In recent years, high-performance computing (HPC) and artificial intelligence (AI) have revolutionized how organizations process and analyze massive amounts of data, allowing them to solve complex problems faster and more accurately. However, legacy Enterprise storage infrastructure can create bottlenecks and limit the ability to move data to where GPUs are available. Consequently, organizations are racing to modernize their data architectures while leveraging as much existing infrastructure investment as possible. 

Building, training, and iterating with effective AI models require access to large amounts of data and extreme performance to feed massive GPU clusters that process the data. Legacy NAS architectures struggle to meet the requirements of broad-based enterprise AI, machine learning, and deep learning initiatives and the widespread rise of GPU computing both on-premises and in the cloud. So, what are the alternatives?

Seeking Order Out of Data Chaos

The use of data analytics, BI applications, and data warehouses for structured data is a mature industry, and the strategies to extract ongoing value from structured data are well known. However, the emerging explosion of generative AI (GenAI) and related deep learning (DL) technologies now promises to extract hidden value from unstructured data.

Such AI workloads will not only help data owners figure out what they have and should keep or can discard, but AI/DL use cases also hold the promise of helping enterprises uncover new value previously hidden in large volumes of unstructured file data. In this way, organizations can finally utilize both their structured and unstructured digital assets to extract business value and drive operational efficiencies.

The problem is that the barriers separating unstructured data silos are a serious limitation to how quickly organizations can implement AI initiatives without costs and complexity spiraling out of control. They need the flexibility to utilize any or all data from across multiple, incompatible storage silos to feed GPU-powered compute clusters for AI/DL workflows. This capability traditionally has meant consolidating files into net-new high-performance storage repositories, which adds a significant cost burden to such projects.

These large capital and operational costs slow down the implementation of AI initiatives and raise serious questions about the anticipated return on investment of such efforts. Organizations simply can’t afford to discard existing infrastructure and migrate their unstructured data to a new platform to implement an AI strategy.

Solving the Silo Problem to Unlock the AI Puzzle

There are multiple phases in AI workflows, and of course, different AI/DL use cases can vary greatly depending on the industry or desired outcome. For unstructured data, in particular, medical image analysis for disease detection will differ from activity recognition in video data or sentiment analysis in text data to determine targeted ad placement. Inferencing workloads for analyzing satellite imagery for crop yields or to drive decisions on irrigation or water management will differ from prediction models used on video and other sensor data for refining autonomous vehicle behavior or streamlining manufacturing automation.

But as varied as the AI use cases are, the common denominator is the need to collect data from many diverse sources and often different locations. The fundamental problem is that access to data by both humans and AI/DL applications is always funneled through a file system at some point. This task is done via the metadata in the file system, which is the interface between the raw data and the file structure that users/applications see.

The issue is that since the introduction of network-attached storage (NAS), file systems have been embedded within each vendor’s storage platform. Even though different vendors will present the file/folder structure via industry-standard NFS or SMB file access protocols, the underlying file systems that contain these metadata are siloed into vendor-specific variations that are incompatible with each other.

This problem is particularly acute for AI/DL workloads, where a critical first step is consolidating data from multiple sources to enable a unified view. AI workloads must have access to the complete dataset to classify and/or label the files as the first step to figuring out which should be pipelined into the process. This requirement adds cost and complexity to AI workflows where access to consolidated datasets is necessary.

Feeding the GPU Beast: How to Fully Utilize GPU-Based Computing to Power AI Pipelines

 But bridging silos alone to get global unified access to an organization’s data is not enough to solve the AI problem. The key challenge is how to feed high-performance GPU clusters that may be on-premises, in the cloud, or at remote sites without exacerbating the data copy and silo problem. That is, how to feed GPU-based computing without requiring organizations to consolidate all their data into a bespoke high-performance (and expensive) new data silo.

 Hyperscale NAS is a new category of NAS architecture, based on open standards available in all standard Linux distributions used in the industry today. Not only does this mean organizations can leverage data in place to feed on-premises or cloud-based GPU clusters for AI workloads, but they can even accelerate their existing scale-out NAS platforms.

Marc Staimer of theCUBE Research explains that “Hyperscale NAS is not a product; it’s a new NAS architecture model based on open standards available in all standard Linux distributions used in the industry today. Within the Linux kernel – on servers and hypervisors in practically every data center – are fundamental elements to enable standards-based HPC-level parallel file system performance on commodity hardware and even existing legacy storage types from any data center or cloud vendor. Specifically parallel NFS v4.2, with the pNFSv4.2 client, Flex Files, and NFSv3.”

Hyperscale NAS is the first architecture that combines the extreme performance and linear scale of HPC parallel file systems with standards-based connectivity, simplicity, and enterprise NAS features. It is ideally suited for powering GPU computing at scale – for use cases such as generative AI training – and well-suited for any use case requiring parallel processing and high throughput, low-latency data access.

Whereas previous NAS architectures were about scaling capacity, hyperscale NAS is about scaling performance and capacity—to feed GPU clusters, CPU clusters, and other use cases that require high-speed parallel data processing at scale. Where traditional scale-out NAS architectures start to hit a performance plateau as data volumes scale, hyperscale NAS performance scales linearly to thousands of storage nodes.

 AI workloads that may consist of data placement to centers of excellence for cleansing, movement to a remote data center for training workloads, or transfer to high-performance computing resources in the cloud or another site for inferencing workloads can all be automated in the background without disrupting user or application access, even on live data.

This capability is particularly important because AI workloads typically require several iterations through multiple independent datasets. For inferencing workloads, this will often require an HPC-class infrastructure with GPU clusters that may be on-premises or part of a temporary cloud-based resource cluster set up for the job. At each step in the process, additional custom metadata tags can be automatically applied that identify the algorithm used or other variables needed to track the results or recreate the workflows to iterate the runs.

In addition, since many industries, such as pharma, financial services, or biotechnology, require training data and the resulting models to be archived for compliance reasons, the ability to automate the placement and protection of these data into low-cost resources is critical. With custom metadata tags tracking data provenance, iteration details, and other steps in the workflow, recalling old model data from archives (for reuse or to apply a new algorithm) is a simple operation that can be automated in the background.

In this way, data scientists may be given direct, self-service control over all stages in the AI pipeline across multiple locations, storage silos, and the cloud without needing to request data retrieval from IT administrators or needing to get into IT infrastructure management themselves. Because the data can be seamlessly accessed from existing storage resources, these workflows can leverage data without replacing legacy storage systems with new infrastructure.

Hammerspace’s Hyperscale NAS, for example, “addresses the complex demands of modern high-performance computing, particularly in AI, machine learning, and GPU-intensive workloads,” according to Steve McDowell, Chief Analyst & CEO of NAND Research. “The proven capability of Hyperscale NAS in AI model training, supporting thousands of storage nodes and GPUs, underscores its potential as a leading solution in this space. It’s not just the performance that makes Hyperscale NAS stand out; it’s also the architecture’s ability to improve the performance of existing NAS systems without requiring modifications.”

In Summary

The rapid shift to accommodate AI/DL workloads has created challenges that exacerbate the silo problems that IT organizations have faced for years. Hyperscale NAS is a fundamentally different NAS architecture that allows organizations to use the best HPC technology without compromising enterprise standards.

Molly Presley, SVP Marketing, brings over 15 years of product and growth marketing leadership experience to the Hammerspace team. She is the host of the Data Unchained podcast and the author of multiple books focused on data strategies. Molly has led the marketing organization and strategy at fast-growing innovators such as Pantheon Platform, Qumulo, Quantum Corporation, DataDirect Networks (DDN), and SpectraLogic. Molly has also served on multiple boards, including SNIA and Women in HPC.  Molly on LinkedIn [email protected]