We’re excited to present a breakthrough from the CLEVER Project: a novel RDMA over InfiniBand Communicator that significantly accelerates GPGPU virtualization using GVirtuS — enabling smarter, faster, and more efficient AI workloads across remote GPUs ⚡🧠
Traditional TCP/IP-based GPU virtualization is plagued by context switches and latency. With the new RDMA Communicator, we slash those overheads and enable up to 82% faster performance on real HPC systems.
🧪 Key Results:
- 35–55% performance gain over TCP/IP
- 10x fewer context switches = lower CPU load
- Seamless integration into GVirtuS plug-in architecture
- Tested on real CUDA workloads: Matrix Multiplication & SAXPY
💡 The RDMA-enhanced GVirtuS now delivers high-speed CUDA offloading with:
- Pre-registered memory regions 🧠
- Polling-based completions for low latency
- Real-time optimized buffer management
Read Zenodo paper to see how the RDMA Communicator stacks up in matrix multiplication across different communication methods 👇
(Lower = Better Execution Time)
🌐 Powered by the CLEVER Project: https://www.cleverproject.eu
Read zenodo paper: https://zenodo.org/records/14717622
