Imagine a large Internet company called "Data Center Story" in a busy data center. The company has been using traditional CPU architecture to process massive amounts of data. However, as the amount of data continues to grow, the CPU architecture has encountered many problems and bottlenecks, which makes "Data Center Story" very distressed.
One day, "Data Center Story" invited a mysterious friend, the magician DPU, who brought a brand new solution. The magician DPU waved his magic wand and handed over some workloads in the data center to DPUs dedicated to handling these tasks. In this way, the CPU can be relieved and focus on processing more complex computing tasks.
"Data Center Story" thinks this Magician DPU is very magical, but Magician DPU said: "That's not all, I have a more magical friend-Super Magician CSD."
As a result, Super Magician CSD also joined the "Data Center Story" and brought a brand new Computational Storage Drive (CSD). This drive combines computing and storage to make data storage more efficient. With the help of Super Magician CSD, the data center of "Data Center Story" becomes more powerful, flexible and efficient.
The traditional data center architecture is mainly CPU-centric, which makes the computing and storage tasks of the data center mainly rely on the processing power of the CPU. However, as the amount of data continues to grow, this architecture gradually reveals some problems and bottlenecks, such as slow processing speed and low efficiency.
At present, the traditional CPU-centered architecture has begun to transform into a new data-centered architecture. With the new architecture processing requirements, a variety of products have emerged that offload the computing power of traditional CPUs, such as DPU, CSD, etc.
The emergence of DPU provides innovative ideas for data-centric computing architecture. It mainly shares the work of other processors in the data center, such as network offloading, computing offloading or data service offloading, etc., to save costs, especially reducing the capital cost of entering the data center and reducing the operating cost of the data center. By offloading and isolating infrastructure services, DPU can offload the workload of the CPU and GPU, allowing the CPU and GPU to focus on processing core computing tasks and improve overall performance. In addition, DPU also uses hardware acceleration technology to process these services at a faster speed, thereby greatly improving the efficiency of the data center.
For example, a BlueField-3 DPU can provide data center infrastructure services equivalent to those achieved by up to 300 CPU cores.
Similarly, CSD emerged to optimize data center performance and efficiency. By combining computing and storage, CSD can leverage hardware computing acceleration engines to integrate storage processing tasks onto disks, freeing up the CPU for higher-value work. This architecture improves storage efficiency and performance.
In the field of computational storage CSD, Samsung and ScaleFlux are the two major players in the industry, and both have made certain research and development and application progress in computational storage drives (CSD).
Samsung SmartSSD Computational Storage Drive (CSD) is the industry's first customizable and programmable computational storage platform that can push computing to where the data is, thereby significantly accelerating data-intensive applications by more than 10 times. SmartSSD CSD can accelerate a variety of applications, including database management, video processing, artificial intelligence layer and virtualization, etc.
At its core is the Xilinx Adaptive Platform, a fully customizable computational storage device created using the programmability of Xilinx FPGAs. Perform high-speed calculations on data where it is stored, and ultra-fast parallel computing frees up the CPU to handle other high-level tasks more efficiently. Not only that, another advantage of SmartSSD CSD is that it can improve storage efficiency and save storage space without sacrificing performance or storage capacity.
ScaleFlux also launched a computational storage SSD called CSD 3000. It uses a hardware computing acceleration engine to integrate storage processing tasks onto the disk, thereby freeing up the CPU for higher-value work, extending service life, and improving overall performance. Its core design is a processing method optimized for flash storage, with transparent compression when data is written and seamless decompression when data is read, which reduces the burden on the CPU. This transparent compression technology can increase the amount of stored data by up to 4 times, thereby improving storage efficiency and saving storage space.
In addition, CSD 3000 optimizes the data writing method by using variable length mapping and write aggregation technology, reducing the number of erase and write operations on flash storage, thereby effectively extending the use of SSD life.
In addition, there are some emerging companies and teams that are also actively engaged in R&D and innovation in the field of computing and storage.
At the same time, the current computing and storage field is also forming an industry ecosystem under the leadership of two major organizations, SNIA and NVME. "Computational Storage Architecture and Programming Model Version 1.0" has been released in August 2022.
Currently, 258 members have joined the SNIA working group, and the construction of the industry ecosystem is in full swing.
The NVME specification has also established a corresponding working group, and the corresponding function/command specifications are also being discussed and formulated.
SNIA defines Computational Storage as a technology that integrates computing and storage resources to support more efficient and smarter data centers. Under this definition, Computational Storage emphasizes integrating computing and storage functions into a unified system to improve data management and processing efficiency.
NVME defines Computational Storage as a technology that converts storage devices into computing resources to support more efficient and flexible data centers. Under this definition, NVMe emphasizes converting storage devices into programmable computing resources to achieve more efficient data processing and management. The NVMe specification focuses on performance optimization and flexibility of storage devices.
Although SNIA and NVMe differ in their definitions of Computational Storage, their focus is on integrating storage and computing resources more effectively to improve data center efficiency and flexibility. SNIA focuses more on the integration of storage and computing, while NVMe focuses more on performance optimization and flexibility of storage devices. Both definitions have value in practical applications and can complement each other.
In general, computing storage is a field that is developing very fast, and the current R&D status in the industry shows a situation where a hundred schools of thought are contending. Each company has its own R&D direction and technological advantages, and competition will become more intense in the future.