Mellanox FDR IB Solution With Nvidia GPUDirect RDMA Technology

Mellanox Technologies, Ltd. announced the next advancement in GPU-to-GPU communications with the launch of its FDR IB solution with support for NVIDIA GPUDirect remote direct memory access (RDMA) technology.

The next generation of NVIDIA GPUDirect technology provides application performance and efficiency for GPU-accelerator based HPC clusters. NVIDIA GPUDirect RDMA technology accelerates communications between GPUs by providing a direct peer-to-peer communication data path between Mellanox’s scalable HPC adapters and NVIDIA GPUs.

This capability provides a decrease in GPU-GPU communication latency and completely offloads the CPU and system memory subsystem from all GPU-GPU communications across the network. The latest performance results from Ohio State University demonstrated MPI latency reduction of 69%, from 19.78us to 6.12us, when moving data between IB-connected GPUs, while overall throughput for small messages increased by 3X and bandwidth performance increased by 26% for larger messages.

"MPI applications with short and medium messages are expected to gain a lot of performance benefits from Mellanox’s IB interconnect solutions and NVIDIA GPUDirect RDMA technology," said Professor Dhableswar K. (DK) Panda, The Ohio State University.

The performance testing was done using MVAPICH2 software from The Ohio State University’s Department of Computer Science and Engineering, which delivers world-class performance, scalability and fault tolerance for high-end computing systems and servers using IB. MVAPICH2 software powers numerous supercomputers in the TOP500 list, including the 7th largest multi-Petaflop TACC Stampede system with 204,900 cores interconnected by Mellanox FDR 56Gb/s IB.

"The ability to transfer data directly to and from GPU memory dramatically speeds up system and application performance, enabling users to run computationally intensive code and get answers faster than ever before," said Gilad Shainer, VP marketing, Mellanox. "Mellanox’s FDR IB solutions with NVIDIA GPUDirect RDMA ensures the highest level of application performance, scalability and efficiency for GPU-based clusters."

"Application scaling on clusters is often limited by an increase in sent messages, at progressively smaller message sizes," said Ian Buck, GM of GPU computing software at NVIDIA. "With MVAPICH2 and GPUDirect RDMA, we see substantial improvements in small message latency and bisection bandwidth between GPUs directly to Mellanox’s IB network fabric."

GPU-based clusters are widely used for computationally-intensive tasks, such as seismic processing, computation fluid dynamics and molecular dynamics. Since the GPUs perform high-performance floating point operations over a very large number of cores, a high-speed interconnect is required to connect between the platforms to deliver the necessary bandwidth and latency for the clustered GPUs to operate efficiently and alleviate any bottlenecks in the GPU-to-GPU communication path.

Mellanox ConnectX and Connect-IB based adapters are IB solutions that provide full offloading capabilities critical to avoiding CPU interrupts, data copies and systems noise, while maintaining high efficiencies for GPU-based clusters. Combined with NVIDIA GPUDirect RDMA technology, Mellanox IB solutions are driving HPC environments to new levels of performance and scalability.

The alpha-code to enable functionality of GPUDirect RDMA is available, including the alpha version of MVAPICH2-GDR release from OSU to enable existing MPI applications.

General availability is expected in the 4Q13.