Skip to content

Build.md

Latest commit

 

History

History
121 lines (94 loc) · 3.12 KB

Build.md

File metadata and controls

121 lines (94 loc) · 3.12 KB

Build Instructions

Repository Structure

bf3-bench/
├── README.md
├── vocab/                    # GPT-2 vocabulary files
│   ├── vocab.json            # Token-to-ID mapping (50,257 tokens)
│   └── merges.txt            # BPE merge rules (49,992 rules)
├── src/
│   ├── dpu_client/           # DPU-side code (ARM)
│   │   ├── dpu_client_sample.c
│   │   ├── dpu_client_main.c
│   │   ├── bpe_tokenizer.c   # Real GPT-2 BPE implementation
│   │   ├── bpe_tokenizer.h
│   │   ├── rdma_common.h
│   │   ├── common.c
│   │   └── meson.build
│   └── host_server/          # Host GPU code (x86_64 + CUDA)
│       ├── host/
│       │   └── gpunetio_rdma_client_server_write_sample.c
│       ├── device/
│       │   └── gpunetio_rdma_client_server_write_kernel.cu
│       ├── gpunetio_rdma_client_server_write_main.c
│       ├── tokenizer_kernel.cu
│       ├── tokenizer_common.h
│       ├── embedding_kernel.cu
│       ├── embedding_common.h
│       ├── rdma_common.c
│       ├── rdma_common.h
│       ├── common.c
│       └── meson.build
├── scripts/
│   └── wordpiece_tokenizer.py  # GPU WordPiece using RAPIDS nvtext
├── results/
│   └── *.csv                 # Benchmark results
├── charts/
│   └── *.png
└── docs/
    └── DPU_Tokenization_Offload_Report.md

Requirements

  • NVIDIA DOCA SDK 2.9 or later
  • CUDA Toolkit 12.0 or later
  • Meson 0.61+
  • Ninja build system

Building the Host Server

cd src/host_server
meson setup build
cd build
ninja

The output binary is doca_gpunetio_rdma_client_server_write.

Building the DPU Client

On the DPU (access via SSH to 192.168.200.2):

cd src/dpu_client
meson setup build
cd build
ninja

The output binary is dpu_rdma_client.

Running the Benchmark

Quick Test

# Terminal 1 (Host): Start server
sudo PAYLOAD_SIZE_KB=8 ./doca_gpunetio_rdma_client_server_write \
    -d mlx5_2 -gpu 2a:00.0 --gid-index 1

# Terminal 2 (DPU via SSH): Run client
sudo PAYLOAD_SIZE_KB=8 ./dpu_rdma_client \
    -d mlx5_2 -s 192.168.200.1 -g 1

Mode Selection

Mode Host Flag DPU Flag Description
DPU-Based (default) (default) DPU tokenizes, GPU embeds only
GPU-Based --gpu-tokenize --send-text DPU sends raw text, GPU tokenizes + embeds

Batch Processing

Both host and DPU support batch processing via the BATCH_SIZE environment variable:

# Terminal 1 (Host): Start server with batch_size=4
sudo PAYLOAD_SIZE_KB=8 BATCH_SIZE=4 ./doca_gpunetio_rdma_client_server_write \
    -d mlx5_2 -gpu 2a:00.0 --gid-index 1

# Terminal 2 (DPU via SSH): Run client with batch_size=4
sudo PAYLOAD_SIZE_KB=8 BATCH_SIZE=4 ./dpu_rdma_client \
    -d mlx5_2 -s 192.168.200.1 -g 1
Parameter Range Description
BATCH_SIZE 1-32 Number of sequences to process in parallel