Artificial intelligence (AI) is becoming solidified as an important tool for competitive advantage, used by organizations of all sizes and across industries. Legacy network-attached storage (NAS) systems, however, are not equipped to provide the levels of throughput that these workloads require, especially as AI implementations scale. As a result, the business is left with either expensive central processing unit (CPU) or graphics processing unit (GPU) cycles that are not used, and/or with an overprovisioning of costly storage infrastructure.
Cost-effectively rolling AI workloads out across the enterprise, at scale necessitates a new approach. IT planners should consider an architecture that can decouple the one-to-one relationship between host compute and storage media, enabling the AI application to access pooled resources across storage devices, without sacrificing performance.
For its part, DataDirect Networks (DDN) has created an architecture that is worth evaluating as a solution to this problem. DDN has created an application programming interface (API) that effectively parallelizes AI workloads across the storage input/output (I/O) communication path. Consequently, the workload is not directed to a singular storage device. Rather, it is automatically spread out across storage devices that share a unified namespace. Because data and workload traffic are automatically balanced across the storage infrastructure, compute and storage resources may be scaled independently. This improves utilization levels. Meanwhile, application performance is accelerated.
To minimize overhead in the shared storage model, DDN’s architecture calls for compute and storage to be connected via an optimized remote direct memory access (RDMA) protocol. RDMA allows the network adapters to transfer data directly to application memory. As a result, the majority of connectivity functions are offloaded from the host compute node to the network connecters, facilitating faster AI workload performance.
Another value inherent in DDN’s architecture is economical scaling. Organizations may still start small with their AI workloads, and flexibly and independently add resources as they are needed. This meshes with how AI projects are typically introduced in the organization; they typically begin concentrated within a particular team and are scaled across the organization from there as proof of concept is established.
AI has tremendous potential in terms of unlocking new, strategic business insights – but getting there must not break the budget. Sven Oehme, DDN Chief Research officer, recently joined George Crump, Storage Switzerland Founder and Lead Analyst, on our Lightboard for further discussion around what to look for in an optimized underlying infrastructure for AI.