Posts on mc · noteshttps://hk.crepuscule.uk/posts/Recent content in Posts on mc · notesHugoen-usMon, 01 Jun 2026 09:00:00 +0800Reading list — June 2026https://hk.crepuscule.uk/posts/reading-2026-06/Mon, 01 Jun 2026 09:00:00 +0800https://hk.crepuscule.uk/posts/reading-2026-06/A few things that crossed my desk this month.Data-parallel training: gradient bucketing and overlaphttps://hk.crepuscule.uk/posts/grad-bucketing/Tue, 19 May 2026 17:40:00 +0800https://hk.crepuscule.uk/posts/grad-bucketing/Why DDP feels like magic until you look at the allreduce schedule.Reading notes: QUIC, 0-RTT and why HTTP/3 is UDPhttps://hk.crepuscule.uk/posts/quic-http3/Wed, 29 Apr 2026 20:25:00 +0800https://hk.crepuscule.uk/posts/quic-http3/Moving HTTP onto UDP was not insane — head-of-line blocking is the reason.A minimal, reproducible Docker setup for ML experimentshttps://hk.crepuscule.uk/posts/docker-ml/Wed, 11 Mar 2026 14:00:00 +0800https://hk.crepuscule.uk/posts/docker-ml/Pin everything, mount data, never bake it. The boring setup that stopped me losing runs.KV cache, and why LLM inference is memory-boundhttps://hk.crepuscule.uk/posts/kv-cache/Sun, 08 Feb 2026 15:20:00 +0800https://hk.crepuscule.uk/posts/kv-cache/The cache that makes autoregressive decoding fast also makes it the thing that runs out of memory first.eBPF: tracing syscalls without a kernel modulehttps://hk.crepuscule.uk/posts/ebpf/Mon, 12 Jan 2026 18:50:00 +0800https://hk.crepuscule.uk/posts/ebpf/Attaching a verified program to a tracepoint, and why this beats strace under load.Tuning TCP BBR and fq on a high-latency linkhttps://hk.crepuscule.uk/posts/bbr/Mon, 03 Nov 2025 22:15:00 +0800https://hk.crepuscule.uk/posts/bbr/Switching off loss-based congestion control on a long-fat path, and the fq gotcha for UDP.WireGuard: why it's so smallhttps://hk.crepuscule.uk/posts/wireguard/Sun, 14 Sep 2025 16:30:00 +0800https://hk.crepuscule.uk/posts/wireguard/~4k lines, one cipher suite, no negotiation. The design choices that make it auditable.The TLS 1.3 handshake, in plain termshttps://hk.crepuscule.uk/posts/tls13/Fri, 18 Jul 2025 19:05:00 +0800https://hk.crepuscule.uk/posts/tls13/One round trip instead of two, and why the ServerHello already carries a key share.A first, honest look at io_uringhttps://hk.crepuscule.uk/posts/io-uring/Thu, 22 May 2025 20:40:00 +0800https://hk.crepuscule.uk/posts/io-uring/Two ring buffers, one syscall, and the mental model that finally made it click.Notes on cgroups v2: memory limits that actually holdhttps://hk.crepuscule.uk/posts/cgroups-v2/Sun, 09 Mar 2025 21:10:00 +0800https://hk.crepuscule.uk/posts/cgroups-v2/Why my OOM kills moved around after switching to the unified hierarchy, and the three knobs that matter.