Three primitives cover almost every locking decision: spinlocks, mutexes, and futexes. They solve the same problem with very different cost models — burning CPU vs trapping to the kernel vs the hybrid approach modern mutexes actually use.
The Linux network stack tops out at ~2 Mpps per core. DPDK reaches 20 Mpps. The difference isn't speed — it's removing work: no syscalls, no allocations, no demux, no locks. Why kernel-bypass exists and when it earns its complexity.
Compilers reorder. CPUs reorder. Single-threaded code never notices — multi-threaded code crashes mysteriously. A clear walk through acquire and release semantics, what the hardware actually does, and the spinlock that ties it all together.
Two threads, two variables, no shared state — and a 4x slowdown. False sharing is the cache coherence tax you pay when unrelated data lands on the same 64-byte line. How to spot it and fix it.
select and poll were O(n). epoll is O(1). The reason isn't smarter scanning — it's that the kernel pushes ready events into a list as they happen. A walk through the data structures that make it work.
How eBPF changed production debugging — attaching probes to kernel functions, tracing syscalls, measuring latency histograms, and XDP packet processing, all without touching application code or rebooting.
Inside ThreadSafeCLogger: how to build a high-performance logging library that doesn't serialize your threads, using per-thread SPSC ring buffers and careful memory ordering.