Technical demonstration of RDS On Yitian ECS

01 RDS database architecture evolution

As shown in the figure above, it shows the development of Alibaba Cloud RDS. In 2009, AliSQL went to the commercial database to realize scripted automatic operation and maintenance. In 2011, Alibaba Cloud RDS was launched, successfully realizing fully automatic deployment, operation and maintenance free, and highly available disaster tolerance capability. In 2014, RDS supported OpenAPI, PostgreSQL, SQL Server and other multiple engines, and successfully entered the service stage.

In 2019, Alibaba Cloud implemented DAS self-diagnosis service, which supports automatic elastic scaling, intelligent decision-making and parameter tuning based on machine learning; At the same time, the group's e-commerce business began to be fully clouded. Since then, Alibaba Cloud RDS has officially entered the intelligent stage.

In 2022, RDS and Alibaba Cloud infrastructure will be deeply integrated, supporting the flexibility of Serverless products, highly integrated with ECS and CIPU, and successfully achieving in-depth optimization on domestic hardware, as well as autonomous controllable kernel. Alibaba Cloud RDS has started to make full use of cloud computing.

Based on basic services such as Alibaba Cloud ECS and ESSD, the RDS database has built a separate architecture of computing and storage, greatly improving product capabilities and customer business value.

Under this architecture, users can use the cloud disk snapshot function to quickly build the backup and recovery capability of the database. In addition, users can also take advantage of ECS's elastic resource capability to easily realize the rapid lifting and expansion of resources.

RDS builds a K8S cluster scheduling system based on Alibaba Cloud IaaS service, and deploys the database container on its core management and control architecture.

The platform supports four RDS engines, including MySQL, PostgreSQL, MariaDB and SQL Server. Based on this architecture, Alibaba Cloud has achieved a number of core business innovations, including Serverless, cloud native Kubernetes multi-tenant deployment, and so on.

In the process of architecture upgrading, Alibaba Cloud's technical team carried out a lot of technical innovation to solve the IO and other related problems after the local disk was migrated to the cloud disk. The team solved the IO delay problem caused by architecture evolution through the transformation of Binlog system and the effective improvement of RTO and RPO.

As one of the most basic mechanisms of MySQL database, Binlog is a deeply dependent function of all MySQL databases on the cloud. However, Binlog mechanism also has many limitations, such as large transaction processing, limitations in high-throughput scenarios, etc. Alibaba Cloud R&D team has significantly improved performance indicators such as throughput and latency through the in-depth transformation of Binlog.

In addition, Alibaba Cloud has conducted in-depth optimization on buffer pool initialization and crash recovery. Although there are a lot of optimization details and code modifications, its kernel is compatible with the kernel capabilities of the open source community.

Through the evolution of the OnECS architecture, RDS provides a lot of optimization possibilities for product capabilities and customer business value. For example, OnECS based on ESSD not only supports conventional PL1, PL2, and PL3 cloud disks, but also supports PL0 and AutoPL, helping customers achieve the goal of reducing costs and increasing efficiency. At the same time, OnECS supports the dynamic degradation of PL level, and the business is insensitive throughout the degradation process.

In addition, OnECS supports cloud disk expansion and capacity reduction capabilities, so that data storage can be dynamically adjusted with customer data and business development stage, and finally achieve the effect of cost reduction and efficiency increase. With the improvement of stability, the OnECS architecture can support the basic version of read-only instances on the basis of highly available read-only instances, and reduce the cost of read-only instances by 50%, further achieving the purpose of reducing costs and increasing efficiency.

02 RDS On Technical Evolution of Yitian ECS

Next, let's talk about the technological evolution of RDS On and Reliance ECS. Through the technical architecture evolution mentioned above, Alibaba Cloud has significantly improved the efficiency of research and development of the database RDS migration from Intel's OnECS architecture to On Reliance ECS.

As shown in the figure above, under the mixed load scenario of MySQL read and write, the performance of RDS On Reliance ECS slightly surpasses that of the seventh generation x86 machine. Under different instance specifications, the performance of RDS On Reliance ECS has been improved by 10% to 25%. In addition, compared with the seventh generation Intel model, the price of Yitian ECS has decreased by 13%, and the overall cost performance has increased by 25%.

In the process of customer business and application migration from x86 architecture database to Yitian ECS architecture database, the amount of business code transformation was zero, and the seamless migration was successfully achieved. The customer only needs to change the connection address of the database from the x86 architecture to the address of RDS On Reliance ECS.

Next, I'd like to share with you the in-depth technical optimization of RDS On Yitian ECS. As a basic PaaS service, RDS database needs to be deeply tuned on the server of ARM architecture. This is a whole stack optimization process, from the chip prefetch strategy optimization to ECS TSO optimization, as well as operating system optimization, compiler optimization, database kernel code optimization and parameter tuning. Among them, the most important part is the code optimization, parameter optimization and compilation environment optimization of the database.

On Yitian ECS supports two traditional open source engine products of RDS, MySQL and PostgreSQL. The above figure shows the performance changes of these two engines before and after optimization. After the above in-depth tuning, in the write-only business scenario, the performance of RDS is 110% of that of the seventh generation ECS server, and the read-only capacity is up to 125%.

The above figure is a screenshot of the RDS sales console on Alibaba Cloud's official website. Users can choose to use x86 architecture or ARM architecture during the purchase process. At present, the product is still in the stage of test invitation.

Yitian 710 is not only a single core performance improvement, but also a performance improvement in business scenarios such as data encryption, data decryption, data compression, and data decompression.