Expert highlights 3 architecture tips for storage environments primed for AI/ML

In his recent thought leadership article, “Three Architecture Tips for Storage Environments Primed for AI/ML,” Ghassan Azzi, Sales Director, Africa at Western Digital, discussed the transformative impact of artificial intelligence (AI) and machine learning (ML) on data analysis, offering insights and automation capabilities. The data fueling these technologies is spread across data warehouses, data lakes, […]

In his recent thought leadership article, “Three Architecture Tips for Storage Environments Primed for AI/ML,” Ghassan Azzi, Sales Director, Africa at Western Digital, discussed the transformative impact of artificial intelligence (AI) and machine learning (ML) on data analysis, offering insights and automation capabilities.

The data fueling these technologies is spread across data warehouses, data lakes, the cloud, and on-premises data centres, ensuring critical information is accessible and analysable for AI initiatives.

Azzi noted that AI’s proliferation has disrupted traditional business models. Organisations increasingly rely on AI to enhance customer experiences, streamline operations, and drive innovation. To maximise AI’s benefits, it is crucial to adopt advanced storage architectures. NVMe over Fabrics provides the low-latency, high-throughput access needed for AI workloads, accelerating performance and reducing potential bottlenecks. Implementing disaggregated storage offers greater flexibility, enabling the independent scaling of storage and compute resources to maximize utilization. Failure to implement suitable architecture and integrate AI can leave businesses behind in a data-driven world.

The sales director noted that organisations are under constant pressure to extract maximum value from their data quickly and cost-efficiently, without disrupting regular business operations. Azzi explained that relying on commodity storage, whether on-premises or in the cloud, is no longer ideal. High-performance, flexible, and scalable compute environments are needed to support the real-time processing demands of modern AI workflows. Efficient, purpose-built data storage is critical, requiring considerations for data volume, velocity, variety, and veracity.

He added that organisations can now build public cloud-like infrastructures in on-premises data centres, providing the flexibility and scalability of the cloud along with the control and cost efficiency of private infrastructure. Properly architected, these environments offer a more efficient way to support the high-performance, scalable requirements of storage environments primed for AI applications. Repatriating AI/ML datasets to on-premises data centers from the cloud can be an ideal option for organisations operating within certain performance or cost limits.

Azzi outlined three key considerations when building on-premises storage environments suited to the needs of today’s AI/ML-powered world. AI applications require significant compute resources to process and analyse ML datasets efficiently, making the selection of a suitable server architecture crucial, with a focus on the ability to scale GPU resources without creating system bottlenecks. It is also important to include high-performance storage networking that can not only meet but exceed the ever-increasing performance demands of GPUs while providing scalable capacity and throughput to meet learning model data set sizes and performance requirements.

Storage solutions that take advantage of direct path technology enable direct GPU-to-storage communication, bypassing the CPU to enhance data transfer speeds, reduce latency, and improve utilization. Finally, solutions should be hardware and protocol agnostic, providing multiple ways to connect servers and storage to the network, as the interoperability of the infrastructure is essential for building a flexible environment primed for AI applications.

He said that building public cloud-like infrastructures on-premises can offer organizations the flexibility and scalability of the cloud while maintaining the control and cost efficiency of private infrastructure. However, the right storage architecture decisions must be made with AI considerations in mind, providing the necessary combination of compute power and storage capacity for AI applications to operate at the speed of business.

According to Azzi, one way to ensure proper resource allocation and reduce bottlenecks is through storage disaggregation, which allows for independent scaling of storage, ensuring GPU saturation and efficient performance in AI/ML workloads.

Western Digital’s RapidFlex technology, Ingrasys’ ES2100 with integrated NVIDIA Spectrum Ethernet switches, and NVIDIA’s GPUs, Magnum IO GPUDirect Storage, and ConnectX SmartNICs combine to offer the performance, scalability, and agnostic architecture required for building on-premises supercomputing environments for AI/ML applications.

These technologies create a direct data path between NVMe-oF storage and GPU memory, driving high performance and efficient utilization of storage and GPU resources. Western Digital’s proof of concept demonstrates simple independent scaling of storage bandwidth to maximize GPU workloads, achieving over 100 GB/s for multiple NVIDIA A100 GPUs.

Azzi’s insights provide valuable guidance for organizations looking to build efficient, scalable storage environments tailored to the demands of AI and ML applications.