add GPU metrics

Dockerfile

@@ -6,7 +6,7 @@ COPY go.sum .
 RUN go mod download
 COPY . .
 ARG VERSION=unknown
-RUN CGO_ENABLED=1 go build -mod=readonly -ldflags "-X 'github.com/coroot/coroot-node-agent/flags.Version=${VERSION}'" -o coroot-node-agent .
+RUN CGO_ENABLED=1 go build -mod=readonly -ldflags "-extldflags='-Wl,-z,lazy' -X 'github.com/coroot/coroot-node-agent/flags.Version=${VERSION}'" -o coroot-node-agent .
 
 FROM registry.access.redhat.com/ubi9/ubi
 

containers/app.go

@@ -120,6 +120,8 @@ func guessApplicationTypeByCmdline(cmdline []byte) string {
 		return "nats"
 	case bytes.HasSuffix(cmd, []byte("java")):
 		return "java"
+	case bytes.HasSuffix(cmd, []byte("ollama")):
+		return "ollama"
 	case bytes.Contains(cmd, []byte("victoria-metrics")) ||
 		bytes.Contains(cmd, []byte("vmstorage")) ||
 		bytes.Contains(cmd, []byte("vminsert")) ||

containers/container.go

@@ -29,6 +29,7 @@ var (
 	gcInterval                = 10 * time.Minute
 	pingTimeout               = 300 * time.Millisecond
 	multilineCollectorTimeout = time.Second
+	gpuStatsWindow            = 15 * time.Second
 )
 
 type ContainerID string
@@ -132,6 +133,8 @@ type Container struct {
 	l7Stats  L7Stats
 	dnsStats *L7Metrics
 
+	gpuStats map[string]*GpuUsage
+
 	oomKills                 int
 	pythonThreadLockWaitTime time.Duration
 
@@ -181,6 +184,8 @@ func NewContainer(id ContainerID, cg *cgroup.Cgroup, md *ContainerMetadata, pid
 		l7Stats:                  L7Stats{},
 		dnsStats:                 &L7Metrics{},
 
+		gpuStats: map[string]*GpuUsage{},
+
 		mounts:     map[string]proc.MountInfo{},
 		seenMounts: map[uint64]struct{}{},
 
@@ -370,7 +375,27 @@ func (c *Container) Collect(ch chan<- prometheus.Metric) {
 				process.dotNetMonitor.Collect(ch)
 			}
 		}
+
+		for _, usage := range c.gpuStats {
+			usage.Reset()
+		}
+		if usage := process.getGPUUsage(); usage != nil {
+			for uuid, u := range usage {
+				tu := c.gpuStats[uuid]
+				if tu == nil {
+					tu = &GpuUsage{}
+					c.gpuStats[uuid] = tu
+				}
+				tu.GPU += u.GPU
+				tu.Memory += u.Memory
+			}
+		}
 	}
+	for uuid, usage := range c.gpuStats {
+		ch <- gauge(metrics.GpuUsagePercent, usage.GPU, uuid)
+		ch <- gauge(metrics.GpuMemoryUsagePercent, usage.Memory, uuid)
+	}
+
 	for appType := range appTypes {
 		ch <- gauge(metrics.ApplicationType, 1, appType)
 	}

containers/metrics.go

@@ -51,6 +51,9 @@ var metrics = struct {
 
 	PythonThreadLockWaitTime *prometheus.Desc
 
+	GpuUsagePercent       *prometheus.Desc
+	GpuMemoryUsagePercent *prometheus.Desc
+
 	Ip2Fqdn *prometheus.Desc
 }{
 	ContainerInfo: metric("container_info", "Meta information about the container", "image", "systemd_triggered_by"),
@@ -100,6 +103,9 @@ var metrics = struct {
 	Ip2Fqdn: metric("ip_to_fqdn", "Mapping IP addresses to FQDNs based on DNS requests initiated by containers", "ip", "fqdn"),
 
 	PythonThreadLockWaitTime: metric("container_python_thread_lock_wait_time_seconds", "Time spent waiting acquiring GIL in seconds"),
+
+	GpuUsagePercent:       metric("container_resources_gpu_usage_percent", "Percent of GPU compute resources used by the container", "gpu_uuid"),
+	GpuMemoryUsagePercent: metric("container_resources_gpu_memory_usage_percent", "Percent of GPU memory used by the container", "gpu_uuid"),
 }
 
 var (

containers/process.go

@@ -8,11 +8,22 @@ import (
 
 	"github.com/cilium/ebpf/link"
 	"github.com/coroot/coroot-node-agent/ebpftracer"
+	"github.com/coroot/coroot-node-agent/gpu"
 	"github.com/coroot/coroot-node-agent/proc"
 	"github.com/jpillora/backoff"
 	"github.com/mdlayher/taskstats"
 )
 
+type GpuUsage struct {
+	GPU    float64
+	Memory float64
+}
+
+func (gu *GpuUsage) Reset() {
+	gu.Memory = 0
+	gu.GPU = 0
+}
+
 type Process struct {
 	Pid       uint32
 	StartedAt time.Time
@@ -29,6 +40,8 @@ type Process struct {
 	goTlsUprobesChecked   bool
 	openSslUprobesChecked bool
 	pythonGilChecked      bool
+
+	gpuUsageSamples []gpu.ProcessUsageSample
 }
 
 func NewProcess(pid uint32, stats *taskstats.Stats, tracer *ebpftracer.Tracer) *Process {
@@ -97,6 +110,43 @@ func (p *Process) instrumentPython(cmdline []byte, tracer *ebpftracer.Tracer) {
 	p.uprobes = append(p.uprobes, tracer.AttachPythonThreadLockProbes(p.Pid)...)
 }
 
+func (p *Process) addGpuUsageSample(sample gpu.ProcessUsageSample) {
+	p.removeOldGpuUsageSamples(sample.Timestamp.Add(-gpuStatsWindow))
+	p.gpuUsageSamples = append(p.gpuUsageSamples, sample)
+}
+
+func (p *Process) getGPUUsage() map[string]*GpuUsage {
+	p.removeOldGpuUsageSamples(time.Now().Add(-gpuStatsWindow))
+	if len(p.gpuUsageSamples) == 0 {
+		return nil
+	}
+	gpuStatsWindowSeconds := gpuStatsWindow.Seconds()
+	res := make(map[string]*GpuUsage)
+	for _, sample := range p.gpuUsageSamples {
+		u := res[sample.UUID]
+		if u == nil {
+			u = &GpuUsage{}
+			res[sample.UUID] = u
+		}
+		u.GPU += float64(sample.GPUPercent) / gpuStatsWindowSeconds
+		u.Memory += float64(sample.MemoryPercent) / gpuStatsWindowSeconds
+	}
+	return res
+}
+
+func (p *Process) removeOldGpuUsageSamples(cutoff time.Time) {
+	i := 0
+	for ; i < len(p.gpuUsageSamples); i++ {
+		if p.gpuUsageSamples[i].Timestamp.After(cutoff) {
+			break
+		}
+	}
+	if i > 0 {
+		copy(p.gpuUsageSamples, p.gpuUsageSamples[i:])
+		p.gpuUsageSamples = p.gpuUsageSamples[:len(p.gpuUsageSamples)-i]
+	}
+}
+
 func (p *Process) Close() {
 	p.cancelFunc()
 	for _, u := range p.uprobes {

containers/registry.go

@@ -14,6 +14,7 @@ import (
 	"github.com/coroot/coroot-node-agent/common"
 	"github.com/coroot/coroot-node-agent/ebpftracer"
 	"github.com/coroot/coroot-node-agent/flags"
+	"github.com/coroot/coroot-node-agent/gpu"
 	"github.com/coroot/coroot-node-agent/proc"
 	"github.com/prometheus/client_golang/prometheus"
 	"github.com/vishvananda/netns"
@@ -59,9 +60,11 @@ type Registry struct {
 	trafficStatsLastUpdated time.Time
 	trafficStatsLock        sync.Mutex
 	trafficStatsUpdateCh    chan *TrafficStatsUpdate
+
+	gpuProcessUsageSampleChan chan gpu.ProcessUsageSample
 }
 
-func NewRegistry(reg prometheus.Registerer, processInfoCh chan<- ProcessInfo) (*Registry, error) {
+func NewRegistry(reg prometheus.Registerer, processInfoCh chan<- ProcessInfo, gpuProcessUsageSampleChan chan gpu.ProcessUsageSample) (*Registry, error) {
 	ns, err := proc.GetSelfNetNs()
 	if err != nil {
 		return nil, err
@@ -113,6 +116,8 @@ func NewRegistry(reg prometheus.Registerer, processInfoCh chan<- ProcessInfo) (*
 		tracer: ebpftracer.NewTracer(hostNetNs, selfNetNs, *flags.DisableL7Tracing),
 
 		trafficStatsUpdateCh: make(chan *TrafficStatsUpdate),
+
+		gpuProcessUsageSampleChan: gpuProcessUsageSampleChan,
 	}
 	if err = reg.Register(r); err != nil {
 		return nil, err
@@ -205,6 +210,12 @@ func (r *Registry) handleEvents(ch <-chan ebpftracer.Event) {
 			if c := r.containersByPid[u.Pid]; c != nil {
 				c.updateTrafficStats(u)
 			}
+		case sample := <-r.gpuProcessUsageSampleChan:
+			if c := r.containersByPid[sample.Pid]; c != nil {
+				if p := c.processes[sample.Pid]; p != nil {
+					p.addGpuUsageSample(sample)
+				}
+			}
 		case e, more := <-ch:
 			if !more {
 				return

go.mod

@@ -5,6 +5,7 @@ go 1.23.8
 require (
 	cloud.google.com/go/compute/metadata v0.5.2
 	github.com/ClickHouse/ch-go v0.62.0
+	github.com/NVIDIA/go-nvml v0.12.4-1
 	github.com/agoda-com/opentelemetry-logs-go v0.4.1
 	github.com/cilium/cilium v1.17.2
 	github.com/cilium/ebpf v0.17.3

go.sum

@@ -70,6 +70,8 @@ github.com/Microsoft/hcsshim v0.9.12 h1:0Wgl1fRF4WmBuqP6EnHk2w3m7CCCumD/KUumZxp7
 github.com/Microsoft/hcsshim v0.9.12/go.mod h1:qAiPvMgZoM0wpkVg6qMdSEu+1VtI6/qHOOPkTGt8ftQ=
 github.com/Microsoft/hcsshim/test v0.0.0-20201218223536-d3e5debf77da/go.mod h1:5hlzMzRKMLyo42nCZ9oml8AdTlq/0cvIaBv6tK1RehU=
 github.com/Microsoft/hcsshim/test v0.0.0-20210227013316-43a75bb4edd3/go.mod h1:mw7qgWloBUl75W/gVH3cQszUg1+gUITj7D6NY7ywVnY=
+github.com/NVIDIA/go-nvml v0.12.4-1 h1:WKUvqshhWSNTfm47ETRhv0A0zJyr1ncCuHiXwoTrBEc=
+github.com/NVIDIA/go-nvml v0.12.4-1/go.mod h1:8Llmj+1Rr+9VGGwZuRer5N/aCjxGuR5nPb/9ebBiIEQ=
 github.com/NYTimes/gziphandler v0.0.0-20170623195520-56545f4a5d46/go.mod h1:3wb06e3pkSAbeQ52E9H9iFoQsEEwGN64994WTCIhntQ=
 github.com/OneOfOne/xxhash v1.2.2/go.mod h1:HSdplMjZKSmBqAxg5vPj2TmRDmfkzw+cTzAElWljhcU=
 github.com/PuerkitoBio/purell v1.0.0/go.mod h1:c11w/QuzBsJSee3cPx9rAFu61PvFxuPbtSwDGJws/X0=