TL;DR A GPU shows 97% utilization in nvidia-smi , but training throughput is a fraction of what benchmarks promise. The GPU is not computing; it is waiting. Data loading workers are starving the training loop because CPU contention, I/O bottlenecks, or scheduling delays prevent data from arriving fast enough. Tracing the full host-to-GPU pipeline via eBPF uprobes reveals exactly where the bubble is. We investigated a case where GPU utilization numbers looked healthy but training was slow, revealing a gap between metric dashboards and actual compute efficiency. The Problem An H100 costs 2.5 Million Problem This isn't hypothetical. A 100-GPU H100 cluster at 60% effective utilization (despite nvidia-smi reporting 95%+) wastes 2.5M. 75% of organizations report GPU utilization below 70% at peak. The gap between what nvidia-smi reports and actual compute efficiency is where millions of dollars disappear. What nvidia-smi Cannot Show nvidia-smi samples GPU state once per second. It reports a binary: "was a kernel running?" It has zero visibility into: Pipeline bubbles: GPU idle between kernel launches while waiting for data CPU scheduling delays: DataLoader workers got preempted by other processes Host memory pressure: Page faults in the data loading path stalling cudaMemcpy I/O bottlenecks: Disk or NFS reads blocking the next batch Scheduling storms: 10,000+ context switches per second on the training process These are all host-side problems causing GPU-side underutilization. nvidia-smi only sees the GPU side. What the Trace Shows The tracer traces both sides: CUDA APIs (what the GPU is doing) and host kernel events (what the CPU is doing), then builds causal chains connecting them. Step 1: Spot the pipeline bubble $ ingero explain --since 120s
System Context: CPU: 94.2% | Memory: 78.1% | Load: 12.3 (8 cores) | Swap: 0 MB
Causal Chains (last 2 min): ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ [HIGH] CPU scheduling contention → CUDA throughput drop Root: 14,504 context switches on training process (PID 3821) Process off-CPU 62 of 120 seconds (51.7% of wall clock) Effect: cudaStreamSync p99 inflated 1,028x (7µs → 7.2ms) CUDA op throughput dropped 47% from peak (1,200 → 640 ops/sec) Contributing: 4 DataLoader workers + 3 background processes competing for 8 cores Fix: pin training to dedicated cores: taskset -c 0-3 python3 train.py set DataLoader persistent_workers=True nice -n 19 background jobs There it is. The training process was off-CPU for 51.7% of the time . The GPU was waiting, not computing. nvidia-smi saw kernels queued and reported "97% utilized," but actual compute throughput was half of what it should be. Step 2: Find who's stealing CPU time Using the MCP server: Engineer : "Which processes caused the most scheduling contention in the last 2 minutes?" SELECT pn . name as process , COUNT ( * ) as context_switches , SUM ( duration_ns ) / 1 e9 as total_off_cpu_sec , MAX ( duration_ns ) / 1 e6 as worst_stall_ms FROM events e JOIN process_names pn ON e . pid = pn . pid WHERE op = 'sched_switch' AND timestamp > ( SELECT MAX ( timestamp ) - 120000000000 FROM events ) GROUP BY pn . name ORDER BY total_off_cpu_sec DESC LIMIT 10 ; process | switches | off_cpu_sec | worst_stall_ms ---------------------|----------|-------------|---------------- python3 (train.py) | 14,504 | 62.0 | 790.3 pt_data_worker:0 | 8,217 | 31.4 | 609.1 pt_data_worker:1 | 7,932 | 29.8 | 642.7 pt_data_worker:2 | 8,104 | 30.1 | 611.3 pt_data_worker:3 | 7,889 | 28.9 | 587.6 prometheus-node-exp | 3,201 | 8.7 | 45.2 fluent-bit | 2,890 | 7.1 | 38.9 The training process and all 4 DataLoader workers are fighting for CPU. And the worst single stall is 790ms , that's almost a full second where the training loop was frozen while the GPU sat idle. Background monitoring agents (Prometheus node exporter, Fluent Bit) are stealing another 15+ seconds of CPU time. Step 3: See it over time SELECT ( timestamp / 10000000000 ) * 10 as window_sec , COUNT ( CASE WHEN op = 'sched_switch' THEN 1 END ) as ctx_switches , COUNT ( CASE WHEN op = 'cudaStreamSync' THEN 1 END ) as sync_calls , AVG ( CASE WHEN op = 'cudaStreamSync' THEN duration_ns END ) / 1000 as sync_avg_us FROM events WHERE timestamp > ( SELECT MAX ( timestamp ) - 120000000000 FROM events ) GROUP BY window_sec ORDER BY window_sec ; window_sec | ctx_switches | sync_calls | sync_avg_us -----------|-------------|------------|------------ 0 | 342 | 89 | 52 ← baseline 10 | 1,205 | 91 | 180 ← contention starts 20 | 2,847 | 78 | 890 ← throughput drops 30 | 3,102 | 64 | 1,420 ← GPU starving 40 | 2,956 | 61 | 2,100 ← worst period 50 | 1,834 | 72 | 780 ← partial recovery At the 40-second mark, context switches hit 3,000/10s and cudaStreamSync average latency is 40x baseline. The GPU is doing 30% fewer sync calls, not because it's working harder, but because it has nothing to sync on. The pipeline is empty. Step 4: Get the Python stack trace With --stack enabled, The tracer captures exactly which Python function was on-CPU when the stall happened: Top cudaStreamSync callers during contention window (t=20-50s): train.py:142 → cudaStreamSync | 89 calls | avg 1.8ms | max 7.2ms ↳ loss.backward() train.py:145 → cudaStreamSync | 34 calls | avg 2.1ms | max 4.9ms ↳ optimizer.step()
Top sched_switch victims: train.py:138 → DataLoader.next() | preempted 4,201 times ↳ waiting for batch from workers The training loop at line 138 is blocked waiting for the next batch. The DataLoader workers themselves are being preempted. The fix is clear. The Fix Pin training to dedicated cores: taskset -c 0-3 python3 train.py: isolate from background processes Use persistent workers: DataLoader(persistent_workers=True): eliminates worker respawn overhead Reduce background noise: nice -n 19 for monitoring agents, or move them to separate cgroups Match worker count to available cores: Don't set num_workers=8 on an 8-core machine; leave 2 cores for the training loop and OS Check I/O: If workers are blocked on disk reads, pre-cache the dataset to tmpfs or use DataLoader(prefetch_factor=4) After applying fixes 1-3, the same training run: Context switches on training process: 14,504 → 890 cudaStreamSync p99: 7.2ms → 45µs (160x improvement) Effective throughput: 200 → 540 samples/sec (2.7x) nvidia-smi still says 97%, but now it's real utilization The Bigger Picture GPU underutilization is a trillion-dollar infrastructure problem hiding behind a misleading metric. Every ML team has hit this wall: training that should take 4 hours takes 12, and nobody can explain why because all the dashboards say the GPU is "fine." The problem is always on the host side: CPU scheduling, data loading, memory pressure, I/O contention. These are Linux kernel events. The only way to see them alongside CUDA behavior is to trace both layers simultaneously. This is what the tracer does: eBPF uprobes on the CUDA libraries plus kernel tracepoints on the scheduler, memory subsystem, and I/O stack. No code changes, no SDK integration, <2% overhead. Production-safe. Try It Yourself No GPU needed to see the pattern: # 1. Build git clone https://github.com/ingero-io/ingero.git cd ingero && make build # 2. Try the demos ./bin/ingero demo cpu-contention # CPU scheduling delays causing GPU stalls ./bin/ingero demo memcpy-bottleneck # Data transfer dominating wall-clock time For real GPU tracing: sudo ./bin/ingero trace --stack --duration 120s # ... run the training job ... ./bin/ingero explain --since 120s GitHub (give us a star!): github.com/ingero-io/ingero . No NVIDIA SDK, no code changes, production-safe by design. We believe that your GPU utilization metrics can be misleading, and we'd love to help! Drop us an issue on GitHub and we will gladly dive into it together. Ingero is free & open source software licensed under Apache 2.0 (user-space) + GPL-2.0/BSD-3 (eBPF kernel-space). One binary, zero dependencies, <2% overhead. Related reading debugging PyTorch GPU out-of-memory errors 124x slower PyTorch DataLoader traced at kernel level GPU incident response in 60 seconds with eBPF