dataloader

Left panel showing the agent's four-step investigation of rank 7 slowdown, right panel showing the raw eBPF kernel event stream that produced the causal chain, MCP arrow linking the two

eBPF, GPU Debugging, GPU Observability, MLOps

What Happens When an AI Agent Gets Kernel-Level GPU Traces

Ingero Team / April 16, 2026

Claude Code with MCP access to kernel-level GPU traces, pointed at a real performance anomaly. The full investigation session: what the agent called, what it found, and where it got things right and wrong.

Stats display showing 200000 context switches, 300000 page allocations, 3676 sched_switch events over a 40 second trace window, explaining why DataLoader runs 124x slower than direct tensor indexing

eBPF, GPU Debugging, GPU Observability, MLOps

124x Slower: What PyTorch DataLoader Actually Does at the Kernel Level

Ingero Team / April 1, 2026

PyTorch DataLoader with 8 workers: 124x slower than direct tensor indexing. Kernel-level tracing of the full execution path revealed CPU over-subscription, context switching storms, and CUDA kernel launch latency spikes up to 356x above baseline.