kernel-dvfs-agentic-latency

Linux Kernel Slowpath Control Plane for Agentic AI: DVFS, cpuidle, IRQ path, and scheduler wakeups.

This repository is a systems research project exploring the Agentic Slowpath Control Plane. It addresses a fundamental impedance mismatch between modern Linux power/scheduling subsystems and the unique execution patterns of AI agents.

Thesis

Agentic AI workloads (coding agents, retrieval systems, robotic control loops) run in tight Wait → Wake → Compute → Repeat loops.

Linux subsystems today are reactive: they wait for utilization to increase before ramping frequency, and they enter deep sleep states as soon as a task blocks. For an agentic workload with thousands of short steps, this creates latency amplification: a massive cumulative delay incurred at every single transition.

Latency Amplification

Every agent step is gated by a sequence of kernel slowpaths:

1. IRQ/Timer/Event completion
2. Workqueue/kworker deferred execution
3. Scheduler wakeup and task placement
4. DVFS (schedutil) frequency ramp-up
5. cpuidle C-state exit latency

If each transition takes 100µs and the agent performs 10,000 steps, the kernel introduces 1 full second of cumulative latency that is completely invisible to traditional "performance tuning."

Kernel Control Path

Our research proposes an integrated control plane that spans the entire vertical path:

Patch Series Overview

This RFC implements a cross-subsystem "agent latency window" abstraction:

• 0000: Core Control Plane (DVFS boosts, cpuidle guards, scheduler wakeup window).
• 0001: Deferred Completions (Workqueue and Timer latency attribution).
• 0002: Asynchronous I/O (io_uring and blk-mq attribution).
• 0003: Memory Subsystem (Page fault and page cache hit/miss tracing).
• 0004: VFS Metadata (Path lookup, open/stat, and dentry cache attribution).
• 0005: Enforcement (Cgroup-level latency budgeting and token buckets).

Observability

We provide a comprehensive suite of eBPF (bpftrace) tools to measure:

• Wakeup-to-run latency
• DVFS ramp timing after agent wakeup
• Workqueue execution delay
• Page cache hit/miss ratios in agent loops
• VFS metadata resolution slowpaths

Latency Budgeting

To prevent mechanism abuse, agent latency behavior is gated by a cgroup token budget. This ensures that latency-sensitive execution remains bounded and does not permanently override power management.

Status

• Phase: Research / RFC (v2)
• Target: Linux Kernel (ARM/x86_64)
• Experimental: All changes are gated behind CONFIG_AGENT_LATENCY.

---

Note: This is a research artifact. It is not intended for upstream LKML submission in its current form.