schedcp

# SchedCP - Automatically Optimize Linux Scheduler with MCP Server > WIP: We are building a benchmark for evaluating the optimizations for OS! SchedCP is an `experimental` project that enables AI optimization of Linux kernel schedulers using the sched-ext framework. It provides e2e automatic scheduler selection/synthesis, workload profiling, and performance optimization without any human intervention or guidance. Paper: [SchedCP: Towards Agentic OS](https://arxiv.org/abs/2509.01245) *The future is not just about letting AI write code for you; the AI agent should act as your system administrator, able to optimize anything for you automatically, without requiring any manual intervention!* It includes the following tools: - **autotune** - AI Agent-powered automatic OS optimization - **schedcp** - MCP server for scheduler management and generation ## Demo Start optimize any workload with AI by simply run: ```sh autotune/target/release/autotune cc "<your workload command>" # example for linux build autotune/target/release/autotune cc "make -C workloads/linux-build-bench/linux clean -j && make -C workloads/linux-build-bench/linux -j" # example for schbench autotune/target/release/autotune cc workloads/basic/schbench/schbench ``` Allow LLM Agent to auto select and config the best scheduler:  Allow LLM Agents to write new schedulers: ## Features & design - Automatic workload profiling - Automatic scheduler selection based on workload characteristics - Performance tracking across different schedulers - Real-time scheduler management and generation  The current MCP tools include: - **list_schedulers** - Get detailed information about all available schedulers - **run_scheduler** - Start schedulers with custom configurations - **stop_scheduler** - Stop running scheduler instances - **get_execution_status** - Monitor scheduler performance and output - **create_and_verify_scheduler** - Create custom BPF schedulers from source code - **system_monitor** - Collect real-time CPU, memory, and scheduler metrics - **workload** - Manage workload profiles and execution history ## Installation ### Requirements - Linux kernel 6.12+ with sched-ext support - Rust toolchain The major dependencies are the dependencies for the sched-ext framework. You can check the [github.com/sched-ext/scx](https://github.com/sched-ext/scx) for more details. You also need to install the deps for the workloads you want to optimize. ### Build ```bash # Clone with submodules git clone https://github.com/eunomia-bpf/schedcp cd schedcp git submodule update --init --recursive scheduler/scx # Build schedulers cd scheduler && make && make install && cd .. # Build autotune cd autotune && cargo build --release && cd .. # Build MCP server cd mcp && cargo build --release && cd .. ``` ## Documentation ### User Guides - **[USAGE_GUIDE.md](document/USAGE_GUIDE.md)** - Complete guide on how to use schedulers - Installation and setup - Using schedulers with CLI, MCP server, and autotune - Understanding scheduler selection - Troubleshooting - **[PROJECT_STRUCTURE.md](document/PROJECT_STRUCTURE.md)** - Detailed project organization - Component overview and responsibilities - Directory structure and file locations - Data flow and integration points - Build system architecture - **[AI_AGENTS.md](document/AI_AGENTS.md)** - AI agent implementation - How observation, planning, execution, and learning agents work - MCP tools and agent capabilities - Example optimization workflows - Integration with Claude and other AI systems ### Design Documents - **[sched-agent-design.md](document/sched-agent-design.md)** - Multi-agent framework design - **[schedcp-design.md](document/schedcp-design.md)** - System control plane architecture - **Research Paper**: [SchedCP: Towards Agentic OS](https://arxiv.org/abs/2509.01245) ## Quick Start **You should run the claude-code on project root directory.** ### Autotune (Recommended) ```bash # Set sudo password export SCHEDCP_SUDO_PASSWORD="your_password" # Optimize any workload ./autotune/target/release/autotune cc "<your workload command>" ``` ### MCP Server check the [.mcp.json](https://github.com/eunomia-bpf/schedcp/blob/master/.mcp.json) for more details. You can just open the claude-code on the ### CLI Tool ```bash export SCHEDCP_SUDO_PASSWORD="your_password" # List schedulers ./mcp/target/release/schedcp-cli list # Run a scheduler ./mcp/target/release/schedcp-cli run scx_rusty --sudo # Monitor system metrics ./mcp/target/release/schedcp-cli monitor ``` For detailed usage instructions, see [USAGE_GUIDE.md](document/USAGE_GUIDE.md). ## Artifact from Paper Artifact for reproducing results from **"Towards Agentic OS: An LLM Agent Framework for Linux Schedulers"** (arXiv:2509.01245). ### Evaluation Workloads #### 1. **Linux Kernel Build** - **Location**: [`workloads/linux-build-bench/`](workloads/linux-build-bench/) - **Run**: `autotune/target/release/autotune cc "make -C workloads/linux-build-bench/linux -j"` #### 2. **schbench** - **Location**: [`workloads/basic/schbench_test/`](workloads/basic/schbench_test/) - **Run**: `autotune/target/release/autotune cc "workloads/basic/schbench/schbench"` #### 3. **Batch Workloads** - **Location**: [`workloads/processing/`](workloads/processing/) - **Run**: See [`workloads/processing/README.md`](workloads/processing/README.md). The schedulers there are already generated by Claude Code. The original version was letting Claude Code generate them first and run the workloads with benchmarks mannually. - Workloads: [workloads/processing/assets](workloads/processing/assets/) - Prompt and testcases: [workloads/processing/scripts](workloads/processing/scripts) - Generated schedulers: [workloads/processing/schedulers](workloads/processing/schedulers) ## Related Projects - [sched-ext](https://github.com/sched-ext/scx) - Linux kernel BPF scheduler framework - [Model Context Protocol](https://modelcontextprotocol.io/) - AI-application integration protocol ## License See [LICENSE](https://github.com/eunomia-bpf/schedcp/blob/master/LICENSE) for details.