MCP Hackathon Tuner

For the MCP hackathon, we pick a simple example having an LLM optimize some code for the stream benchmark. The code is written in C, with an associated makefile in the folder benchmark. We can start with things like having the compiler tweak the flags for gcc. Then we can ask it to consider some actual hand written optimizations.

Building

You can build the docker container: Build the Docker image:

docker build -t lj_benchmark .

Then verify everything is good by typing

python utils/test_mcp.py

Connecting to MCP Clients

If that works, you can modify ~/.gemini/settings.json. Make sure to use absolute paths. Add in the mcpServers section.

    "lj": {
      "name": "lj",
      "type": "stdio",
      "command": "docker",
      "args": [
        "run", "-i", "--rm", "lj_benchmark"
      ]
    }

For claude code, you can put the following into a .mcp.json file of the current directory.

{
  "mcpServers": {
    "lj_benchmark": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "lj_benchmark"]
    }
  }
}

Running Over HTTP (Remote Server)

The MCP server supports both local (stdio) and remote (HTTP) transports.

Using Docker:

docker run -p 8000:8000 lj_benchmark python3 server.py --transport http --host 0.0.0.0 --port 8000

Using Singularity/Apptainer:

# Download the .sif file from GitHub Actions artifacts, then:
apptainer exec --pwd /app stream_benchmark_amd64.sif python3 server.py --transport http --host 0.0.0.0 --port 8000

For ARM systems, use stream_benchmark_arm64.sif instead.

Use with R-CCS Cloud

Connect to the R_CCS cloud and reserve a job in interactive mode.

srun -p fx700 --ntasks=1 --cpus-per-task=48 --pty bash

Start the server:

singularity exec --pwd /app stream_benchmark_arm64.sif python3 server.py --transport http --host 0.0.0.0 --port 8002

Now in a different terminal on your machine, connect again to the cloud machine.

ssh -L 8002:fx01:8002 login.cloud.r-ccs.riken.jp

Replace fx01 with the name of the compute node you actually got.