Public Sectors

Scenarios

The supercomputer system is the most cutting-edge information technology complex that showcases the technological strength of a country. Supercomputing provides essential technical support for scientific research. Computers with powerful computing, communication, and data processing capabilities are used to process data and online transactions, provide information services, and perform scientific engineering computing. Supercomputing plays an important role in the scientific engineering computing tasks of multiple disciplines, including new material design, new nanostructure, global climate change research, industrial engineering design, aerospace vehicle manufacturing, and AI research.

Provinces or cities in China that intend to stand out in technology competition are endeavoring to build a national or regional HPC center for public sectors. In the short run, the HPC center provides scenario-specific HPC and AI capabilities to supercharge regional economy, industry upgrade, scientific research, and urban services. In the long run, the HPC center provides optimal services, such as HPC, AI, and big data processing, to streamline the value chain and enhance collaboration for an ecosystem of shared success. The HPC center is also a new driver for digital economy and the foundation for City, Industrial, and Scientific Intelligent Twins.

The HPC center for public sectors provides resources and tool services to support diversified businesses and customers at the application layer. The majority of them are heavy-load businesses, such as scientific research, public services, and security-sensitive industries. Resources need to be flexibly scheduled based on loads to reduce power consumption and improve resource utilization. The government HPC center applies to a broad array of fields, including quantum chemistry, molecular simulation, weather forecast, weather research, oil and gas exploration, hydrodynamics, structural mechanics, and nuclear reaction. With economic development and social progress, fields such as scientific research, economic construction, and national defense have increasingly high requirements for HPC infrastructure and environments. Not only is the demand for HPC applications surging, but their scope is also expanding from traditional sectors into numerous fields vital to economic and social development, such as resources and environment, aerospace, new materials, new energy, healthcare, finance, Internet, and cultural industries.

Challenges

• Diversified computing resources

  As there are various types of HPC centers applications that deal with complex requirements, the CPU+GPU heterogeneous computing technology is needed to meet different requirements. In this case, storage with higher bandwidth and lower latency is required to fully unleash the computing potential. In addition, the HPC system needs to adapt to different performance models to cope with diversified service loads, such as the support for both bandwidth and operations per second (OPS).

• Mainstream networks: InfiniBand (IB)/RoCE

  While not all applications are network-intensive, a vast number of them rely on high-bandwidth, low-latency IB or RoCE networks to achieve large-scale parallel computing and scalability. As the rapid evolution of multi-core CPUs boosts single-node performance, the pressure on inter-node communication increases significantly. Consequently, high-speed networking like IB or RoCE has become a de facto standard for modern HPC applications. Furthermore, since the HPC center utilizes shared file systems, these high-bandwidth networks are essential for sustaining rapid data access performance.

• Parallelization and tiering of file systems

  The HPC center incorporates a large number of applications with strong computing capabilities. In addition to some I/O-intensive applications, the concurrent read/write of massive tasks also poses heavy load on the shared file systems. To handle massive data volumes and unified image files, the HPC center needs a single partition with extreme scalability. Currently, the common solution is the parallel file system that uses software to implement single partitioning and concurrent read/write of multiple storage spaces. This solution breaks the bottleneck of hardware resource design and greatly improves scalability and performance.

  Given the extensive user base, access privileges are typically categorized into multiple tiers, resulting in varying data security levels tailored to each user's profile. In addition, the type of files that can be accessed also varies with applications. Some are large files, and some are a large number of small files. Therefore, it is recommended that tiered storage be used in the solution design.

• Refined management and scheduling systems

  The HPC center has a large number of users with different permission levels, which increases management complexity. Since O&M teams primarily serve these users, they require robust mechanisms for access control, accounting, and general administration. Therefore, the HPC center demands for sophisticated management and scheduling software. Beyond core scheduling, these platforms must support a wide range of functions, such as flexible policy and permissions allocation, job accounting, user preemption, login restrictions, alarm reporting, and rapid system recovery. In addition, certain rules and regulations need to be formulated to standardize the requesting, use, and allocation of resources.

• Growing requirements for energy-efficient HPC centers

  Generally an HPC center, due to its large scale, will consume tens of thousands or even hundreds of thousands of kWh of power a year, and the electricity expenses remain high. Therefore, improving energy efficiency will help greatly reduce the energy consumption of devices and the consequent O&M costs. Low-consumption processors, energy-saving software, and infrastructure with high cooling efficiency (water-cooled units or closed cooling cabinets) are used to reduce energy consumption. In addition, as the liquid cooling technology has been mature in recent years, more and more users start to use liquid-cooled servers.

• Increasing attention to data security
  A large amount of user data is stored in a high-performance cluster, which requires close attention to user data security. In particular, to deal with the increasing malicious attacks on clusters in recent years, the security protection has become an important part of the HPC system R&D. Data security involves the following aspects:

  • The system may be vulnerable to remote exploits, internal data theft, or breaches resulting from compromised credentials. To mitigate these risks, the solution implements firewalls, encrypted file systems, and encrypted authentication to ensure robust security protection.
  • To address data loss caused by device failures, earthquakes, or fires, the solution implements data backup and recovery mechanisms.