March 15, 2024 — Xinnor, a software storage solutions developer, has announced the release of a white paper titled “Saturating infiniBand bandwidth with xiRAID to keep NVIDIA DGX busy.” The paper details the collaboration between Xinnor and DELTA Computer Products GMBH, a leading system integrator based in Germany, to develop a high-performance storage solution tailored explicitly for AI and HPC tasks.
The white paper showcases the collaborative effort’s key components, including integrating high-performance NVMe drives from Micron, efficient software RAID from Xinnor, and 400Gbit InfiniBand controllers from NVIDIA. With a 2U dual sockets server equipped with 24x 7400 NVMe 15.36 drives from Micron, the solution offers storage capacity of up to 368TB and theoretical access speeds of up to 50GBps.
Central to the white paper is the detailed explanation of setting up the system with xiRAID to saturate the InfiniBand bandwidth, ensuring optimal performance for NVIDIA DGX H100 systems using the NFSoRDMA interface. Additionally, the white paper highlights the capabilities of xiRAID software, offering a range of features tailored to diverse storage needs. By providing a comprehensive instruction manual, the white paper empowers users to achieve optimal and consistent performance across various deployments.
The white paper demonstrates the system’s performance under synchronous and asynchronous file access modes through rigorous testing methodologies, showcasing its ability to maintain stability and data integrity across diverse scenarios. Performance tests conducted locally and over the network using NFSoRDMA protocols illustrate the solution’s scalability and reliability, even during drive failures.
Key conclusions drawn from the white paper include
- The solution’s ability to saturate network bandwidth, optimize DGX H100 utilization and ensure fast flushing and checkpoint execution.
- Unaffected storage performance in the event of drive failures, eliminating the need for resource overprovisioning and minimizing system downtime.
- Support for both synchronous and asynchronous operation modes, with customizable settings to optimize performance for various scenarios and load patterns.
“We are thrilled to unveil our latest white paper, representing a significant milestone in our ongoing commitment to revolutionizing storage solutions for AI and HPC tasks,” said Davide Villa, CRO at Xinnor. “Through our collaboration with DELTA Computer Products GMBH and the integration of cutting-edge technologies from Micron and NVIDIA, we have engineered a solution that meets and exceeds modern AI innovations’ demands. This white paper showcases the power of our xiRAID software and underscores our dedication to providing users with the tools they need to achieve unparalleled performance and efficiency in their deployments.“
For more detailed information, readers are encouraged to access the white paper, available on Xinnor’s blog.
Source: Xinnor