COCA DPU Open COCA White Paper 2023

Mondo Technology Updated on 2024-01-24

Today's sharing is:《opencoca***2023)》

Featured Report**: The world of fantasy films

Summary of the content of the research reportAt present, the new generation of information technology represented by cloud computing, artificial intelligence, and big data is booming, and the integration of traditional industries and emerging technologies is accelerating, promoting the rapid growth of the digital economy. As the core capability on which the operation of information systems in various industries depends, computing infrastructure is indispensable in economic and social operations.

As a new type of productivity, computing power mainly includes information computing power, data storage power and other elements, and provides services to the society through computing infrastructure.

cocaThe software-hardware integrated on-chip computing architecture creates a national-level independent and controllable computing infrastructure

COCA takes building inclusive high-performance computing power as its core goal and the grand vision of building an independent and controllable high-performance computing infrastructure to help build a digital China.

Following the concept of "software-defined, hardware-accelerated", COCA is mainly composed of Coca-GPU modules, Coca-DPU modules, and Coca-HPN modules. Among them, the COCA-DPU module forms "five offloading engines" around computing, storage, network, security, and control, and reconstructs the computing infrastructure of the data center based on the integration of software and hardwareThe Coca-GPU module focuses on AI abstraction, AI acceleration, and AI pooling technologies to connect the upstream and downstream of the GPU industry and jointly solve the problem of "fragmentation" of different GPU platforms

The Coca-HPN module aims at large-scale clusters that require massive GPU computing power on the one hand, and more serious network congestion problems on the other hand, improving computing efficiency and breaking through the bottleneck of computing power interconnection. Based on DPU, COCA reconstructs computing infrastructure, co-creates high-performance network technology, and jointly builds an independent DPU+GPU industry alliance through the deep integration of HPN and domestic GPU ecosystems, driving the mature development of the localized intelligent computing industry.

coca-dpuReimagine the computing architecture

A DPU is a processor that provides data center infrastructure services that offloads and accelerates basic functions such as networking, storage, security, and governance, freeing up more CPU computing power for customers to use[4]. The DPU is usually composed of a general-purpose processing unit that processes the control plane business, and a dedicated acceleration engine that ensures the processing performance of the data plane, and breaks through the bottleneck of data processing performance of general-purpose infrastructure virtualization while ensuring versatility. Extending the virtualization software framework from a single CPU platform to a dual CPU + DPU platform can greatly enhance the data processing capacity of cloud infrastructure.

The COCA-DPU module can accelerate and offload functions such as computing, storage, network, security, and control by reconstructing the data center of computing infrastructure with integrated software and hardware. The Coca-DPU module realizes the standard access to the DPU through the abstract driver adaptation layer, which can be divided into five engines: computing, storage, network, security, and control, among which the computing engine provides standardized Virtio-Net (Virtual IO Network) and VirTIO-BLK (VirTIUAL IO Block) back-end interfaces to realize the acceleration and offloading of the data plane and control plane of the virtualization IOThe storage engine implements the backend of the storage interface on the DPU, and reads and writes block storage by loading standard virtio-blk or NVMe (non-volatile memory express) drivers without the need for additional vendor-specific drivers

The network engine uses standard offloading interfaces and flow tables to offload and accelerate network trafficThe security engine secures systems and multi-tenant networks through a root of trust mechanism and standard encrypted communication protocols such as IPsec, and provides an effective offload scheme based on the DPUThe management engine shields the differences in product forms of bare metal, virtual machines, and containers, so as to achieve unified management of DPU resources and full-link management and operation and maintenance.

Report produced by China Mobile).

The following is an excerpt from the original report:

This article is for informational purposes only and does not represent any investment advice from us. The world of fantasy filmsOrganizing and sharing information is only recommended for reading, usersInformation obtainedFor personal Xi only, please refer to the original report for use. The report totals:31 pages。Due to space limitations, only some of the contents are listed.

Related Pages