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

A few days ago, at the Fifth OCP China Day 2023 (Open Computing China Technology Summit), Chaochao Information officially launched the fusion architecture 3.0 prototype system, which completely decoupled and pooled core IT resources such as computing resources, storage resources, memory resources and heterogeneous acceleration resources with groundbreaking system architecture design. Support for asynchronous upgrade of pooled resources, support for fine-grained multi-host sharing of high concurrent storage, sub-microsecond remote memory sharing access and other features, through software definition to achieve "a set of systems, N applications", effectively alleviate the current data center "memory wall", "I / O wall", "power consumption wall" and other bottlenecks.

The release of the converged architecture 3.0 prototype system is expected to develop a new hardware infrastructure that is fully decoupled, fully pooled, highly scalable, easy to deploy and easy to manage, achieve a high degree of coordination between software and hardware, and accelerate the release of digital productivity in data centers. promote the development of the digital economy and deep integration with the real economy.

In the era of intelligent computing, there is an urgent need for a breakthrough in computing architecture.

At present, the transformation of digitalization and intelligence has become the rigid demand of enterprise development, scientific research innovation and social governance, and has also given birth to the vigorous development of digital technologies such as cloud computing, big data, artificial intelligence and so on. However, more and more diversified applications have different requirements for underlying hardware resources, resulting in the independence of cloud, digital, intelligent, edge, end and other technology platforms using traditional architecture, and it is difficult to share and reuse hardware resources. it not only causes a waste of resources, but also makes the difficulty of operation and maintenance management surge.

For example, the AIGC technology represented by the large model needs to be based on massive data sets, and the distributed training of the AI large model with hundreds of billions of parameters on the cluster with hundreds of AI acceleration cards requires higher heterogeneous computing power; scientific computing requires higher computing accuracy and higher demand for general computing power. On the other hand, memory computing hopes to let more application data reside in memory, so that the data and computing power are closer to each other, so as to improve the processing speed and require higher memory capacity. However, under the traditional architecture, the expansion of IT resources is completed in the form of the whole machine, even if users urgently need some specific resources, they still need to pay for the additional resources attached to the whole machine, which is bound to increase IT expenditure and cause idle waste of resources.

At the same time, with the gradual slowdown of Moore's law on the computing supply side and the end of Dennard's scaling law, the congenital deficiency of the existing computing architecture has been magnified exponentially, and the innovation of data center computing architecture is imminent.

Zhao Shuai, general manager of Chaochao Information Server product line, said: "the current phenomena such as' memory wall','I / O wall 'and' power consumption wall 'encountered in data centers do not exist in isolation. They are the embodiment of the lack of magnification of the existing computing architecture. Only through the overall innovation of computing architecture can we completely solve the challenges brought by various bottlenecks."

Converged Architecture 3.0: a New data-centric Architecture

Under this background, Tide Information launched the prototype system of Fusion Architecture 3.0, which broke the previous design concept of "taking CPU as the center". Instead, it decoupled and reconstructed the server system through system architecture innovation, and completely decoupled and pooled core IT resources such as computing resources, storage resources, memory resources, heterogeneous acceleration resources and so on. It can support the collaborative computing of a variety of general processor platforms and a variety of heterogeneous acceleration units such as GPU, FPGA, DPU and so on, and can realize resource collaborative dynamic scheduling through software definition.

This new generation of infrastructure, developed based on hardware reconfiguration technology, will achieve freer definition of resources on demand, provide better flexibility for upper-level software definition systems, and enable them to allocate and reconfigure hardware resources in a highly automated manner according to application characteristics. it is no longer limited by the non-dynamic hardware infrastructure. Let the artificial intelligence, scientific computing, cloud computing, big data and other applications in the data center run on the same architecture, achieve the integration of multi-technology platforms, and accelerate business innovation and digital transformation.

Different from the traditional CPU-centric computing architecture, the fusion architecture 3.0 prototype system takes data as the center, and realizes the sharing of memory data, unified addressing and collaborative work among various computing chips within the computing node. The distributed interconnection and exchange between nodes is formed through intelligent data processing unit and high-speed network, which realizes the computing cooperation of various acceleration chips such as CPU, GPU, FPGA, memory pooling and new storage resource pooling, which has the advantages of extremely low data access delay between nodes and supporting efficient elastic expansion. In addition, the fusion architecture system can achieve more flexible resource reconfiguration, and provide powerful computing support for artificial intelligence, big data and other application scenarios.

Memory decoupling and pooling has always been a hot and difficult point in the industry. with the emergence of serial cache consistency bus represented by CXL, it provides a low-latency access path and cache consistency guarantee between host and remote shared memory, which makes it possible for large-scale memory expansion and memory resource pooling. The fusion architecture 3.0 prototype system breaks through the key technology of memory decoupling pooling, and develops a new type of memory module and memory pooling system that applies serial cache consistency bus and its switching technology to ensure the application requirements of large capacity and high bandwidth memory in the host system.

Zhao Shuai said that the fusion architecture 3.0 prototype system pioneered the design of JBOM independent memory resource pool, innovated to achieve high-density memory expansion scheme, and the host system remote memory expansion technology led the industry. Through software definition system design and CXL high-performance switching technology, it is the first to realize memory resource pooling and fine-grained multi-host sharing.

In the aspect of system interconnection design, decoupling and pooling bring new interconnection challenges. The whole system realizes the overall operation of the decoupling unit through the design of power supply control, reset, clock locking and other cooperative work. In addition, with the continuous rise of the data rate and the more complex system links, the interconnection extension of the decoupled pooled system interconnection is close to the limit. The system carries out high-precision fitting simulation research for the high-speed interconnection of complex links. Accurately analyze the diversified topology of system interconnection links and the limit of transmission rate.

In addition, the Fusion Architecture 3.0 prototype system develops a software definition management system to achieve advanced functions such as topology switching, port dynamic management, multi-host resource sharing and resource dynamic partition, and resource management software. to achieve device utilization monitoring, device allocation configuration and management, I / O throughput monitoring and link health diagnosis to ensure the dynamic deployment and efficient management of host system hardware resources.

Zhao Shuai said: "the efficiency of the converged architecture 3.0 prototype system can be improved by one or two orders of magnitude compared with the previous generation software virtualization system, the scalability is increased by 2 to 4 times, and the system latency is reduced by 90%. The score PUE is less than 1.1. With the continuous development of digital economy and artificial intelligence, the business of enterprises is becoming more and more dependent on data and its value, and computing technology also needs to evolve constantly. The release of the integrated architecture 3.0 prototype system will help enterprises to improve the efficiency of data management and maximize the value of data. "

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.