Archive
Hands-on fsync vs fdatasync vs none: per-write and every-N sync on 256B and 4KiB records. Real p50/p95 show about 100-200x durability tax on localhost
Hands-on nice/ionice lab: victim nice19 under 6 hogs ~5.7× slower (p95 ~7.4s); ionice idle O_DIRECT read ~6.1× under writers. Real numbers, no Docker.
Hands-on TCP_QUICKACK lab: default QA=1 on loopback; WWR set-once ~18 µs; reassert QA=1 under Nagle p95 ~48 ms, OutSegs ~13×. Real numbers, no Docker.
Hands-on ctx-switch lab: process pipe RTT p50 3.2 us (~0.50M sw/s); thread pipe 2.85 us; Event 9.3 us; 6 hogs blow p95 to 19 us. Real numbers, no Docker.
Hands-on asyncio vs ThreadPool lab: sleep 58.7× vs 44×; delayed TCP 13.8× vs serial; fast-echo trap loses; CPU flat. Real localhost numbers, no Docker.
Hands-on flock lab: uncontended LOCK_EX p50 0.33 us; waiter ~50 ms hold; w8 p50 ~36 ms; LOCK_NB 99% fail; LOCK_SH ~6.7x wall. Real numbers, no Docker.
Hands-on TLS handshake lab: Python full HS p50 TLS1.3 1.51 ms vs TLS1.2 1.20 ms (~72× TCP); openssl resume ~1.5× / 4.1×. Localhost RSA numbers, no Docker.
Hands-on stdout buffering lab: full buffering maximizes throughput, while line and unbuffered modes restore live visibility on pipes.
Hands-on ProcessPool vs ThreadPool lab: pure-Python threads lose to processes, hashlib threads help, and sleep I/O favors threads.
Hands-on epoll vs select toy: ready-wait flat ~3.5 µs (8→900 FDs); select scales 4.6→16.9 µs then hits fd≥1024 cliff. Python numbers, no Docker.
Hands-on sendfile vs read+sendall lab: 256 MiB p50 4.41 vs 2.34 GB/s (~1.9×); copy collapses at 4 KiB. Warm localhost TCP measured, no Docker.
Hands-on mmap vs buffered read lab: 256 MiB p50 11.5 vs 5.4 GB/s; O_DIRECT ~0.9 GB/s warm-cache bypass. Fake endpoint-only GB/s disclosed. No Docker.