"Please don't do this disgusting thing... proper hierarchical filesystems have been available since 1965." That was Linus Torvalds' reaction last week when the main set of sched_ext changes was merged into Linux 7.2. The framework promises to make kernel scheduling fully programmable from user space via BPF programs. What infuriated Torvalds, however, wasn't the architecture but something far more mundane: the layout of the new C source files, which he deemed unbefitting a project with half a century of history.

What sched_ext is and why on-prem AI should care

sched_ext (Extensible Scheduler) is an infrastructure that lets developers load custom scheduling policies, written in BPF, directly from user space without patching or recompiling the kernel. The loading mechanism resembles that of kernel modules but operates through the BPF subsystem, offering stronger isolation and safety. The stated goal is to allow experimentation with queuing, balancing, and preemption logic for processes, bypassing the lengthy review and merge cycles of the vanilla kernel.

For teams running LLM inference or training workloads on on-premise hardware, granular scheduling control is an often overlooked lever. Deciding when and how a GPU-bound process gets preempted, or which cores handle the serving thread, can reduce latency and improve resource utilization. sched_ext makes these experiments more accessible, lowering the barrier for teams without deep kernel development expertise.

The clash between innovation and maintainability

The core of Torvalds' fury had nothing to do with performance. The submitted code wasn't organized into nested directories following established practice; instead, files were arranged in an unorthodox, flat manner. Torvalds pointed out that this wasn't a defensible technical choice: hierarchical file systems have existed since the mid-1960s precisely to keep complex codebases maintainable. The message is clear: innovation cannot ignore the basic structural discipline that makes a project as vast as Linux sustainable in the long term.

This isn't the first time the Linux creator has lashed out at development habits he considers sloppy. The community knows that his blessing requires readability and organizational standards that go beyond mere functionality. For a framework like sched_ext, destined to deeply influence kernel behavior, the demand for a clean layout carries even greater weight: code transparency is a form of documentation and a defense against hidden bugs.

Implications for onsite LLM inference

The sched_ext episode signals a trend relevant to those managing clusters for training and inference of large language models on their own premises. The ability to customize scheduling at the kernel level, without leaving the controlled infrastructure, aligns with the pursuit of better energy efficiency and more predictable latency. While sched_ext remains experimental and demands BPF development skills, its official integration into the kernel opens scenarios where AI teams can test scheduling policies tailored to specific GPU–CPU configurations, reducing I/O bottlenecks and optimizing token flow in low-latency serving contexts.

Of course, no public benchmarks yet quantify gains on LLM workloads. But the direction is evident: as interfaces like sched_ext mature, on-premise infrastructure will gain another tool to compete with cloud platforms in flexibility, while retaining full system control. Torvalds' pointed request, harsh as it may seem, is also a reminder: power must be channeled into code that can be read and improved by many, not just its original authors.