Linux Kernel Journey
This article explains the fundamentals of CUDA performance tuning, covering GPU architectures from Kepler to Volta, the role of SMX, warp schedulers, registers and memory hierarchies, and provides practical guidance on launch configuration, latency hiding, and thread‑block sizing to maximize throughput.
This article explains the technical principles behind Linux kernel tracing, covering static tracepoints and dynamic kprobes, demonstrating their use with ftrace and perf, and detailing the underlying macro implementations and low‑level mechanisms that make kernel source tracking possible.
This guide explains Linux's virtual‑physical memory model, shows how to inspect system and process memory via /proc files, demonstrates a 100 MiB allocation test, and provides a step‑by‑step workflow—including ps sorting and Valgrind usage—to locate and fix memory leaks.
The article traces the evolution of Linux CPU scheduling from CFS to the EEVDF algorithm, explains EEVDF’s deadline‑driven allocation with concrete task examples, and details Zang Chunxin’s patch that enables shorter slices and wake‑up preemption to reduce latency.
The article delves into the Linux /sys/bus directory of the sysfs virtual file system, detailing how the kernel registers bus types, organizes devices and drivers, explains core structures like struct bus_type, walks through the device‑add and driver‑probe workflow with concrete USB examples, and demonstrates practical inspection using NanoCode.
The article explains how the /sys/bus directory in the Linux sysfs virtual file system represents registered bus types, details its internal kernel structures such as struct bus_type, describes the bus registration and device‑driver matching process, and demonstrates practical inspection using NanoCode with concrete examples like USB storage devices.
The article examines Linux's page‑cache writeback mechanism, explains why the single‑threaded writeback path becomes a bottleneck under heavy writes, and details Vivo's multithreaded writeback patches—including inode‑to‑context mapping and sysfs‑controlled thread counts—that achieve up to 2.4 GB/s on XFS and a 22 % speedup on F2FS, while also discussing fragmentation trade‑offs and optimal thread‑to‑allocation‑group ratios.
The 20th China Linux Kernel (CLK) Conference in Shenzhen gathered nearly 500 developers on‑site and over 170,000 online viewers, showcasing five technical sub‑forums, keynote insights on AI‑driven kernel challenges, RISC‑V adoption, and a 31‑fold rise in Chinese kernel contributors, underscoring the rapid evolution of China’s kernel ecosystem.
The article examines how Huawei’s ModelFS makes prefetch and memory reclamation programmable and how vivo adapts F2FS to support uncached buffer I/O, enabling AI model weights to be loaded without lingering in the page cache and eliminating kswapd overhead.
This tutorial explains how to leverage Linux kernel tracepoints with eBPF and bpftrace to capture real‑time GPU driver activity—including job scheduling, memory management, and command submission—across Intel, AMD, Nouveau, and NVIDIA GPUs, providing detailed examples, scripts, and analysis of the resulting data.
