The article explains CUDA’s architecture and SM basics, describes how the warp scheduler and dispatch units assign thread blocks to SMs, and concludes that external control cannot target a specific SM, while mentioning the NanoFlow intra‑device parallelism approach as a possible indirect optimization.
The article explains how Linux builds sched_domain and sched_group hierarchies based on physical CPU topology—sockets, dies, clusters, and NUMA nodes—illustrating load‑balancing (BALANCE) versus affinity (AFFINE) with concrete examples, kernel code references, and QEMU‑based experiments.
Linux will continue supporting AMD Zen 2 processors and improve performance by optimizing L3 cache usage, while the new select_idle_sibling() patch expands idle‑thread selection beyond the local cache, addressing inefficiencies in the kernel’s load‑balancing on multi‑CCD systems.
This article introduces TencentOS Server, a cloud‑native Linux distribution developed by Tencent, detailing its general OS architecture, IaaS usage, cloud‑native scheduling (TCNS, BT, VMF, ECFS), resource QoS (RUE), quality monitoring, service‑level indicators, and integration with Kubernetes to enhance performance, isolation, and resource utilization.
The article explains how kernel‑level load balancing works, why naïve thread‑pool wake‑up can cause CPU imbalance, and proposes using thread affinity, priority layers and a custom layered condition variable to achieve better performance in multithreaded server applications on Linux.
