The east wind of large models is sweeping the world, and some industries have begun to reconstruct based on large models. In the future world of digital intelligence, AI will undoubtedly play an important role. By combining with different basic technologies and products, AI will promote continuous technological innovation in various fields, and the operating system as the core basic software is no exception. "Rise in the digital era and lead the future of digital intelligence." At the Operating System Conference & openEuler Summit 2023, we found that openEuler has made plans in advance for the future of digital intelligence.
OpenEuler has always focused on investing in root technologies to build a solid software foundation for thousands of industries. Currently, the cumulative installed base of openEuler has exceeded 6.1 million units, making it the first open source operating system to change the landscape of software operating systems. According to IDC predictions, openEuler will rank first in China's server operating system market share in 2023, reaching 36.8%. Facing the future of digital intelligence, openEuler is fully integrated with AI to create a full-scenario operating system for digital infrastructure, providing technical convenience to more users.
01Facing
the new era of intelligence
openEuler makes AI more efficient
In the past four years, the openEuler community has grown rapidly. According to Xiong Wei, vice chairman of TOC of the Open Atomic Open Source Foundation and executive director of the openEuler Committee, when the openEuler community was established, there were only two to three hundred active developers every day, but now it has gathered more than 16,800 open source contributors, with more than 4,259 daily active users. people. Developers provide a lot of help with system features, innovations, and bug fixes. In addition, the openEuler community has accumulated more than 1,300+ enterprises and partners.
Xiong Wei, Vice Chairman of TOC of Open Atomic Open Source Foundation and Executive Director of openEuler Committee
When many users and partners join the openEuler community, where will openEuler go in the future? In the past, openEuler had two technical pillars, full scenarios and diverse computing power. With the continuous development of new technologies such as large models, openEuler adds a third technical pillar - intelligence. Digital openEuler will gradually evolve into smart openEuler and grow into a next-generation OS for the era of diverse computing power and intelligence.
The boundaries of intelligence continue to expand. As a bridge between hardware and applications, operating systems embrace the future of digital intelligence and combine with AI to be an inevitable trend.
Hu Xinwei, chairman of the openEuler Technical Committee, said that under the current hardware trend, more and more intelligent and diverse computing power needs to be efficiently coordinated, and more and more ubiquitous intelligent applications need to be quickly intervened. In order to realize this vision, openEuler hopes to achieve the effects of zero threshold, zero loss and zero intervention. For ordinary users to deploy and install smart applications, it helps users achieve the "0" threshold; for smart applications running on diverse computing power, the operating system allows "0" loss of smart computing power; during the daily operation and maintenance of smart applications, the system The administrator has "0" involvement in the daily work of the system.
Therefore, openEuler implements full-stack AI enablement through "openEuler for AI" and "AI for openEuler" to create a containerized, ready-to-use solution.
"openEuler for AI" is an expansion of ecological boundaries. In the future, openEuler will comprehensively enhance its compatibility with AI, support more mainstream intelligent applications and large language models, and achieve intelligent capability upgrades. It is not easy to build an AI ecosystem. At present, we can see that some AI products, technology ecosystems, infrastructure, etc. are fighting independently. However, as the market gradually matures, AI full-stack enablement is an inevitable result of future development.
Full-stack enablement will allow openEuler users to use the containerization solution provided by openEuler out of the box when using smart applications in the future. At the same time, it can also automatically adapt to the hardware environment to achieve optimal performance. openEuler's full-stack enablement also includes two levels of meaning. The first is the optimization of support for various applications, models, tools, and frameworks, such as vector databases such as AquilaDB, popular large models in the industry such as Llama and ChatGLM, tool chains such as cuda, rocm, and openvino, and AI frameworks such as PyTorch and TensorFlow. In addition, due to the complexity of AI hardware, openEuler also provides extensive support for AI hardware compatibility. A series of upgrades will greatly improve users' AI development and usage efficiency.
The second is support for developers. openEuler continues to optimize AI usability and adaptability, and has upgraded containerized packaging of training and inference push environments, and one-click image pulling, providing developers with instant solutions. Use the ability to achieve the "0" threshold for AI environment deployment.
Currently, we have entered an era of diverse computing power, and the joint development of software and hardware gives room for redesign of operating system scheduling. Through the SMT architecture of the CPU, the operating system is given the opportunity to pre-deploy tasks, and the GPU can also avoid problems such as pauses through hardware warp switching. Therefore, how to bring together independent heterogeneous devices, uniformly allocate resources, and solve the waste of computing power and the complexity of heterogeneous memory programming is an important issue.
In the traditional sense, different heterogeneous computing powers are completely separated from each other in terms of memory management, and it is difficult to share them with each other. Another important value of "openEuler for AI" is heterogeneous fusion. Its core idea is to bring together independent heterogeneous devices and uniformly allocate resources to solve the two problems of serious waste of computing power and complex heterogeneous memory programming. Through the heterogeneous kernel management introduced in openEuler, unified addressing can be achieved between the CPU and NPU by sharing page tables, allowing both parties to use memory with each other to achieve "transparent" memory expansion and super-resolution, which can improve the throughput of inference scenarios. The volume is increased by 50%, and the demand and cost of memory management are greatly simplified. Multiple sets of memory interfaces in the past are simplified into one set, and heterogeneous driver code can be reduced from 10,000 lines to 100 lines.
02
Collaboration between large models and OS has become a trend
AI empowers openEuler to be smarter
Judging from this year's technological development, large models allow more people to see new possibilities in operating systems. For example, in 2023, Microsoft announced that it would embed the GPT-4 large model into Windows to upgrade the operating system from graphical interaction to natural language interaction, which is expected to change the system interaction method that has been used for nearly thirty years. Windows Copilot will land on the taskbar in Windows 11 to help users find and change settings more easily, avoiding the constraints of cumbersome options and operating procedures.
On the cloud, traditional complex operation and maintenance work is also expected to be changed by large models. The analysis of cloud fault root causes and the formulation of mitigation measures based on large models have performed well, and have been recognized by more than 70% of operation and maintenance personnel; in terms of task collaboration, the software open API is provided to AI to control the completion of complex tasks, achieving a goal beyond that of assistants and The excellent performance of the tool can increase the overall system value.
As artificial intelligence innovation technologies represented by large models and large computing power continue to develop, AI continues to accelerate its entry into thousands of industries, and operating systems also need to continue to evolve towards AI. In this context, collaborative optimization of large models and OS has become a trend, the operating system will undergo major changes, and digital openEuler will evolve into smart openEuler.
Smart openEuler brings changes in system interaction. For example, developers may use various programming languages in their daily work, but no one of them is as natural and convenient as our native language. So, is it possible to implement natural language and operating system interaction through tools, thereby reducing the intensity of writing scripts, configuring parameters, or debugging? openEuler uses the ChatGLM basic model and trains EulerCopilot based on a large amount of openEuler code and data. It initially implements functions such as code-assisted generation, intelligent problem analysis, and system-assisted operation and maintenance, making openEuler more intelligent. EulerCopilot will bring about tremendous changes in the interaction between humans and machines, and this is also an important change in "AI for openEuler".
If "openEuler for AI" has brought new vitality, then "AI for openEuler" has injected new vitality. Based on the technology accumulated by the openEuler community, EulerCopilot provides users with more convenient artificial intelligence capabilities and rich job portals. Users can interact through public accounts, WEB interfaces, SHELL, IDE, etc. EulerCopilot integrates massive knowledge in the OS field and can answer various professional questions from developers, automatically complete unfinished code segments, and even complete requests such as "system performance diagnosis" and automatically generate diagnostic reports and tuning opinions.
03A
vast space worth imagining
In the future of digital intelligence, openEuler will serve as a carrier to continuously input AI capabilities into multiple application scenarios such as servers, cloud, edge computing, and embedded systems, promoting the digital upgrading of thousands of industries. Of course, the future of digital intelligence requires different infrastructure. Wu Fengguang, a member of the openEuler open source community technical committee, said: "In the era of AI, the community is moving towards intelligent collaboration, and we are exploring the use of AI to empower community collaboration." OpenEuler built Infrastructure 2.0 has been implemented to support global development, full-scenario construction, and full-link collaboration to help more people acquire AI capabilities.
In addition, openEuler’s important development direction is overseas expansion and globalization. Wu Fengguang said that when openEuler was founded, it decided to fully globalize. After focusing on independent innovation in China, it would go overseas to gather global developers and come to the openEuler community for native development. Therefore, the Open Atomic Foundation has formed connections with many foundation organizations and upstream community projects to conduct version certifications and complete project work for each other. Only in this way can we continue to gather global open source forces and contribute Chinese wisdom to the world's open source.
Since the OS source code adopts a new architecture, it needs to be converted from SPEC to YAML, which will help build full-scenario capabilities. The advantage of YAML is its versatility and low threshold. This means that its developer base is better, almost everyone can use it, and it can quickly gain a mass base. The universal configuration language can also be used to create out-of-the-box customization capabilities, supporting any YAML field customization and adapting to upstream software in many formats. The OS supports hierarchical customization and can build software for multiple scenarios through EulerMaker and EulerTest.
In order to solve the problem of difficulties in promoting upstream software to users and poor transmission of user voices to upstream, openEuler has opened up the link from upstream to users through the application software platform in terms of infrastructure, achieving full-link collaboration in the true sense. This is obviously not the ultimate form of openEuler. It is expected that in May 2024, openEuler will release the 24.03LTS version of the new kernel to further improve its full-scenario capabilities. Using Linux6.6 as the kernel to achieve ecological unification. Create a new EEVDF scheduler and folio memory management mechanism to further improve scheduling and memory usage efficiency. There are comprehensive improvements in IO management, new network standard support and support for CXL.
For different usage scenarios, the 24.03LTS version will also provide different optimization upgrades. In cloud computing scenarios, low-load computing power is intelligently aggregated through CPU aggregation scheduling to achieve load and computing power synergy; a cloud-native minimum set release supporting openEuler is provided to support one-click deployment; in embedded scenarios, server-oriented releases will be released BMC's original open source project MetaBMC provides a standardized development board "openEuler Pie" with native built-in openEuler and an industrial robot framework "openEuler arm".
Of course, in terms of AI, openEuler will continue to evolve. In addition to EulerCopilot, it will also implement "self-optimization" intelligent tuning of the operating system, which will be of great help to developers. By sensing business characteristics while the application is running, it can dynamically adjust optimization strategies such as scheduling priorities and configuration parameters. Achieve scene-based performance improvement of more than 15%.
As far as I can see, I can see friends coming from afar. The development of AI has never been driven by one company or one community. openEuler will promote industry development by establishing learning groups, formulating new specifications and other measures, and is committed to becoming a leader in the AI field. When AI and OS achieve each other, future developers will be expected to create greater value and complete more technological innovations.
Recommended reading
↑Limited time 50% discount↑