On May 30, 2024, the CCF Information Storage Technology Professional Committee (CCF-TCIST Special Committee) was successfully held in Inspur Information, which was hosted by the China Computer Federation and undertaken by the CCF Information Storage Technology Professional Committee and Inspur Information.
This event invited Professor Wang Yijie, Deputy Director of CCF Information Storage Technology Committee, Winner of the National Natural Science Foundation of China Outstanding Youth Fund, Deputy Director of CCF Information Storage Technology Professional Committee, Associate Professor Zhang Guangyan of Tsinghua University, Professor Liu Haikun and Associate Professor Zhou You of Huazhong University of Science and Technology, Associate Professor Lu Youyou of Tsinghua University, Professor Shi Liang of East China Normal University, Sun Bin, Chief Architect of Inspur Storage, Professor Wu Chentao of Shanghai Jiao Tong University, Assistant Professor Zhang Jie of Peking University, Professor Zhang Qiming of University of Shanghai for Science and Technology, Professor Zhang Xiao, a member of the Standing Committee of CCF Information Storage Technology Committee, and experts and scholars in related fields participated. The whole event was divided into "opening speech", "invited report" &"special thinking" in the morning and afternoon, and "enterprise visit" and other exchange and discussion activities, and the participants had a warm technical discussion and exchange.
Professor Mao Bo of Xiamen University, head of the CCF Information Storage Technology Professional Committee, presided over the morning seminar, and he expressed his gratitude to the Inspur Information and Storage Alliance for its strong support for this meeting, and also expressed his gratitude to all teachers and colleagues for participating in this meeting. Mr. Li Hui, general manager of Inspur storage product line, delivered an opening speech, extended a warm welcome to all teachers, and expressed his gratitude to the CCF Information Storage Technology Professional Committee for entering the first phase of the enterprise and placing it in Inspur Information.
Mr. Li Hui said that facing the storage needs of the national strategic layout and large model application, advocating the in-depth docking and exchange between the industry and the academic community, especially in the current situation at home and abroad in terms of key technology and talent demand, academia and industry have a strong desire for cooperation.
The "Invited Presentation" of the morning event was first shared by Associate Professor Zhang Guangyan of Tsinghua University on "Coding Storage System: Architecture, Theory and Method", which mainly proposed a global resource sharing architecture for the current problems of coding storage, and introduced the online recovery of storage arrays, stable and low-latency SSD arrays, and erasure code storage for public clouds.
Assistant Professor Zhang Jie from Peking University shared his views on "Research Progress in Near-Data Computing". This paper mainly puts forward the optimization idea for the huge cost of data copy under the traditional CPU/GPU architecture, and introduces the work of delegating GNN applications to SSD and the research on using DPU+CSD to accelerate aggregation in federated learning (FL). Finally, how to realize the versatility of near-data computing is introduced, and the design of the user-mode all-flash array engine and the combination of software and hardware of page cache are introduced.
Professor Shi Liang from East China Normal University shared the "Research on SSD Failure Prediction and Deployment Optimization Technology for Large-scale Storage Scenarios", which mainly optimized the current problems of poor fault prediction performance of large-scale flash storage systems, fault recovery mechanism occupying computing resources and slow speed, and introduced differentiated machine learning SSD failure prediction research and RAID rapid recovery technology based on predictive backup.
Professor Zhang Qiming from the University of Shanghai for Science and Technology shared the "Progress of Super-resolution Optical Information Storage Technology", mainly proposing a writing technology that breaks through the diffraction limit for massive optical storage limited by optical diffraction, and introduced an efficient reading technology for image detection using optical neural networks.
Sun Bin, Chief Architect of Inspur Storage, shared the "New Storage System for Future Data Centers", mainly aiming at the current data storage needs and technical challenges, and proposed that the new storage system for future data centers needs to be gradually platformized, relying on a set of architectures, breaking through key technologies, and reconstructing multi-form storage systems. It also introduces the industrial practice and exploration of Inspur Storage, including the construction of a distributed converged storage platform for data centers and the layout and exploration of new technologies of Inspur Storage.
In the morning session of "Data-centric Storage Technology Collaborative Innovation", Professor Mao Bo of Xiamen University first threw out questions from the perspectives of architecture evolution, storage independence and data transactions, and discussed how to break through the traditional computing-centric design of Von Neumann architecture from the perspectives of storage-computing integration and storage-computing decoupling. In terms of storage independence, we have conducted extensive and in-depth thinking and discussion on whether it is necessary to establish an independent data storage center, so as to reflect the value of data, provide data services, and make a good data storage foundation platform and foundation for future data transactions. There is a consensus on the need to build an independent data storage center, but there are still many challenging questions about how to build it, who will build it, and how to develop relevant regulations and systems.
The "Invited Report" of the afternoon event was first shared by Professor Wang Yijie, Deputy Director of the CCF Information Storage Technology Committee, on "Cloud Storage in the Era of Computing Power Network", which proposed an inter-cloud computing architecture model that supports multi-party peer-to-peer collaboration under the background of continuous national project support for inter-cloud computing, breaking through the constraints of multi-stakeholder peer-to-peer collaboration followed by the centralized control architecture followed by traditional computing systems. Four works related to inter-cloud storage services are introduced: low-cloud repair traffic erasure code construction based on active incremental support vector machine, inter-cloud erasure code writing based on generation matrix transformation, elastically scalable unified redundant storage, and time-sensitive explainable anomaly detection.
Professor Liu Haikun from Huazhong University of Science and Technology shared the "Large Memory System Based on High-speed Interconnect Bus", introduced the stand-alone memory expansion and distributed memory expansion based on the new high-speed interconnection bus CXL (Compute Express Link), put forward the key challenges of building a large memory pool based on CXL, and introduced the work of a large memory system that integrates CXL/RDMA. Finally, two future development trends of CXL are expounded: 1. CXL/UB bus will subvert the future data center architecture (storage-computing convergence → storage-computing separation); 2. Enable data-centric peer-to-peer computing between CPUs, GPUs, and FPGAs.
Professor Wu Chentao of Shanghai Jiao Tong University shared the "Research on Distributed Machine Learning High Reliability Data Recovery Mechanism", mainly focusing on the problems existing in the existing recovery methods in the context of the current large model, and introduced the research on the acceleration of hard disk failure recovery in the process of distributed learning and the hybrid global graph recovery method of fault and busy nodes.
Associate Professor Lu Youyou from Tsinghua University shared "AI Data Infrastructure: Data Management for Cross-Layer Storage", mainly based on the important role of storage systems in the context of current large models, and introduced the two works of Fleche to achieve multi-level memory management and SuperFS to achieve high-performance file systems. Finally, new storage methods are looked forward to in the future, including vector storage and parameter storage.
Professor Zhang Xiao from the CCF Information Storage Technology Professional Committee shared the "Research on Storage System Optimization in the Era of Large Models", mainly introducing the development history of large models, the needs and challenges of large model training and pushing, the current progress of large model storage technology optimization and the future challenges of large models.
In the afternoon of the special thinking "storage technology and industrial development in the era of large models", Dr. Sun Lingchuan of Inspur Storage started from the current development of large models and the business background of enterprises, first from the current storage architecture, whether various storage products can support nearly 100 exabytes of large-scale infrastructure platforms for extensive and in-depth thinking and discussion, and then everyone further discussed how to evolve domestic storage technology, and agreed that it is necessary to accelerate research and development at the hardware level to meet the needs of all walks of life in the context of large models. It also expands the discussion on TLC/QLC/PLC, PCIe Gen5/6 technology and future storage architecture. Finally, we think and discuss which storage requirements are urgent to solve at this stage in the era of large models, and there are still many challenging problems in the development of storage technology and industry in the era of large models, which need to be solved together.
At the end of the meeting, the participants visited the Inspur Intelligent Computing Innovation Center. So far, the CCF Information Storage Technology Professional Committee has entered the first phase of the company's @Inspur Information activity has been a complete success.
Company Profile:
Inspur Information is the world's leading provider of IT infrastructure products, solutions and services, with 8 R&D centers, 14 production bases and 50 business branches, adhering to the concept of "computing power is productivity, intelligent computing power is innovation", and is committed to promoting the innovation and application of intelligent computing technology and accelerating the integration of digital reality. Inspur Information adheres to the innovation-driven development strategy, provides customers with more advanced cloud computing, big data, artificial intelligence, edge computing and other innovative products and solutions, and actively participates in open computing technology innovation to accelerate the open integration process of the global computing ecosystem; Implement the concept of green and sustainable development, promote the upgrading of data center construction and operation mode, improve the energy utilization level of data centers, and reduce carbon emissions.
Inspur Information Storage is committed to becoming a leading provider of storage products in China, mainly for the private cloud and hybrid cloud application scenarios of cloud data centers, developing and selling storage products, and providing professional and leading storage products, solutions and services for governments, operators, finance, large enterprises, energy and transportation, education, medical and other industries.