Back to Interests

Performance Engineering

Deep dives into system optimization, profiling, and latency hunting.

When perf Goes Blind

Unified On/Off-CPU Profiling with Blocked Samples

Blocked Samples hooks the Linux scheduler to emit synthetic Off-CPU samples alongside perf's On-CPU data — giving bperf unified profiles and BCOZ causal what-if analysis, all at 1.6% overhead.

Beyond the Flame Graph

How wPerf Finds the Bottlenecks Your Profiler Can't See

wPerf uses graph theory — wait-for graphs, Knot detection, and cascaded redistribution — to find Off-CPU bottlenecks that duration-based profilers miss entirely.

No Frame Pointers, No Problem

How eBPF Profiling Works Without DWARF at Runtime

A deep dive into the opentelemetry-ebpf-profiler codebase — DWARF pre-compilation, three-level map hierarchies, tail call chains, and the engineering decisions behind production-grade continuous profiling.