File Storage CPFS Best Practices
Introduction to Qingzhou Zhihang
Qingzhou Zhihang is an automatic driving general solution company with the mission of "bringing unmanned driving into reality". Relying on the dual engine strategy, on the one hand, it advocates a cost-effective front loading mass production solution, and is committed to creating a L4 level experience city+high-speed NOA solution to meet the automatic driving mass production needs of different customers at different levels. On the other hand, we will work together with the government to create a "mobile technology business card for cities" and strive to become the leader of Robobus and the popularization of Robotaxi.
Qingzhou Zhihang owns more than 100 invention patents and software copyrights, fully covering the field of unmanned driving technology research and development and commercialization, and has won dozens of awards in various top events and competitions, including the champion of CVPR 2021 Argoverse Sports Prediction Challenge and other international top events.
Business Scenario Introduction
The "Light Boat Matrix", an automatic driving R&D tool chain with simulation as the core of Light Boat Intelligent Navigation, has opened up the whole process from data processing, labeling, training, large-scale simulation to technology output, realized efficient data flow and closed-loop verification, improved data utilization and technology iteration efficiency, and can also be fully used for urban NOA capacity building, greatly accelerating the development rhythm. The ability of data closed-loop enables the real vehicle test, data transmission, problem diagnosis, model analysis and data mining of Light Boat Intelligent Navigation to achieve a day level closed-loop and iteration.
In terms of database establishment, Qingzhou Zhihang has tested 1.12 million kilometers of urban roads, with a large number of multi-sensor data and long-term accumulated driver driving behavior data. The scale of these data will continue to expand, and can be used in a completely reduced dimension, enabling urban NOA, so that driving ability can be rapidly improved. Not only that, the massive simulation test mileage can reach more than 100 times of the actual vehicle test mileage. In this process, Qingzhou Zhihang also continues to conduct scene mining and continuously strengthen the construction of the scene library. At present, it has covered more than 100000 scenes, and the cumulative simulation mileage has reached hundreds of millions of kilometers. It can build and cover more long tail issues and improve the security of the system. Through data-driven, Qingzhou Intelligent Navigation has also realized the development capability of efficient perception, prediction and planning control, which also means that Qingzhou has basically opened up the application of data-driven in the entire automatic driving research and development system, leading the innovation of research and development paradigm.
Business pain points
1. How to improve the utilization efficiency of data resources
Figure 1 The use of traditional autopilot tool chain involves NAS/HDFS/object storage and multiple data islands
The data generated by the autopilot business has changed from tens of petabytes in the test phase to ZB in the mass production phase. The efficient storage and management of massive data has become an unprecedented problem. If the traditional autopilot scheme is used, multiple sets of storage needs to be deployed for data collection, screening, tagging, training and simulation, resulting in data islands, data migration and low business efficiency.
2. How to maximize GPU utilization and improve computing efficiency?
Qingzhou is committed to achieving L4 unmanned driving, which requires more than billions of kilometers of data testing, and more than 99% of the future test mileage will be completed through simulation. The light boat matrix can build simulation scenarios based on real road test and generated data, which not only reduces the test cost to less than 1% of the pure road test, but also can generate millions of extreme scenarios (Corner Case). Extract millions of frames of effective data every day, complete training, test verification and iterative optimization, and how to transmit millions of frames of data to GPU at high speed for calculation? This poses a great challenge to the small file throughput bandwidth of storage. There is a performance bottleneck in the access of traditional file storage schemes, which leads to the problem of insufficient GPUs and waste of computing resources.
3. There are peaks and valleys in the business. How to save costs when the peak is low?
Qingzhou has realized the comprehensive containerization of business systems, and fully enjoys the elastic advantages of public cloud computing resources through container technology. Fast capacity expansion at peak time can shorten the task running time, and low peak capacity reduction can reduce the cost of computing power. The traditional storage system can't bear the requirements of rapid mounting of storage system and tens of thousands of POD parallel access due to large-scale expansion and contraction of containers.
Alibaba Cloud storage solutions
Aiming at the low efficiency of multi service data storage in traditional automatic driving scheme. Alibaba Cloud's file storage CPFS and object storage OSS data lake storage and data free flow solutions meet the requirements of data automation from massive data collection to cleaning, labeling, training to archiving, provide a unified data platform for automated driving research and development cloud, and greatly improve research and development efficiency.
● Massive small file carrying capacity: CPFS single file system can provide 4 billion file carrying capacity and million OPS capacity
● Ultra high performance: The CPFS single file system provides sub ms level read and write latency, 280W IOPS, and hundreds of thousands of metadata operations OPS
● CPFS and OSS data flow: OSS data is pulled on demand at the data block level, without pre reading or waiting for the complete OSS object to be imported. Data automatically sinks to OSS after cooling, reducing storage costs
● Large scale expansion and contraction of containers: CPFS supports K8S CSI interface, and can support tens of thousands of POD simultaneous access and large-scale elastic expansion and contraction
● Unified data base OSS: a set of system realizes data processing, labeling and persistent storage, and data is copied without time saving by 30%; Seamless connection with EMR and other computing engines and Hadoop and other open source ecosystems
Through the establishment of cooperation with Alibaba Cloud, Qingzhou Zhihang can focus more on business scenario research and development. By using Alibaba Cloud CPFS and OSS data flow solutions, Qingzhou Intelligent Airlines has built an integrated autopilot data base, and supported the rapid growth of its business by relying on the high performance and high throughput of CPFS.
Knowledge Base Team
Knowledge Base Team
Knowledge Base Team
Knowledge Base Team
Explore More Special Offers
50,000 email package starts as low as USD 1.99, 120 short messages start at only USD 1.00