Interpretation of the revolutionary container cluster CCE Turbo: all-round acceleration of computing, network, and scheduling

Abstract: CCE Turbo is a revolutionary container cluster launched by Huawei Cloud.

On May 31st, at the Huawei Cloud Techwave Cloud Infrastructure Technology Special Day, Huawei Cloud Container Batch Computing Chief Architect Motors gave an in-depth interpretation of the technical insider of CCE Turbo. CCE Turbo is a revolutionary container cluster launched by Huawei Cloud. , Officially released at the Huawei Developer Conference (Cloud) 2021, accelerates enterprise application innovation through the all-round acceleration of computing, network, and scheduling.

Software and hardware collaboration to accelerate computing

In order to solve the problem that the performance of cluster servers cannot be fully utilized, HUAWEI CLOUD, based on the software and hardware collaboration capabilities of the DynaSky architecture, introduced the industry's exclusive container offloading technology and applied it to CCE Turbo, allowing 100% of cluster resources to be used for business processing. At the same time, by slimming the engine, optimizing internal execution logic and rewriting core modules, such as rewriting shimv2 and agent based on the Rust language, reducing the number of processes, reducing memory consumption through code optimization, and improving container startup performance and Cgroup management capabilities, At the same time, combined with DynaSky's high-performance processing capabilities, the overall performance of the cluster is increased by 40%, and resource usage costs are reduced by 30%.

Trunkport, to accelerate the network

Network connection speed and forwarding efficiency have always been challenges faced by businesses when dealing with large traffic. Based on the concept of cloud native 2.0 "IN Cloud", HUAWEI CLOUD has created a new cloud native network, using Trunkport technology to comprehensively accelerate the network.

Trunkport technology enables containers to directly connect to the VPC network, transforming the original two-layer model of "container network + virtual machine network" into one layer, shortening the network resource connection time by half, and effectively supporting the business expansion of 1000 container instances within 30 seconds, which is easy to deal with Traffic surges also reduce the network communication delay by 40%, making application access smoother;

CCE Turbo is also the first in the industry to configure independent security groups and QoS for containers. Compared with the current solutions of other vendors' containers sharing security groups with cluster nodes, it not only enhances container communication security, but also improves the forwarding efficiency of large traffic. In addition, CCE Turbo also extends Kubernetes objects based on the CRD mechanism to implement various complex security isolation requirements and further simplify the configuration of container security groups.

Volcano, scheduling acceleration

After an enterprise's core business is fully cloud-native, how to improve scheduling efficiency and increase cluster utilization is a problem that plagues many enterprises. CCE Turbo implements three core scheduling capabilities based on Volcano:

• Online and Offline Hybrid Scheduling: CCE Turbo mixes enterprise offline and online services in the same cluster. Compared with the previous deployment of different services into clusters, it greatly reduces the workload of operation and maintenance. At the same time, it depends on the difference between online and offline services. Flexible scheduling is required. For example, when online business visits are low, CCE Turbo can use idle resources to run offline computing services (such as offline analysis, model training, etc.), and when the business peak comes, it will automatically release offline business occupancy Resources to ensure the online business’s demands for resources;
• Application-aware intelligent scheduling: In order to further improve the utilization of clusters after hybrid deployment, the application models (such as web applications, Tensorflow PS and workers, Spark Driver and executor, etc.) are used to meet the resource demands of different application models. , Application load conditions, reduce the idle proportion of cluster resources through mechanisms such as resource on-demand preemption and time-sharing multiplexing; and through sensing the topology of tasks, each task is scheduled to the best node to reduce network bottlenecks and data spans. The time loss caused by node transmission, etc., can in turn increase the cluster utilization by 2 times;
• Large-scale Distributed Scheduling: In order to ensure the problem of concurrent scheduling of massive tasks after the mixed deployment of services, CCE Turbo has launched a task scheduler with a distributed architecture to improve throughput and pruning the scheduling algorithm to reduce addressing depth and The breadth, combined with the scheduling decision multiplexing mechanism, can shorten the scheduling addressing time by more than 10 times, and achieve large-scale concurrent scheduling of 10,000 containers per second.

CCE Turbo speeds up VIPKID's audio and video services

As an early user of CCE Turbo, VIPKID has fully enjoyed the business value brought by the three major accelerations of CCE Turbo. On this special day, VIPKID senior expert on back-end research and development Ci Yiheng said in the sharing, “After using CCE Turbo, VIPKID audio and video services have been optimized in various aspects in various aspects. The performance of clusters of the same specification has been improved by 40%, and the service interaction delay has been reduced by 40%, which further improves the user experience and saves about 43% of the entire business cost. , It has well controlled the growth rate of IT costs during the period of rapid business growth."

As the first batch of vendors to devote themselves to cloud-native technology, Huawei Cloud is the only startup member of the Cloud Native Computing Foundation (CNCF) in Asia, ranking first in Asia for community code contribution and maintainer seats, and contributing the first cloud-native intelligence The edge project KubeEdge and the batch computing project Volcano. At the Huawei Developer Conference 2021 (Cloud), Huawei Cloud also opened up the cloud-native multi-cloud container orchestration project Karmada with a number of companies, continuing to lead the development direction of cloud-native technology; in terms of product innovation, HUAWEI CLOUD has launched a series of cloud-native products and solutions for the first time in the industry since 2016. In Forrester's product capability assessment, it has won full marks for two consecutive years, and its container software market ranking has ranked first in China; in terms of prosperous industries, HUAWEI CLOUD has not only released a cloud native 2.0 white paper in cooperation with China Academy of Information and Communications Technology, fully interpreting the core concept of cloud native 2.0, but also established a global cloud-native communication platform with CNCF and China Academy of Information and Communications Technology-Chuangyuan Conference. Industry cloud-native elites work together to explore cutting-edge cloud-native technologies, share industry landing practices, and use cloud-native technologies to fully empower enterprises' digital transformation.

Click to follow and learn about Huawei Cloud's fresh technology for the first time~