Shulou(Shulou.com)11/24 Report--

A few days ago, at the 20th Huawei Global analysts Conference held in Shenzhen, China, Feng Dan, a special professor of scholars from the Yangtze River and dean of the School of computer Science and Technology of Huazhong University of Science and Technology, made an authoritative interpretation of the Diskless architecture of the data center and the development trend of near data processing, and put forward a new technical solution direction for the waste of resources caused by the incomplete separation of computing and storage domains such as traditional cloud storage and big data. Help cloud and Internet data centers reduce costs and increase efficiency, and achieve high-quality evolution of data centers.

Professor Feng Dan mentioned that under the traditional deposit and accounting integration architecture, enterprises usually adopt the server integration architecture in which the application is coupled with the local disk, which has typical problems such as low storage efficiency, invalid data migration, waste of resources, and so on.

1. Computing domain and storage domain are not completely separated, and there is still independent local storage, which is often configured according to the maximum capacity preset, and the storage utilization is low.

two。 The general computing power is used for data processing, the IO efficiency is low, the storage domain is still centered on CPU, and the access performance can not be brought into full play.

3. Data needs to be moved more than 8 times from disk to computing domain CPU. Invalid relocation brings waste of network and computing resources.

4. In the large-scale cluster with thousands of nodes, the ratio of computing power and capacity space is not balanced, and additional data migration is needed during capacity expansion, which not only brings existing business risks, but also wastes investment.

5. With the explosive growth of data, the data reduction strategy for structured data is not completely suitable for unstructured scenarios.

With the development of diversified computing power, high-performance networks and storage media, the data center architecture has evolved from a CPU center to a data-centric Diskless architecture. Under the Diskless architecture, the computing power of a single CPU has developed to a variety of computing power such as GPU, DPU, XPU, and so on, and the computing efficiency has been brought into full play; the network has developed from IP and NOF to high-throughput CXL, which has laid the foundation for the decoupling of computing power and storage power; and the storage power has also improved the performance and power consumption from traditional HDD to SSD. With the active promotion of Intel, Huawei, KIOXIA, Aliyun and other industry chip, storage, cloud and Internet manufacturers, Diskless architecture uses dedicated data processing chips such as DPU / IPU and dedicated memory modules such as OceanDisk to further improve the efficiency of data processing and storage and fully mobilize data center resources.

Professor Feng Dan said that the Diskless architecture of resource pooling and sharing should have the following key technologies:

1. Network storage cooperation, DPU hardware unloading acceleration, efficient cooperative data processing. Through DPU and other proprietary data processing modules, the server CPU is all for application services, the infrastructure is decoupled from application tasks, and CPU computing tasks such as VirtIO and NVMe-oF are transferred to the special accelerator of DPU to achieve flexible expansion of resources.

two。 CNC separation + IO pass-through, to achieve the separation of control flow and IO flow, IO straight-through to the storage media to further improve storage efficiency. Through the numerical control separation, the extreme IO access delay can be achieved. Through the separation of the IO read and write data stream from the IO control processing, the read and write IO is directly connected to the storage medium, which reduces the IO processing path, reduces the processing complexity, and greatly improves the storage domain access performance. Computing domain can use DPU bypass CPU to reduce data copy overhead and improve IO access performance.

3. Disk control coordination, disk chip and controller chip cooperate to improve integration and achieve the best cost and energy consumption. The key of disk control collaborative technology is to achieve a high-performance, low-cost market, and accelerate the algorithms in the disk control system through hardware, such as using FPGA to achieve FTL / DIF / online compression / encryption / EC / GC hardening to improve performance; all kinds of drivers and cache layers in the disk and frame are integrated to realize the integration of disk and frame, and build a large disk storage with extreme cost and energy consumption.

4. Operator unloading, EC, data reduction and other algorithms are offloaded to the storage side by hardware to achieve near data processing. Through the unloading of computing semantics, the storage goes straight out of the semantic interface, the host side software stack is thinned, the number of IO interactions is reduced by 80%, and the CPU and network bandwidth of the whole system are saved. At the same time, in the big data / data warehouse scenario, operations such as engine push-down interface, push-down predicate and projection can be operated to the storage resource pool to reduce the movement of invalid data.

Professor Feng Dan said: "from the perspective of technological trends, the development of high-performance heterogeneous computing, high-speed networks, new disk frames with large storage power, operator offloading and other technologies have brought about changes in the data center architecture, accelerating the CPU-centric coupling architecture to a data-centric Diskless architecture with complete storage separation and resource pooling, which further simplifies the construction of data center infrastructure. To be able to achieve intensive and efficient development of survival and computing resources has become an important technological trend in the development of large-scale data centers. "

The OceanDisk smart disk frame product released by Huawei last year is a standard new type of disk frame product with large survival capacity. OceanDisk connects the Diskless server through NoF + high-speed network to achieve independent flexible expansion of computing and storage, helping customers improve resource utilization, reducing cabinet space and device power consumption by 40%, and combining software and hardware technologies such as built-in FlashLink ®intelligent disk frame coordination algorithm, resulting in high performance with up to 2.5 million IOPS,50GB / s bandwidth. Relying on Huawei's years of storage research and development strength, OceanDisk smart disk frame provides hard disk sub-health management, intelligent slow disk optimization and other technologies to maintain the stable performance of 100,000-level hard drives in large-scale data centers and greatly reduce the difficulty of operation and maintenance. As the industry's first new disk frame for Diskless architecture, OceanDisk will redefine the next generation cloud and Internet data center storage architecture with high performance, high reliability and green intensity.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.