Construct the Second Technology Curve

01 Alibaba Cloud core products are fully serverless

New changes are taking place in the software R&D paradigm. Serverless is one of the most important trends. Alibaba Cloud will firmly promote the comprehensive serverless core products to help customers better realize agile innovation.

"We hope that users can do less and gain more. Through Serverless, using the cloud is as simple as using electricity." Zhang Jianfeng said that Serverless will make cloud computing from a resource to a capability. In the future, the cloud will be fully Serverless, closer to the "grid" mode, and pay according to the number of calls.

In Supermarket Lianhua, a retailer, the membership system and transaction system are all processed on Alibaba Cloud's function computing. It says goodbye to relying on extended machines to support massive business. The preparation time for promotion is shortened from week level to hour level, and the cost is reduced by 40%;

In Pumpkin Movie: When there are a large number of users, Pumpkin Movie can switch to the Alibaba Cloud Serverless application engine SAE in a short time, automatically expand capacity in seconds at peak hours, and automatically shrink capacity on demand at peak and valley hours. The total cost will be reduced by 40% to deal with sudden traffic.

At present, Alibaba Cloud has more than 20 Serverless products, including functional computing FC, Serverless application engine SAE, Serverless container service ASK, PolarDB database, AnalyticDB data warehouse, etc.

Among them, the number of daily calls of function calculation exceeded 20 billion, effectively supporting the double 11 million QPS peak over the years, and the average annual growth rate of business exceeded 300%, with the overall scale ranking the first in China.

02 Yitian+Feitian+CIPU combination

Increase performance by more than 20%

In order to provide better products and services, Alibaba Cloud has further optimized the efficiency of data processing and transmission, built a new data-centric cloud computing architecture, and provided accelerated support for near-network computing and near-storage computing.

With the support of Apsara+CIPU, the cache QPS performance of Redis/Air database can be improved by 80%, the master-slave replication latency of the database can be reduced by 50%, the global consistent read performance can be improved by 5~10 times, the big data Spark performance can be improved by 30%, the AI in-depth learning graphics classification training performance can be improved by 30%, the OLAP real-time data warehouse query performance can be improved by more than 2 times, the cold and hot storage types can be layered, and a single data can support multiple computing engines

Under this system, the combined performance of Yitian+Feitian+CIPU is also very bright. It is generally superior to similar products in the industry in the scenarios of database, big data and AI, high-performance computing, video codec, etc., and the performance can be improved by more than 20%.

The AliCloud elastic computing instance G8M, which is equipped with the Reliance chip, will be officially launched on November 15. At the same time, CIPU will also be introduced into the form of private cloud for the first time, providing government and enterprise customers with the ability of a cloud multi-core and cloud multi-region architecture with separate storage and calculation.

03 Alibaba Cloud database moves towards "four modernizations"

All in Serverless

At the current Yunqi Conference, Alibaba Cloud announced that core database products such as PolarDB, AnalyticDB (hereinafter referred to as ADB), RDS, and Lindorm have been deeply integrated and innovated with the self-developed CPU, CIPU, and Apsaras operating system, and are fully serverless to better provide one-stop data management and services externally.

• RDS database is based on the software and hardware co-optimization of self-developed CPU, and the overall cost performance is improved by 25%;

• PolarDB achieves zero loss of performance and data storage costs can be reduced by up to 60% by integrating the self-developed Smart-SSD smart solid-state drive technology.

PolarDB is the first cloud database in the industry to support cross-machine serverless services. It breaks through the two major technical difficulties of sensorless switching and high-performance global consistency. Compared with the cost of configuring capacity according to peak load, it can save up to 95% of the database cost.

At the same time, the latest commercialized cloud native HTAP function of PolarDB database, under the support of the IMCI technology of the inventory index, has achieved a significant lead in the performance of TPC-C and TPC-H.

The core database products of Alibaba Cloud, represented by PolarDB and ADB, have also made efforts in cutting-edge technologies such as HTAP and offline integration to further improve the usability of database products and further improve performance.

Among them, ADB provides the ability of offline and online integrated data processing and analysis. Based on a data fusion two computing models (MPP+BSP), it supports offline processing and online analysis at the same time, solves the problem of data consistency and timeliness caused by data synchronization, and reduces the resource cost by more than 30%.

Facing the future, Alibaba Cloud database will develop in the direction of "four modernizations" as a whole:

• Cloud biochemistry: resource decoupling, serverless

• Flat platform: ability to build data platform based on cloud, OpenAPI standardization

• Integration: processing and analysis integration, off-line integration, centralized distribution integration, multimode processing integration

• Intelligence: AI for DB simplifies operation and maintenance, In-DB ML mining data value

04ODPS new upgrade

Meet users' diversified data computing needs

To meet users' diversified data computing needs, Alibaba Cloud ODPS has been upgraded to an integrated open big data platform. The upgraded ODPS supports large-scale batch computing, real-time analysis and other services, provides real-time streaming computing, machine learning and other computing capabilities, and can simultaneously schedule more than 100000 servers for parallel computing.

ODPS realizes the integration of storage, scheduling and metadata management, supports the efficient processing of data in intelligent transportation, life science, risk actuary and other fields, and provides consistent data development and governance services through DataWorks.

In the TPC-H 30000GB benchmark test just released, ODPS-Hologres broke the world record, scoring more than 27.86 million points, 23% ahead of the second place. ODPS-MaxCompute has remained the first in terms of performance and cost performance for six consecutive years in the TPCx-BB 100TB standard test.

05 Panjiu ultra-high performance network, significantly reducing latency by 90%

In order to break through the limitations of computing performance and memory resources on the training efficiency of AI models, Alibaba Cloud engineers have created a set of "Panjiu Ultra High Performance Network", which uses the self-developed Solar-RDA high-speed network protocol to enable the processor to access the memory of any other server. It is very suitable for in-depth learning of the interactive form of neural networks in the model. Compared with the traditional model, it can reduce the long tail delay by more than 90%, and the minimum can be as low as 2 microseconds.

Based on this set of ultra-high performance network technology and the ability to integrate software and hardware, Alibaba Cloud launched a new generation of intelligent computing product, Lingjun, in June this year, which can minimize all non-computing costs and achieve a five-fold improvement in communication performance. The efficiency of kilo-card parallel computing is up to 90%, and the performance of large-scale AI can be improved by 2-5 times.

06 Model as a service

ModelScope of Magic Community lowers AI threshold

In order to lower the threshold of AI use and give full play to the potential of AI, Alibaba Dharma Academy and CCF Open Source Development Committee jointly launched the "magic" ModelScope of AI model community. The first batch of cooperative institutions in the community include Lanzhou Technology, Shenshi Technology, Zhipu AI, China University of Science and Technology, etc.

The magic community practices the concept of Model as a Service, and goes further on the basis of open source code, turning AI code into model and model into service. The community has achieved zero threshold AI model experience, one-line code model reasoning, and ten-line code model tuning.

Dharma Academy and others have taken the lead in contributing more than 300 proven high-quality AI models to the community, more than one third of which are Chinese models. Using these basic models, developers can quickly put them into use with only a little tuning for specific scenarios.

07 Hyundai C908 achieved a breakthrough in energy efficiency of RISC-V architecture

The newly released black iron C908 processor has achieved a breakthrough in energy efficiency of RISC-V architecture, which is more than 20% higher than that of the industry's same-energy processors.

In order to further improve AI computing power and energy efficiency, the black iron C908 upgraded the micro-architecture, adopted the RISC-V Vector 1.0 standard for the first time, and added the DOT command based on the standard; Launch INT4 data type; The AI operator and algorithm library are comprehensively optimized, and the performance of typical neural network computing is more than 50% higher than that of the previous generation of products.

As a result, the black iron C908 also surpassed the black iron C906, which had won the first prize in the international authoritative AI test MLPerf Tiny V0.7, and its performance in image classification tasks was further improved by 3.5 times. Dark iron C908 can be used in intelligent interaction, AR/VR, wireless communication and other scenarios.

08 Invisible upgrade to enhance the integrated security of cloud network:

Coverage of terminal productivity scenarios

Wuying recently launched productivity solutions in multiple scenarios, including research and development, branch offices, RPA, education, etc., which are often in the field of data sensitivity. In addition to providing better cloud flow protocols for terminal R&D manufacturers, Wuyi further integrates security capabilities.

Whether users use existing applications, desktop instance pools, or other scheduling, they can set permissions and policies that meet their own scenarios in sensitive data detection, behavior log processing, high-risk operation interception, security situation awareness, external device permissions, network and application security monitoring and other aspects.

In the past, in the enterprise office scenario, the security policies of the computing terminal and the intranet were often fragmented and could not form an overall collaboration, especially the security of enterprise cross-organization collaboration.