The Development of Data-Centric Computing with Computational Storage

The next step in the evolution of data-centric computing, Computational Storage, improves data and computation localization and scale economies. The goal is to federate data processing by bringing application-specific processing closer to the data instead of transmitting the data to the application. It enhances the overall data center environment by freeing up host CPU and Memory for running client applications, reducing network and I/O traffic by shifting processing to data, and enhancing security by minimizing data migration, reducing the total carbon footprint.

In the previous decade, "Compute Acceleration" emphasized Application Acceleration and AI/ML, leading to an industry emphasis on GPUs, ASICs, AI/ML frameworks, and AI-enabled apps. As a result, GPUs and ASICs are now widely used in AI/ML application cases. The objective of the second phase of data-centric computing was "Network and Storage Acceleration." It began with a focus on FPGAs and SmartNICs and has expanded over the past two years to include DPUs (Data Processing Units) and IPUs (Infrastructure Processing Units).

These allow for the disaggregation of data center physical infrastructure and software services to facilitate logically composable systems and efficient dataflows. DPUs/IPUs are widely deployed in cloud environments and gaining traction in enterprise implementations. Data Acceleration is the next stage in the evolution of data-centric computing, and Computational Storage is the crucial enabling technology.

Computational Storage solutions are designed to bring computation closer to data by maximizing silicon diversity and distributed computing. It will facilitate the transition from today's "data storage" systems to future "data-aware" systems for more efficient data discovery, processing, transformation, and analytics. In the coming years, it will attain the same level of maturity and industry momentum as GPUs/ASICs for AI/ML and DPUs/IPUs for network/storage processing.

Computational Storage Drives (CSD), Computational Storage Processors (CSP), and Data Processing Units (DPU) are enabling technologies for moving data processing to hardware and improving data center deployment economics. FPGAs (Field-Programmable Gate Arrays) will provide a software-programmable element for application-specific processing and future innovation. These are integrated into CSDs and CSPs for application-specific processing with excellent performance.

In the last two years, startups, system manufacturers, solution vendors, and cloud service providers have been involved in developing computational storage solutions. The difficulty is the integration of computational storage interfaces with applications and the widespread availability of storage devices and platforms with hardware acceleration capabilities.

Multiple NVM Express and SNIA efforts are underway to unify the block storage architecture model and command set. SNIA architecture for computational storage includes CSD, CSP, and CSA (Computational Storage Array). A CSA typically consists of one or more CSDs, CSPs, and the software required to identify, manage, and utilize the underlying CSDs / CSPs. Integrated solutions are an illustration of CSA. Since most applications access and store data via files and objects, standardization and open-source initiatives will continue to progress toward the object and file protocols.

Since computation can only be relocated to a place where an application-level context of data exists or is created, computational interfaces will also arise for file and object storage systems. There is potential to expand the file and object access methods to federate application-specific processing closer to the data and send only the results to the application. Integration with developing software-defined databases and data-lake architectures will make the solution transparent to user applications that run on top of the data lake, enhancing the solution's performance and economics.

Increased acceptance of Edge installations gives further opportunity to federate application-specific processing to data-generating Edge locations. Computational Memory is emerging as an adjacent technology to shift computation to data in Memory. This will allow computing in future fabrics with Persistent Memory. HBM (High Bandwidth Memory) would be helpful for GPUs and the data transformation engines embedded into storage devices.

The data operations will have both fixed and configurable functions. Modern storage systems are constructed with a software-defined architecture and containerized microservices. This enables the execution of application-specific computation in the form of protected microservices on horizontally scalable storage system nodes or entirely on a computational storage disc drive or persistent computational Memory. Future databases and data lake architectures will utilize computational storage principles to improve data processing, discovery, and classification.

Conclusion?

In the future, "data storage" systems will grow into "data-aware" systems. These technologies will automatically identify, classify, and transform data based on policies, allowing enterprises to transition from digital-first to data-first. In addition, application-specific data processing will federate closer to the data and optimize the economics of data centers and edge-to-cloud architectures.