27 @@ func (r MetricsReceiverDcgm) Pipelines(_ context.Context) ([]otel.ReceiverPipeli
 					"dcgm.gpu.profiling.dram_utilization",
 				),
 				otel.RenameMetric(
-					"gpu.dcgm.nvlink.traffic",
+					"gpu.dcgm.nvlink.io",
 					"dcgm.gpu.profiling.nvlink_traffic_rate",
+					otel.RenameLabel("network.io.direction", "direction"),
+					otel.RenameLabelValues("direction", map[string]string{
+						"receive":  "rx",
+						"transmit": "tx",
+					}),
 				),
 				otel.RenameMetric(
-					"gpu.dcgm.pcie.traffic",
+					"gpu.dcgm.pcie.io",
 					"dcgm.gpu.profiling.pcie_traffic_rate",
+					otel.RenameLabel("network.io.direction", "direction"),
+					otel.RenameLabelValues("direction", map[string]string{
+						"receive":  "rx",
+						"transmit": "tx",
+					}),
 				),
 				otel.RenameMetric(
 					"gpu.dcgm.pipe.utilization",
 					"dcgm.gpu.profiling.pipe_utilization",
+					otel.RenameLabel("gpu.pipe", "pipe"),
 				),
 				otel.RenameMetric(
 					"gpu.dcgm.sm.occupancy",

diff --git a/confgenerator/testdata/goldens/metrics-receiver_dcgm/golden/linux-gpu/otel.yaml b/confgenerator/testdata/goldens/metrics-receiver_dcgm/golden/linux-gpu/otel.yaml
@@ -67,14 +67,40 @@ processors:
       include: gpu.dcgm.memory.bandwidth_utilization
       new_name: dcgm.gpu.profiling.dram_utilization
     - action: update
-      include: gpu.dcgm.nvlink.traffic
+      include: gpu.dcgm.nvlink.io
       new_name: dcgm.gpu.profiling.nvlink_traffic_rate
+      operations:
+      - action: update_label
+        label: network.io.direction
+        new_label: direction
+      - action: update_label
+        label: direction
+        value_actions:
+        - new_value: rx
+          value: receive
+        - new_value: tx
+          value: transmit
     - action: update
-      include: gpu.dcgm.pcie.traffic
+      include: gpu.dcgm.pcie.io
       new_name: dcgm.gpu.profiling.pcie_traffic_rate
+      operations:
+      - action: update_label
+        label: network.io.direction
+        new_label: direction
+      - action: update_label
+        label: direction
+        value_actions:
+        - new_value: rx
+          value: receive
+        - new_value: tx
+          value: transmit
     - action: update
       include: gpu.dcgm.pipe.utilization
       new_name: dcgm.gpu.profiling.pipe_utilization
+      operations:
+      - action: update_label
+        label: gpu.pipe
+        new_label: pipe
     - action: update
       include: gpu.dcgm.sm.occupancy
       new_name: dcgm.gpu.profiling.sm_occupancy

diff --git a/confgenerator/testdata/goldens/metrics-receiver_dcgm/golden/linux/otel.yaml b/confgenerator/testdata/goldens/metrics-receiver_dcgm/golden/linux/otel.yaml
@@ -62,14 +62,40 @@ processors:
       include: gpu.dcgm.memory.bandwidth_utilization
       new_name: dcgm.gpu.profiling.dram_utilization
     - action: update
-      include: gpu.dcgm.nvlink.traffic
+      include: gpu.dcgm.nvlink.io
       new_name: dcgm.gpu.profiling.nvlink_traffic_rate
+      operations:
+      - action: update_label
+        label: network.io.direction
+        new_label: direction
+      - action: update_label
+        label: direction
+        value_actions:
+        - new_value: rx
+          value: receive
+        - new_value: tx
+          value: transmit
     - action: update
-      include: gpu.dcgm.pcie.traffic
+      include: gpu.dcgm.pcie.io
       new_name: dcgm.gpu.profiling.pcie_traffic_rate
+      operations:
+      - action: update_label
+        label: network.io.direction
+        new_label: direction
+      - action: update_label
+        label: direction
+        value_actions:
+        - new_value: rx
+          value: receive
+        - new_value: tx
+          value: transmit
     - action: update
       include: gpu.dcgm.pipe.utilization
       new_name: dcgm.gpu.profiling.pipe_utilization
+      operations:
+      - action: update_label
+        label: gpu.pipe
+        new_label: pipe
     - action: update
       include: gpu.dcgm.sm.occupancy
       new_name: dcgm.gpu.profiling.sm_occupancy

diff --git a/confgenerator/testdata/goldens/metrics-receiver_dcgm_v2/golden/linux-gpu/otel.yaml b/confgenerator/testdata/goldens/metrics-receiver_dcgm_v2/golden/linux-gpu/otel.yaml
@@ -58,6 +58,50 @@ processors:
         - otelcol_grpc_io_client_completed_rpcs
         - otelcol_googlecloudmonitoring_point_count
   metricstransform/dcgm_0:
+    transforms:
+    - action: update
+      include: gpu.dcgm.pipe.utilization
+      operations:
+      - action: update_label
+        label: gpu.pipe
+        new_label: pipe
+    - action: update
+      include: gpu.dcgm.memory.bytes_used
+      operations:
+      - action: update_label
+        label: gpu.memory.state
+        new_label: state
+    - action: update
+      include: gpu.dcgm.nvlink.io
+      operations:
+      - action: update_label
+        label: network.io.direction
+        new_label: direction
+    - action: update
+      include: gpu.dcgm.pcie.io
+      operations:
+      - action: update_label
+        label: network.io.direction
+        new_label: direction
+    - action: update
+      include: gpu.dcgm.clock.throttle_duration.time
+      operations:
+      - action: update_label
+        label: gpu.clock.violation
+        new_label: violation
+    - action: update
+      include: gpu.dcgm.ecc_errors
+      operations:
+      - action: update_label
+        label: gpu.error.type
+        new_label: error_type
+    - action: update
+      include: gpu.dcgm.xid_errors
+      operations:
+      - action: update_label
+        label: gpu.error.xid
+        new_label: xid
+  metricstransform/dcgm_1:
     transforms:
     - action: update
       include: ^(.*)$$
@@ -461,13 +505,13 @@ processors:
       include: ^(.*)$$
       match_type: regexp
       new_name: agent.googleapis.com/$${1}
-  modifyscope/dcgm_2:
+  modifyscope/dcgm_3:
     override_scope_name: agent.googleapis.com/dcgm
     override_scope_version: "2.0"
   resourcedetection/_global_0:
     detectors:
     - gcp
-  transform/dcgm_1:
+  transform/dcgm_2:
     metric_statements:
       context: datapoint
       statements:
@@ -517,8 +561,9 @@ service:
       - googlecloud/otel
       processors:
       - metricstransform/dcgm_0
-      - transform/dcgm_1
-      - modifyscope/dcgm_2
+      - metricstransform/dcgm_1
+      - transform/dcgm_2
+      - modifyscope/dcgm_3
       - resourcedetection/_global_0
       receivers:
       - dcgm/dcgm

diff --git a/confgenerator/testdata/goldens/metrics-receiver_dcgm_v2/golden/linux/otel.yaml b/confgenerator/testdata/goldens/metrics-receiver_dcgm_v2/golden/linux/otel.yaml
@@ -53,6 +53,50 @@ processors:
         - otelcol_grpc_io_client_completed_rpcs
         - otelcol_googlecloudmonitoring_point_count
   metricstransform/dcgm_0:
+    transforms:
+    - action: update
+      include: gpu.dcgm.pipe.utilization
+      operations:
+      - action: update_label
+        label: gpu.pipe
+        new_label: pipe
+    - action: update
+      include: gpu.dcgm.memory.bytes_used
+      operations:
+      - action: update_label
+        label: gpu.memory.state
+        new_label: state
+    - action: update
+      include: gpu.dcgm.nvlink.io
+      operations:
+      - action: update_label
+        label: network.io.direction
+        new_label: direction
+    - action: update
+      include: gpu.dcgm.pcie.io
+      operations:
+      - action: update_label
+        label: network.io.direction
+        new_label: direction
+    - action: update
+      include: gpu.dcgm.clock.throttle_duration.time
+      operations:
+      - action: update_label
+        label: gpu.clock.violation
+        new_label: violation
+    - action: update
+      include: gpu.dcgm.ecc_errors
+      operations:
+      - action: update_label
+        label: gpu.error.type
+        new_label: error_type
+    - action: update
+      include: gpu.dcgm.xid_errors
+      operations:
+      - action: update_label
+        label: gpu.error.xid
+        new_label: xid
+  metricstransform/dcgm_1:
     transforms:
     - action: update
       include: ^(.*)$$
@@ -432,13 +476,13 @@ processors:
       include: ^(.*)$$
       match_type: regexp
       new_name: agent.googleapis.com/$${1}
-  modifyscope/dcgm_2:
+  modifyscope/dcgm_3:
     override_scope_name: agent.googleapis.com/dcgm
     override_scope_version: "2.0"
   resourcedetection/_global_0:
     detectors:
     - gcp
-  transform/dcgm_1:
+  transform/dcgm_2:
     metric_statements:
       context: datapoint
       statements:
@@ -486,8 +530,9 @@ service:
       - googlecloud/otel
       processors:
       - metricstransform/dcgm_0
-      - transform/dcgm_1
-      - modifyscope/dcgm_2
+      - metricstransform/dcgm_1
+      - transform/dcgm_2
+      - modifyscope/dcgm_3
       - resourcedetection/_global_0
       receivers:
       - dcgm/dcgm