MorphSched

MorphSched is a workload-aware adaptive resource-control agent that profiles a Linux workload under EEVDF, constructs a workload-specific objective, generates a Scheduler IR, and synthesizes a complete Rust+BPF sched_ext scheduler. Each candidate is built, loaded, executed, benchmarked, and refined using implementation or performance feedback.

Submission contents

• llm/: strategy, Scheduler IR, scheduler generation, feedback, and memory modules.
• scripts/: workload runners, monitoring, result parsing, and plotting utilities.
• bpf/: optional runnable-to-running latency monitor.
• configs/: workload configuration files.
• tools/: aggregate ablation data and plotting tools.
• scx/: upstream sched-ext/scx source plus the scx_gensched generation target. Build artifacts are excluded.
• docs/: anonymized paper, project presentation, and project specification.
• run_demo_redis_*.sh: recording-friendly Redis demo launchers.

Environment

A Linux kernel with compatible sched_ext, BTF, and eBPF struct_ops support is required. Rust, LLVM/Clang, libbpf, and the dependencies documented in scx/INSTALL.md and scx/CARGO_BUILD.md are also required.

Install workload dependencies with:

bash scripts/setup_deps.sh

Create a local LLM configuration before running the workflow:

cp llm/config.example.yaml llm/config.yaml

Fill in the endpoint, model, and API key locally. Never commit or distribute llm/config.yaml.

Demo

Run the Redis plus CPU-stress demonstration with:

./run_demo_redis_cpu.sh

Shorten the benchmark and search budget for a smoke run with:

DURATION=20 MAX_ITERS=3 ./run_demo_redis_cpu.sh

All runtime artifacts are written to runs/, results/, and recordings/; these directories are intentionally excluded from the submission archive.

Current validation boundary

The repository provides foreground parsers for Redis, Nginx, Memcached, and Sysbench. CPU-stress scenarios also have a standardized background bogo-ops parser and can enforce foreground/background joint acceptance. Other interferer runners preserve raw logs, but their standardized background-progress parsers still need to be completed before claiming equivalent joint acceptance.

Privacy and reproducibility

The submission archive contains no API credentials, local absolute paths, raw LLM responses, run logs, build outputs, Python caches, or editor metadata. The included PowerPoint copy has its Office author/editor metadata removed. Third-party source and attribution are documented in THIRD_PARTY.md.