POSIX SIGSTOP vs. Process Killing: Why Restarting Local LLMs Destroys Developer Velocity

The Developer Velocity Tax: Cold Starts

When an engineer's Mac begins to crawl during a build, the instinctual reaction is to open Terminal and run killall ollama or close LM Studio. This instantly drops memory pressure and lets the compile proceed smoothly.

The problem arises 90 seconds later when the build finishes. You switch back to your editor to test your code, press Tab for an AI autocomplete or query your local assistant, and... nothing happens. The background server must cold-start from disk:

1. The executable binary loads and initializes CUDA/Metal compute contexts (~1.5s).
2. The 16GB GGUF model file is mapped from SSD into unified memory (~8.0s).
3. Compute graphs are built and memory buffers allocated (~2.5s).
4. Total latency before your first token: 12 to 18 seconds.

Over an 8-hour workday with 20 build cycles, this cold-start tax wastes over 5 minutes of dead time every day, while fracturing developer focus.

The POSIX Solution: Suspending Without Evicting

Unix systems have included the perfect primitive for this problem since the 1970s: POSIX Job Control Signals.

// How POSIX signals control process execution
kill(pid, SIGSTOP); // Halts execution immediately; process remains in memory
kill(pid, SIGCONT); // Resumes execution from exact instruction pointer

Unlike SIGTERM or SIGKILL, SIGSTOP cannot be caught, blocked, or ignored by the target process. When sent to Ollama or a local AI runner:

• Thread Execution Freezes: All CPU threads stop immediately. No CPU cycles, no thread scheduling overhead, no thermal throttling.
• Metal GPU Queues Flush: Active GPU command buffers complete, and no new compute shaders are dispatched. The Metal GPU is completely surrendered to other processes.
• Weights Remain Paged: The tens of gigabytes of model weights stay mapped in physical RAM. No SSD reads or reallocations are required.

Microsecond Resume Latency

When ContextWarden detects that your compiler or build task has terminated, it dispatches SIGCONT:

# Resuming execution
$ kill -CONT 

Because the process state and memory buffers are intact, the scheduler resumes the threads in less than 2 milliseconds (0.002 seconds). The developer experiences zero lag, zero wait time, and uninterrupted AI assistance.