schedcp
by eunomia-bpf
Overview
Develop, evaluate, and dynamically manage custom eBPF-based CPU schedulers for Linux, particularly focusing on optimizing long-tail and memory-intensive workloads (like AI/ML, I/O, distributed processing).
Installation
cargo run -p mcp -- --socket-path /tmp/schedcp.sock --storage-path ./dataEnvironment Variables
- CUDA_MANAGED_FORCE_DEVICE_ALLOC
- LD_PRELOAD
- SCHEDCP_CLI_PASSWORD
Security Notes
The server component (MCP) allows remote creation and compilation of new eBPF schedulers from provided source code. This is a critical remote code execution vulnerability if the input source code is not thoroughly sanitized and verified in a robust sandbox environment. Compiling and loading eBPF programs grants kernel-level privileges. While eBPF programs are subject to kernel verifier safety checks, logical flaws in custom schedulers could still lead to denial-of-service or unintended resource allocation that impacts system stability or fairness. The use of `LD_PRELOAD` in certain workload scripts (`ktransformers`) presents a high risk for arbitrary code injection, although this appears to be part of a specific workload's optimization rather than the core scheduler management. Running `sudo` commands for scheduler management (as indicated in `schedcp-cli run`) is inherent to kernel-level operations but elevates the risk if the `mcp` server or its client (`schedcp-cli`) is compromised.
Similar Servers
mcp_massive
An AI agent orchestration server, likely interacting with LLMs and managing multi-agent workflows.
asya
A microservices platform for orchestrating asynchronous, event-driven AI/ML workflows via an MCP JSON-RPC gateway.
nettune
This server acts as an MCP (Model Context Protocol) stdio client wrapper for the 'nettune' TCP network optimization tool, enabling AI-assisted network diagnostics and configuration on Linux servers.
a2a
AI agent for multi-cluster Kubernetes management, enabling workload distribution and operational tasks across KubeStellar environments.