[gpum] Report gpu.sm_active from GPM GR engine activity

[martavicentenavarro]

What does this PR do?

Adds a source for gpu.sm_active from GPM_METRIC_GRAPHICS_UTIL (already collected as gpu.gr_engine_active) on physical GPUs that don't have MIG mode enabled.

• Low priority by default, so it's only used as a fallback when no other gpu.sm_active source is available (the NVML sampling source is Medium, and eBPF wins the tie at Low).
• High priority with the new gpu.prefer_gr_engine_sm_active option.
• Not used when gpu.legacy_sm_active is enabled (the legacy value takes precedence), nor on MIG devices or MIG-enabled GPUs, where it hasn't been validated.
• RemoveDuplicateSamples now resolves equal-priority ties by collector name instead of map iteration order, so sm_active can't alternate between the ebpf and gpm sources at Low. No other metric has an equal-priority tie across collectors, so their output doesn't change.

No new NVML calls: GRAPHICS_UTIL is already queried for gpu.gr_engine_active.

Motivation

gpu.sm_active should be the percentage of time at least one SM was active. GRAPHICS_UTIL is a time-based measure of the GR engine being busy, unlike GPM_METRIC_SM_UTIL, which averages SM activity over all SMs (25% when a quarter of the SMs are busy all the time).

This is an alternative to #57928, which derives sm_active from the SM cycle counters and turned out to report the same value as SM_UTIL. Both PRs carry the same RemoveDuplicateSamples change.

Describe how you validated your changes

Unit tests in pkg/collector/corechecks/gpu/nvidia.

On an H100 80GB HBM3 (driver 595.91.07), GRAPHICS_UTIL was compared with a ground truth computed from per-kernel GPU timestamps (%globaltimer), as the percentage of time with a kernel running:

Workload	Ground truth	GRAPHICS_UTIL	SM_UTIL	NVML GPU util
33 of 132 SMs busy continuously	100.0	100.0	25.0	99.4
All SMs, 50% duty cycle (10 ms)	49.5	49.6	49.6	49
1 SM, 50% duty cycle	49.5	49.7	0.4	49
1 SM, 10% duty cycle	9.9	9.9	0.1	9
20 µs kernels, synchronizing after each	71.4	80.3	71.7	80
2 µs kernels, synchronizing after each	20.5	45.3	21.1	45
memcpy only, no kernels	0.0	0.0	0.0	99

GRAPHICS_UTIL matches the ground truth within ±0.2 points for kernels of 1 ms or longer. It overestimates by about 2.5 µs per kernel launch, as NVML GPU utilization does. It doesn't count memory copies, which NVML GPU and process utilization (the current sampling source) report as ~99–100% busy.

With the Agent built from this branch (agent check gpu, three runs 15 s apart, GPU 0):

Workload	Default config	gpu.prefer_gr_engine_sm_active: true	With gpu.legacy_sm_active too
Idle	0	0	-
gpu-burner (--run_time 300 auto --target_sm 60)	61–62 (sampling)	61.6–62.0 (= gr_engine_active)	59.7 (= sm_utilization)
33 of 132 SMs busy	100 (sampling)	99.96–99.97 (= gr_engine_active)	-
memcpy only	100 (sampling)	0.01 (= gr_engine_active)	-
2 µs kernels	44 (sampling)	45 (= gr_engine_active)	-

Additional Notes

• With the default config, the new sm_active sample is produced on every run and dropped by deduplication when another source wins. This raises the duplicate_metrics telemetry by one per device per run.
• GRAPHICS_UTIL measures the GR engine, so it includes the time spent launching kernels: workloads made of thousands of very short kernels read higher than the actual SM activity (the current sampling source behaves the same way).
• Not validated on MIG (excluded) nor with system-probe/eBPF on hardware.

[github-actions]

AI review by Codex (OpenAI) - workflow run

patch is correct. No actionable regressions found in the source selection, MIG exclusions, configuration wiring, or deterministic deduplication. Tests were not run in the read-only environment.

[datadog-prod-us1-4]

🎯 Code Coverage (details)
• Patch Coverage: 100.00%
• Overall Coverage: 58.62% (+0.02%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: d5ebf55 | Docs | Give us feedback!}

[dd-octo-sts]

Files inventory check summary

File checks results against ancestor 8e99ce18:

Results for datadog-agent_7.86.0~devel.git.328.d5ebf55.pipeline.143810155-1_amd64.deb:

No change detected

Results for datadog-iot-agent_7.86.0~devel.git.328.d5ebf55.pipeline.143810155-1_amd64.deb:

No change detected

[dd-octo-sts]

Static quality checks

✅ Please find below the results from static quality gates
Comparison made with ancestor 8e99ce1
📊 Static Quality Gates Dashboard
🔗 SQG Job

Successful checks

Info

	Quality gate	Change	Size (prev → curr → max)
✅	agent_deb_amd64	+8.19 KiB (0.00% increase, -0.15% of buffer)	755.293 → 755.301 → 760.520
✅	agent_deb_amd64_fips	+8.52 KiB (0.00% increase, -0.16% of buffer)	688.271 → 688.280 → 693.570
✅	agent_msi	+2.05 KiB (0.00% increase, -1.01% of buffer)	653.991 → 653.993 → 654.190
✅	agent_rpm_amd64	+8.19 KiB (0.00% increase, -0.15% of buffer)	755.276 → 755.284 → 760.500
✅	agent_rpm_amd64_fips	+8.52 KiB (0.00% increase, -0.16% of buffer)	688.255 → 688.263 → 693.550
✅	agent_rpm_arm64	+5.23 KiB (0.00% increase, -0.10% of buffer)	725.822 → 725.827 → 731.100
✅	agent_rpm_arm64_fips	+8.5 KiB (0.00% increase, -0.15% of buffer)	664.366 → 664.374 → 669.740
✅	agent_suse_amd64	+8.19 KiB (0.00% increase, -0.15% of buffer)	755.276 → 755.284 → 760.500
✅	agent_suse_amd64_fips	+8.52 KiB (0.00% increase, -0.16% of buffer)	688.255 → 688.263 → 693.550
✅	agent_suse_arm64	+5.23 KiB (0.00% increase, -0.10% of buffer)	725.822 → 725.827 → 731.100
✅	agent_suse_arm64_fips	+8.5 KiB (0.00% increase, -0.15% of buffer)	664.366 → 664.374 → 669.740
✅	docker_agent_amd64	+8.19 KiB (0.00% increase, -0.15% of buffer)	811.455 → 811.463 → 816.720
✅	docker_agent_arm64	+5.22 KiB (0.00% increase, -0.10% of buffer)	806.628 → 806.633 → 811.970
✅	docker_agent_jmx_amd64	+8.19 KiB (0.00% increase, -0.15% of buffer)	1002.325 → 1002.333 → 1007.590
✅	docker_agent_jmx_arm64	+5.22 KiB (0.00% increase, -0.10% of buffer)	986.278 → 986.283 → 991.620

18 successful checks with minimal change (< 2 KiB)

	Quality gate	Current Size
✅	agent_heroku_amd64	319.102 MiB
✅	docker_cws_instrumentation_amd64	7.443 MiB
✅	docker_cws_instrumentation_arm64	6.877 MiB
✅	docker_dogstatsd_amd64	39.566 MiB
✅	docker_dogstatsd_arm64	37.628 MiB
✅	docker_host_profiler_amd64	308.725 MiB
✅	docker_host_profiler_arm64	319.865 MiB
✅	dogstatsd_deb_amd64	30.309 MiB
✅	dogstatsd_deb_arm64	28.294 MiB
✅	dogstatsd_rpm_amd64	30.309 MiB
✅	dogstatsd_suse_amd64	30.309 MiB
✅	iot_agent_deb_amd64	47.394 MiB
✅	iot_agent_deb_arm64	43.857 MiB
✅	iot_agent_deb_armhf	44.668 MiB
✅	iot_agent_rpm_amd64	47.395 MiB
✅	iot_agent_suse_amd64	47.394 MiB
✅	docker_cluster_agent_amd64	211.240 MiB
✅	docker_cluster_agent_arm64	224.225 MiB

[cit-pr-commenter-54b7da]

Regression Detector

Regression Detector Results

Metrics dashboard
Target profiles
Job ID: 4708c803-c7f9-40a3-8740-b2d6cea16024

Baseline: 8e99ce1
Comparison: d5ebf55
Diff

Optimization Goals: ✅ No significant changes detected

Fine details of change detection per experiment

perf	experiment	goal	Δ mean %	Δ mean % CI	trials	links
➖	dsd_uds_client_drop_detector_cpu	% cpu utilization	+0.21	[-0.25, +0.67]	1	Logs
➖	quality_gate_idle_all_features	memory utilization	+0.03	[-0.05, +0.11]	1	Logs bounds checks dashboard
➖	quality_gate_security_no_fs_load	memory utilization	-0.09	[-0.16, -0.01]	1	Logs bounds checks dashboard
➖	quality_gate_idle	memory utilization	-0.15	[-0.19, -0.10]	1	Logs bounds checks dashboard
➖	python_openmetrics	% cpu utilization	-0.18	[-0.86, +0.49]	1	Logs bounds checks dashboard
➖	quality_gate_security_idle	memory utilization	-0.35	[-0.38, -0.31]	1	Logs bounds checks dashboard
➖	quality_gate_private_action_runner	memory utilization	-0.35	[-0.47, -0.22]	1	Logs bounds checks dashboard
➖	dsd_uds_10mb_3k_timestamped_contexts_memory	memory utilization	-0.51	[-0.73, -0.30]	1	Logs
➖	quality_gate_security_mean_fs_load	memory utilization	-0.55	[-0.59, -0.50]	1	Logs bounds checks dashboard
➖	quality_gate_logs	% cpu utilization	-0.64	[-1.50, +0.23]	1	Logs bounds checks dashboard
➖	dsd_uds_10mb_3k_timestamped_contexts_cpu	% cpu utilization	-0.91	[-1.15, -0.67]	1	Logs
➖	quality_gate_metrics_logs	memory utilization	-0.99	[-1.23, -0.75]	1	Logs bounds checks dashboard

Bounds Checks: ✅ Passed

perf	experiment	bounds_check_name	replicates_passed	observed_value	links
✅	python_openmetrics	checks_execution_time	10/10	80.53 ≤ 100	bounds checks dashboard
✅	python_openmetrics	cpu_usage	10/10	1333.32 ≤ 1500	bounds checks dashboard
✅	python_openmetrics	memory_usage	10/10	4.31GiB ≤ 4.75GiB	bounds checks dashboard
✅	quality_gate_idle	intake_connections	10/10	4 ≤ 5	bounds checks dashboard
✅	quality_gate_idle	memory_usage	10/10	178.18MiB ≤ 181MiB	bounds checks dashboard
✅	quality_gate_idle	total_bytes_received	10/10	758.77KiB ≤ 819.20KiB	bounds checks dashboard
✅	quality_gate_idle_all_features	intake_connections	10/10	2 ≤ 5	bounds checks dashboard
✅	quality_gate_idle_all_features	memory_usage	10/10	475.58MiB ≤ 542MiB	bounds checks dashboard
✅	quality_gate_idle_all_features	total_bytes_received	10/10	1.14MiB ≤ 1.25MiB	bounds checks dashboard
✅	quality_gate_logs	intake_connections	10/10	17 ≤ 40	bounds checks dashboard
✅	quality_gate_logs	memory_usage	10/10	215.18MiB ≤ 228MiB	bounds checks dashboard
✅	quality_gate_logs	missed_bytes	10/10	0B = 0B	bounds checks dashboard
✅	quality_gate_logs	total_bytes_received	10/10	263.89MiB ≤ 292MiB	bounds checks dashboard
✅	quality_gate_metrics_logs	cpu_usage	10/10	389.75 ≤ 2000	bounds checks dashboard
✅	quality_gate_metrics_logs	intake_connections	10/10	19 ≤ 40	bounds checks dashboard
✅	quality_gate_metrics_logs	memory_usage	10/10	447.47MiB ≤ 455MiB	bounds checks dashboard
✅	quality_gate_metrics_logs	missed_bytes	10/10	0B = 0B	bounds checks dashboard
✅	quality_gate_metrics_logs	total_bytes_received	10/10	0.95GiB ≤ 1.04GiB	bounds checks dashboard
✅	quality_gate_private_action_runner	memory_usage	10/10	74.40MiB ≤ 77MiB	bounds checks dashboard
✅	quality_gate_security_idle	cpu_usage	10/10	30.73 ≤ 100	bounds checks dashboard
✅	quality_gate_security_idle	memory_usage	10/10	320.27MiB ≤ 357MiB	bounds checks dashboard
✅	quality_gate_security_mean_fs_load	cpu_usage	10/10	70.09 ≤ 200	bounds checks dashboard
✅	quality_gate_security_mean_fs_load	memory_usage	10/10	310.23MiB ≤ 337MiB	bounds checks dashboard
✅	quality_gate_security_no_fs_load	cpu_usage	10/10	23.69 ≤ 100	bounds checks dashboard
✅	quality_gate_security_no_fs_load	memory_usage	10/10	330.47MiB ≤ 348MiB	bounds checks dashboard

Explanation

Confidence level: 90.00%
Effect size tolerance: |Δ mean %| ≥ 5.00%

Performance changes are noted in the perf column of each table:

• ✅ = significantly better comparison variant performance
• ❌ = significantly worse comparison variant performance
• ➖ = no significant change in performance

A regression test is an A/B test of target performance in a repeatable rig, where "performance" is measured as "comparison variant minus baseline variant" for an optimization goal (e.g., ingress throughput). Due to intrinsic variability in measuring that goal, we can only estimate its mean value for each experiment; we report uncertainty in that value as a 90.00% confidence interval denoted "Δ mean % CI".

For each experiment, we decide whether a change in performance is a "regression" -- a change worth investigating further -- if all of the following criteria are true:

1. Its estimated |Δ mean %| ≥ 5.00%, indicating the change is big enough to merit a closer look.

2. Its 90.00% confidence interval "Δ mean % CI" does not contain zero, indicating that if our statistical model is accurate, there is at least a 90.00% chance there is a difference in performance between baseline and comparison variants.

3. Its configuration does not mark it "erratic".

CI Pass/Fail Decision

✅ Passed. All Quality Gates passed.

• quality_gate_logs, bounds check missed_bytes: 10/10 replicas passed. Gate passed.
• quality_gate_logs, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_logs, bounds check intake_connections: 10/10 replicas passed. Gate passed.
• quality_gate_logs, bounds check total_bytes_received: 10/10 replicas passed. Gate passed.
• quality_gate_security_mean_fs_load, bounds check cpu_usage: 10/10 replicas passed. Gate passed.
• quality_gate_security_mean_fs_load, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_idle_all_features, bounds check intake_connections: 10/10 replicas passed. Gate passed.
• quality_gate_idle_all_features, bounds check total_bytes_received: 10/10 replicas passed. Gate passed.
• quality_gate_idle_all_features, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_metrics_logs, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_metrics_logs, bounds check total_bytes_received: 10/10 replicas passed. Gate passed.
• quality_gate_metrics_logs, bounds check intake_connections: 10/10 replicas passed. Gate passed.
• quality_gate_metrics_logs, bounds check cpu_usage: 10/10 replicas passed. Gate passed.
• quality_gate_metrics_logs, bounds check missed_bytes: 10/10 replicas passed. Gate passed.
• quality_gate_idle, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_idle, bounds check total_bytes_received: 10/10 replicas passed. Gate passed.
• quality_gate_idle, bounds check intake_connections: 10/10 replicas passed. Gate passed.
• quality_gate_private_action_runner, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_security_no_fs_load, bounds check cpu_usage: 10/10 replicas passed. Gate passed.
• quality_gate_security_no_fs_load, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_security_idle, bounds check memory_usage: 10/10 replicas passed. Gate passed.
• quality_gate_security_idle, bounds check cpu_usage: 10/10 replicas passed. Gate passed.