softwaredefined
Search documents
DDN Infinia on OCI: High-Performance AI Storage
DDN· 2025-11-11 18:56
Performance Overview - DDN Infinia demonstrates excellent performance in Oracle Cloud Infrastructure (OCI) with a small six-node cluster [7] - Achieved a consistent 5 milliseconds Time To First Byte (TTFB), which is excellent for S3 object IO [6] Throughput Metrics - Achieved approximately 30 GB/s of put throughput during object population [5] - Each client and Infinia node processed puts at roughly 5 GB/s [5] - Sustained approximately 37.5 GB/s of get throughput during the get benchmark [6] - Load was evenly distributed across all clients and Infinia nodes at around 6.5 GB/s of throughput during get operations [6] Infrastructure and Configuration - The test used six BM dense ioe5 compute instances as hosts for the Infinia cluster [2] - Six BM standard E5.192 instances with single 100 GB connections were used for the clients to avoid networking bottlenecks [2] - Only 32 out of the 128 cores available in the dense ioe5 instances were utilized for the Infinia software [2] - DDN is investigating other OCI instances to prevent overallocation of hardware [3] Technology and Architecture - Infinia architecture provides capabilities for data management, including data IO paths, object file querying, scale-out KV store, always-on encryption, and data reduction [2] - Infinia is fully software-defined and containerized, enabling it to run on physical or virtualized hardware with Intel, AMD, or ARM processors [2] - Implemented high-performance eraser coding, custom fall domains, and the ability to use both TLC and QLC flash [2] Testing Methodology - IO generation was performed using warp in distributed benchmarking mode to ensure a full mesh of IO across all clients and Infinia cluster nodes [3] - Parallel warp was used across all six clients and six Infinia nodes during the put and get tests [4][5][6] Disclaimer - The information presented is for potential future integrations and is a tech preview [1] - The overall capabilities, including the performance of this feature, can and will change [1] - No timelines for delivering this capability should be inferred from this demo [1]