NVIDIA proposes sysfs interface for user-space CPU core preference control
A writable sysfs interface would let user-space daemons like Thermald steer the Linux fair scheduler toward or away from specific CPU cores based on thermal conditions, workload type, or other external factors.
Summary
NVIDIA engineer Yury Norov has posted a request-for-comments patch series that adds a writable sysfs control for preferred CPU cores, extending the preferred CPU mask infrastructure merged in Linux 7.4 for virtual machine steal-time governance. The proposed interface at /sys/devices/system/cpu/cpu[0-9]*/preferred would allow user-space utilities to communicate core preferences directly to the kernel scheduler, complementing the existing in-kernel steal governor with externally-driven policy decisions.
The patches target heterogeneous and contended environments such as servers running both AI training and inference VMs, where latency-sensitive inference workloads could retain more preferred vCPUs during contention. Thermal management is another immediate use case: Thermald or similar daemons could idle hotter cores by removing them from the preferred set. The aggregate preferred core list would also be readable at /sys/devices/system/cpu/preferred.
Scheduler interface and sysfs design
- sysfs path design: Per-CPU writable files at /sys/devices/system/cpu/cpu[0-9]*/preferred plus a global read-only aggregate at /sys/devices/system/cpu/preferred expose both individual and system-wide preferred core state.
- Steal governor extension: The work builds on the preferred CPU mask added in Linux 7.4's new steal governor for VMs, converting what was an internal kernel mechanism into a user-space controllable policy lever.
- Fair scheduler integration: Preferred cpumask serves as a hint to the CFS (Completely Fair Scheduler) for task placement decisions under CPU contention and vCPU overcommitment scenarios.
- Thermal and workload policy: The interface enables dynamic core preference changes from daemons like Thermald based on temperature sensors, or from orchestrators aware of VM workload roles such as training versus inference.
- RFC stage: The patch series is currently under review on the Linux kernel mailing list and has not been merged into a mainline release.
Why it matters
This interface would give Linux distributions and cloud operators a standardized, kernel-native mechanism for user-space-driven power and performance management without requiring per-driver or per-platform hacks. For data centers running mixed AI workloads and for Steam Deck-like thermally constrained devices, the ability to dynamically shift load away from hot cores through common tools like Thermald could improve sustained performance and hardware longevity.
Original reference: www.phoronix.com →