feat(exporter): add process info

XuehaiPan · Dec 28, 2024 · ad8a454 · ad8a454
1 parent 25b6d61
commit ad8a454
Showing 1 changed file with 35 additions and 7 deletions.
diff --git a/nvitop-exporter/nvitop_exporter/exporter.py b/nvitop-exporter/nvitop_exporter/exporter.py
@@ -405,6 +405,12 @@ def __init__(  # pylint: disable=too-many-statements
         )
 
         # Create gauges for process metrics
+        self.process_info = Info(
+            name='process_info',
+            documentation='Process information.',
+            labelnames=['hostname', 'index', 'devicename', 'uuid', 'pid', 'username'],
+            registry=self.registry,
+        )
         self.process_running_time = Gauge(
             name='process_running_time',
             documentation='Process running time (s).',
@@ -592,19 +598,40 @@ def update_device(self, device: Device) -> None:  # pylint: disable=too-many-loc
         alive_pids.clear()
 
         with GpuProcess.failsafe():
+            host_snapshots = {}
             for pid, process in device.processes().items():
                 with process.oneshot():
                     username = process.username()
-                    running_time = process.running_time()
                     alive_pids.add((pid, username))
+                    if (pid, username) not in host_snapshots:  # noqa: SIM401,RUF100
+                        host_snapshot = host_snapshots[(pid, username)] = process.host_snapshot()
+                    else:
+                        host_snapshot = host_snapshots[(pid, username)]
+                    self.process_info.labels(
+                        hostname=self.hostname,
+                        index=index,
+                        devicename=name,
+                        uuid=uuid,
+                        pid=pid,
+                        username=username,
+                    ).info(
+                        {
+                            'status': host_snapshot.status,
+                            'command': host_snapshot.command,
+                        },
+                    )
                     for gauge, value in (
                         (
                             self.process_running_time,
-                            running_time.total_seconds() if running_time else math.nan,
+                            (
+                                host_snapshot.running_time.total_seconds()
+                                if host_snapshot.running_time
+                                else math.nan
+                            ),
                         ),
-                        (self.process_cpu_percent, process.cpu_percent()),
-                        (self.process_rss_memory, process.host_memory() / MiB),
-                        (self.process_memory_percent, float(process.memory_percent())),
+                        (self.process_cpu_percent, host_snapshot.cpu_percent),
+                        (self.process_rss_memory, host_snapshot.host_memory / MiB),
+                        (self.process_memory_percent, float(host_snapshot.memory_percent)),
                         (self.process_gpu_memory, process.gpu_memory() / MiB),
                         (
                             self.process_gpu_sm_utilization,
@@ -633,7 +660,8 @@ def update_device(self, device: Device) -> None:  # pylint: disable=too-many-loc
                         ).set(value)
 
         for pid, username in previous_alive_pids.difference(alive_pids):
-            for gauge in (
+            for collector in (
+                self.process_info,
                 self.process_running_time,
                 self.process_cpu_percent,
                 self.process_rss_memory,
@@ -645,7 +673,7 @@ def update_device(self, device: Device) -> None:  # pylint: disable=too-many-loc
                 self.process_gpu_decoder_utilization,
             ):
                 try:
-                    gauge.remove(
+                    collector.remove(
                         self.hostname,
                         index,
                         name,