RDMA KV Cache：基于 GPUDirect 的分离式 LLM 推理加速方案

rdma-kv-cache project implements a separated LLM inference architecture, using GPUDirect RDMA technology to achieve zero-copy KV Cache transfer between GPUs, significantly reducing large model inference latency.

With the continuous growth of large language model scale, traditional monolithic inference architecture faces resource imbalance issues—prefill stage is compute-intensive but short, while decode stage is memory-intensive but low in compute density. Separated inference architecture decouples these two stages to dedicated GPU clusters, using high-speed interconnect to transfer KV Cache for more efficient resource utilization.

The system consists of four core components:

• Orchestrator: Request entry, routes requests to prefill nodes, manages node registration and load balancing.
• Prefill Nodes: Run custom vLLM, handle input prompts to generate initial KV Cache, and send it via RDMA.
• Decode Nodes: Run custom vLLM, receive KV Cache via RDMA and perform autoregressive token generation.
• GPUDirect RDMA: Technical foundation, uses Mellanox nvidia-peermem kernel module to enable direct GPU memory access, bypassing CPU with microsecond-level latency.

• vLLM Integration: Uses RDMAConnector module with C++ RDMA engine (exposed via pybind11 bindings) to retain vLLM's scheduling optimizations.
• Zero-Copy Transfer: Eliminates intermediate memory copies (GPU→system→network→system→GPU) by directly registering vLLM memory regions and sending to decode nodes' GPU buffers.
• Deployment Tools: Includes install scripts (install_all_nodes.sh), deployment scripts (smart_deploy.sh), run/stop scripts (smart_run.sh, smart_stop.sh), and verification script (verify_gpudirect.sh).

• Performance Indicators: Prefill stage takes 50-200ms (depending on prompt length), RDMA transfer latency is at microsecond level; benchmark tools like benchmark_client are provided.
• Optimization: Adjust key parameters (KV buffer size, RDMA block size, queue depth) via YAML config; model configs (memory utilization, max sequence length) can be tuned via environment variables.

• Hardware: NVIDIA A100 GPU, Mellanox ConnectX InfiniBand NIC.
• Software: Ubuntu 24.04, CUDA 12.6, Python 3.10+, CMake 3.18+.
• Testing Environment: Developed and tested on Clemson CloudLab cluster.

• Suitable Scenarios: High-concurrency online services (prefill handles many initial requests), long-sequence inference (independent decode scaling), heterogeneous GPU environments (different GPUs for prefill/decode).
• Value: Provides open-source reference implementation for separated inference architecture; GPUDirect RDMA integration is valuable for researchers and engineers pursuing extreme inference performance.