coldtrace

coldtrace is a small Rust CLI for learning Linux process/runtime behavior by tracing cold starts of ML-ish workloads. You can find a write up about this here.

The first target workload is:

python3 -c 'import torch'

Current version: 0.0.7

What It Does Today

coldtrace is currently a direct process wrapper.

At a high level, it:

1. Parses the child command after --.
2. Creates a run directory under runs/.
3. Starts the child process.
4. Redirects the child's stdout and stderr into log files.
5. Measures wall-clock time from the parent process.
6. Waits for the child to exit.
7. Records the child PID, exit code or signal, argv, cwd, output byte counts, and artifact paths.
8. Writes everything into record.json.

That means this:

coldtrace run -- python3 -c 'print("hello")'

creates something like:

runs/
  20260622T022115.254Z-python3-print-hello/
    record.json
    stdout.log
    stderr.log

The tool is intentionally boring at this stage. It is not trying to be a profiler yet. It is building the reliable parent-process measurement foundation first.

Usage

From the Lima VM:

cd /Users/arjunpherwani/dev/coldtrace
cargo run -- run -- python3 -c 'print("hello")'

With an explicit run name:

cargo run -- run --name torch-import -- python3 -c 'import torch'

The CLI prints the run directory:

runs/20260622T022115.254Z-torch-import

Inspect the artifacts:

cat runs/<run-id>/stdout.log
cat runs/<run-id>/stderr.log
python3 -m json.tool runs/<run-id>/record.json

Docs

• C0 details
• Version log