A Traffic Platform

Till the end of 2018, the number of permanent residents in Shanghai totaled 24.2378 million. Plus the non-native vehicles that have been running in Shanghai for a long time, the number of passenger cars totaled 5.11 million and the number of electric bikes amounted to 9 million. In contrast, the density of the road network in Shanghai is 2.93km/km2 and the average road area per person is only 12 m2. This problem is very difficult for the traffic platform to handle.

Although many cities face the challenge of large population, plenty of cars, and a small number of roads, the same problem is much more severe in Shanghai. Plus the multitudes of urban constructions, it is very challenging to solve the problem in a short period of time. Traffic jams occur very often. To ease traffic jams, the traffic platform has implemented many practices. Among these practices, Intelligent Traffic Light System is the most successful one. In addition to providing stop or pass-by signals, Intelligent Traffic Light System perceives and collects transportation data from multiple channels. The combination of movement control via signals and transportation scheduling is the best solution to the preceding problem faced by Shanghai.  

Deeper down into the Intelligent Traffic Light System, the key to the success of the system lies in the real-time massive data processing and analysis capabilities, which are expected to be provided by a cloud computing platform. The platform must be capable of collecting, processing, and analyzing massive non-structured data such as videos and images in real time. This requires that the platform has strong resource scheduling capabilities and computing powers that can be allocated to multiple real-time analysis algorithms and applications. In addition, considering the traffic light business characteristics, the underlying cloud platform must be stable and reliable enough to ensure the high availability of applications.

To satisfy the preceding business needs, ZStack Cloud offers a private cloud solution that supports CPU general computing and GPU heterogeneous computing:

• GPU-powered VM instances are provided for deep learning and AI reasoning scenarios such as video and image analysis. Each physical server is equipped with several NVIDIA Tesla T4 cards. Based on the number of VM GPUs required by reasoning algorithms and application systems, each host can provide several high-performance GPU-powered VM instances, which greatly improves video and image analysis efficiencies and allows flexible and auto-scalable GPU resource allocation.

• General-purpose x86 servers and all-flash SSD disks construct a distributed storage pool, which provides high-performance volumes for VM instances and eliminates IO bottlenecks.

• General-purpose VM instances are provided for CPU-intensive applications. This type of VM instances can be deployed on the same host as that of GPU-powered VM instances. The cloud platform can dynamically schedule resources based on the workloads of the host and improve host resource utilization without compromising the high performance of VM instances. 

• ZStack Cloud is super-lightweight and incurs low resource overheads. The performance of VM instances after virtualization is compromised by less than 5%. This provides strong and stable infrastructure support for high-performance computing scenarios.

• The cloud platform allows easy configurations of GPU virtualization so that GPU resources can be flexibly allocated and scaled. This feature makes full use of GPU computing powers and can support deep learning and AI reasoning applications.

• The distributed storage constructed by general-purpose x86 servers and all-flash SSD disks provides low-latency and ultra-high I/O storage space that can support real-time analysis and transactional workloads. It saves the high cost of purchasing all-flash storage arrays and avoids vendor lock-in. It allows online scale-in or-out and scale-up or-down and offers near-linear scaling of performance and capacity while scaling storage clusters.