Mercury HPC RPC Example

This project demonstrates how to use Mercury for RPC communication in an HPC environment.

Prerequisites

• CMake 3.10+
• Mercury HPC library
• C++17 compatible compiler
• spdlog library

Building

mkdir build
cd build
cmake ..
make
make install

Running

1. Create a nodelist.conf file (optional - will be created automatically if not present):

hostname1:8080
hostname2:8080
hostname3:8080

2. Run the server:

./mercury_server <role> <num_threads>

Examples:

# Start a worker with 4 threads
./mercury_server worker 4

# Start a master with 2 threads
./mercury_server master 2

Node Discovery

The system supports multiple methods for node discovery:

1. SLURM environment (automatic if running under SLURM)
2. nodelist.conf file (fallback if no SLURM environment)
3. Auto-registration (if no existing nodes found)

Transport Protocols

Mercury supports multiple transport protocols. This example uses:

• na+sm: Shared memory protocol (default)

Other available protocols:

• ofi+tcp: TCP/IP over libfabric
• ofi+verbs: InfiniBand verbs over libfabric
• na+cxi: Cray/HPE Slingshot

To use a different protocol, modify the address string in main.cpp:

std::string address = "ofi+tcp://" + std::to_string(8080 + i);

Debugging

The application uses spdlog for detailed logging. Log levels can be controlled at runtime:

spdlog::set_level(spdlog::level::debug);  // Most verbose
spdlog::set_level(spdlog::level::info);   // Normal operation
spdlog::set_level(spdlog::level::warn);   // Warnings and errors only

Log format includes:

• Timestamp
• Log level
• Thread ID
• Component name
• Message

Request Types

• HEARTBEAT: Basic alive check
• GET_ROLE: Query node's assigned role
• ECHO: Echo back the received message
• SHUTDOWN: Graceful shutdown of the node

Architecture

The system implements a multi-threaded RPC server with:

• Multiple RPC handlers per node
• Role-based request handling
• SLURM integration
• Automatic node discovery
• Comprehensive logging