NCCL Mesh Plugin

Custom NCCL network plugin enabling distributed ML over direct-connect RDMA mesh topologies.

🎯 What This Does

This plugin enables NCCL (NVIDIA Collective Communications Library) to work with direct-connect mesh topologies where each node pair is on a different subnet. Standard NCCL plugins assume either:

• A switched InfiniBand fabric (all nodes on same subnet)
• TCP/IP networking (slow, high latency)

Neither works for direct-cabled RDMA meshes. This plugin does.

🔧 The Problem We Solved

                    ┌─────────────┐
                    │   Spark-A   │
                    │  (titanic)  │
                    └──────┬──────┘
           192.168.101.x   │   192.168.100.x
              (100Gbps)    │      (100Gbps)
                    ┌──────┴──────┐
                    │             │
              ┌─────┴─────┐ ┌─────┴─────┐
              │  Spark-B  │ │  Spark-C  │
              │ (iceberg) │ │(carpathia)│
              └─────┬─────┘ └─────┬─────┘
                    │             │
                    └──────┬──────┘
                    192.168.102.x
                      (100Gbps)

Three NVIDIA DGX Spark workstations connected in a triangle mesh with direct 100Gbps RDMA cables. Each link is on a different subnet - a configuration NVIDIA never intended to support.

🚀 Results

Metric	Value
Effective Bandwidth	8+ GB/s
Line Rate Utilization	~64%
Topology	3-node triangle mesh
Link Speed	100 Gbps per link

Successfully ran distributed LLM inference (Mistral-7B) across all 3 nodes using NCCL over this custom topology.

⚡ Unified Memory Advantage

On Grace Hopper / DGX Spark systems, the GPU and CPU share the same physical memory via NVLink-C2C. This unified memory architecture means:

• No staging copies - RDMA operates directly on GPU-accessible memory
• GPUDirect-like performance - Without additional kernel modules or configuration
• Simplified memory management - Register once, use everywhere

The 8+ GB/s bandwidth is the real deal, not bottlenecked by GPU↔Host transfers.

🏗️ Architecture

Key Innovations

1. Multi-Address Handle Exchange

   • Each node advertises ALL its subnet IPs in the NCCL handle
   • Connector searches for reachable addresses by subnet matching

2. Subnet-Aware NIC Selection

   • connect() finds the local NIC on the same subnet as the peer
   • Automatic routing without IP forwarding or bridges

3. Background Handshake Thread

   • Eliminates deadlock when both ranks call connect() simultaneously
   • TCP-based QP info exchange runs asynchronously

4. Bidirectional QP Exchange

   • Each connection creates fresh Queue Pairs on both sides
   • No QP reuse across multiple NCCL channels

RDMA Implementation

• Raw InfiniBand Verbs API (libibverbs)
• Reliable Connected (RC) Queue Pairs
• RoCE v2 over Ethernet
• Zero-copy on unified memory systems

📦 Installation

Prerequisites

# Ubuntu/Debian
sudo apt-get install libibverbs-dev librdmacm-dev

# Verify RDMA devices
ibv_devices

Build

git clone https://github.com/autoscriptlabs/nccl-mesh-plugin.git
cd nccl-mesh-plugin
make

Use

export LD_LIBRARY_PATH=$(pwd):$LD_LIBRARY_PATH
export NCCL_NET_PLUGIN=mesh
export NCCL_DEBUG=INFO  # or WARN for less output

# Run your distributed job
python your_distributed_script.py

🧪 Testing

Basic All-Reduce Test

import torch
import torch.distributed as dist

dist.init_process_group('nccl', rank=RANK, world_size=3,
    init_method='tcp://MASTER_IP:29500')

t = torch.ones(1000, device='cuda')
dist.all_reduce(t)
print(f'Result: {t[0]}')  # Should print 3.0

dist.destroy_process_group()

Bandwidth Benchmark

import torch
import torch.distributed as dist
import time

dist.init_process_group('nccl', rank=RANK, world_size=3,
    init_method='tcp://MASTER_IP:29500')

t = torch.ones(1024*1024*64, device='cuda')  # 256MB

# Warmup
for _ in range(5):
    dist.all_reduce(t)
torch.cuda.synchronize()

# Benchmark
start = time.time()
for _ in range(20):
    dist.all_reduce(t)
torch.cuda.synchronize()
elapsed = time.time() - start

print(f'Bandwidth: {(256*20/1024)/elapsed:.2f} GB/s')

🔬 How It Works

Connection Flow

Rank 0 (listen)                    Rank 1 (connect)
     │                                   │
     ▼                                   │
 listen()                                │
 ├─ Create QPs on ALL NICs               │
 ├─ Start handshake thread               │
 ├─ Return handle with all IPs           │
     │                                   │
     │◄──────── handle exchange ────────►│
     │                                   │
     │                                   ▼
     │                              connect()
     │                              ├─ Find matching subnet
     │                              ├─ Create QP on that NIC
     │                              ├─ TCP handshake ──────────►│
     │                                   │                      │
     │◄────────────────────────────────────────── QP info ─────┤
     │                                   │                      │
     ▼                                   ▼                      ▼
 accept()                           Connect QP            [handshake thread]
 ├─ Get QP from queue               to peer's QP          ├─ Accept TCP
 └─ Return recv_comm                     │                ├─ Create new QP
                                         │                ├─ Connect QPs
                                         │                └─ Queue for accept()
                                         │
                                    ┌────┴────┐
                                    │ RDMA OK │
                                    └─────────┘

Subnet Matching

// For each peer address in handle
for (int i = 0; i < handle->num_addrs; i++) {
    uint32_t peer_ip = handle->addrs[i].ip;
    
    // Find local NIC on same subnet
    for (int j = 0; j < num_nics; j++) {
        if ((peer_ip & nic[j].netmask) == nic[j].subnet) {
            // Found matching NIC!
            selected_nic = &nic[j];
            break;
        }
    }
}

⚙️ Configuration

Environment Variable	Default	Description
NCCL_NET_PLUGIN	-	Set to mesh to use this plugin
NCCL_DEBUG	WARN	Set to INFO for detailed logs
NCCL_MESH_GID_INDEX	3	RoCE GID index to use
NCCL_MESH_DEBUG	0	Enable plugin debug output

🚧 Current Limitations

• Single QP per connection - No multi-rail aggregation yet
• No relay routing - Non-adjacent nodes can't communicate (fine for fully-connected mesh)
• RoCE v2 only - Ethernet-based RDMA, no native InfiniBand support

🗺️ Roadmap

• Multi-QP per connection for higher bandwidth
• Adaptive routing for partial mesh topologies
• Performance tuning (inline data, selective signaling)
• Support for non-unified-memory systems with explicit GPUDirect RDMA

🛠️ Hardware Tested

Component	Specification
Nodes	3x NVIDIA DGX Spark
CPU	NVIDIA Grace (ARM64)
GPU	NVIDIA Blackwell
Memory	Unified (NVLink-C2C)
NICs	ConnectX-7 (100GbE)
Cables	Direct-attach QSFP56

📚 References

• NCCL Documentation
• RDMA Aware Networks Programming User Manual
• InfiniBand Verbs API

📄 License

MIT License - see LICENSE file.

🙏 Acknowledgments

Built to connect three DGX Spark workstations that NVIDIA never intended to cluster. Sometimes the best solutions come from ignoring "supported configurations."

---

"The future of distributed AI computing is here." — Mistral-7B, running distributed inference on this very plugin