PR comments

pkg/collector/corechecks/gpu/gpu.go

@@ -161,6 +161,8 @@ func (c *Check) Run() error {
 	// Commit the metrics even in case of an error
 	defer snd.Commit()
 
+	// build the mapping of GPU devices -> containers to allow tagging device
+	// metrics with the tags of containers that are using them
 	gpuToContainersMap := c.getGPUToContainersMap()
 
 	if err := c.emitSysprobeMetrics(snd, gpuToContainersMap); err != nil {

pkg/collector/corechecks/gpu/model/model.go

@@ -20,8 +20,11 @@ type MemoryMetrics struct {
 
 // UtilizationMetrics contains the GPU stats for a given device and process
 type UtilizationMetrics struct {
-	UsedCores float64       `json:"used_cores"`
-	Memory    MemoryMetrics `json:"memory"`
+	// UsedCores stores the average number of GPU cores used by this process in the interval
+	UsedCores float64 `json:"used_cores"`
+
+	// Memory stores the memory stats for the process during the interval
+	Memory MemoryMetrics `json:"memory"`
 }
 
 // StatsKey is the key used to identify a GPUStats object

pkg/gpu/aggregator.go

@@ -67,7 +67,7 @@ func (agg *aggregator) processKernelSpan(span *kernelSpan) {
 	// threads that were enqueued.
 	//
 	// An example of a situation where this distinction is important: say we
-	// have a kernel launch with 100 threads, but the GPU can only run 500
+	// have a kernel launch with 100 threads, but the GPU can only run 50
 	// threads, and assume this kernel runs for 1 second and that we want to
 	// report utilization for the last 2 seconds. If we were looking at the
 	// actual GPU utilization in real-time, we'd see 100% utilization for the