Realize 500 billion times breakthrough hp help Huazhong University of Hust to build a leading grid computing platform

In the Chinese Educational Research grid plan, Huazhong University of Science and Technology is the lead unit, and is one of the first batch of 12 institutions participating in the plan. The plan requires all institutions to integrate computing platform, but by the current network conditions in China Limited, so that each node has a relatively large computing node, the first batch of universities to join the program requires 500 billion computing power of the Supercomputing node.

In the traditional sense, High-performance computing applications are CPU and memory-intensive applications, it has a number of major requirements for computer architecture (Super computer architecture): Floating-point computing capabilities (especially 64-bit double-precision floating-point operations), memory bandwidth and memory capacity, as well as the system framework. In high-performance computing, these factors are interrelated. Since high performance computing requires a variety of mathematical equations to establish models and simulate physical phenomena, and as various models become larger and more complex, the size of the dataset has increased dramatically. therefore Supercomputers that handle such a large workload require faster CPUs, higher-performance memory and I/O subsystems to achieve the highest computational speed, multi-level parallel processing technology that can handle a task with dozens of or even thousands of processors, and high performance interconnect equipment and system design, To achieve the highest performance.

High-performance computing needs clear

At present, the basic design idea of the global High-performance supercomputer is through high scalability, high-bandwidth, low-latency interconnected devices link less than dozens of, thousands of commercial SMP servers, supported by the corresponding system software, with the support of popular industrial standard languages and tools, to form an easily scalable, Super computer systems that are easy to parallel programming.

Based on the above thinking, Huazhong University of Hust's grid computing solution for High-performance computing environment, the following requirements: Multiple high-performance SMP computing nodes and management nodes composed of high-performance computing cluster; Each compute node is configured with 2GB of memory; The internal system and file system of each compute node is configured with a high speed hard drive The computing nodes adopt high bandwidth and low latency interconnection equipment, secure and efficient network interconnection, open common Linux operating system, common efficient parallel mode, support OpenMP, MPI, multi-level task scheduling management system, support industrial standard language environment, including C, C + + and FORTRAN, etc. A general and extended mathematical library for High-performance computing applications, including basic linear algebra, matrix operations, fast Fourier transform and other scientific Computing program Library software.

A powerful and flexible grid computing platform

After a comprehensive analysis of the high performance computing requirements of HUST, HP designed a high-performance computing system that enables users to achieve maximum performance and flexibility in the same investment situation.

The mainframe system uses 48 HP Integrity rx2600 Kinetic servers based on the Intel? 2 architecture as compute nodes, 1 hp Integrity rx2600 kinetic energy servers based on the Intel Itanium 2 architecture as management nodes, Each compute node server is configured with a Gigabit Ethernet as the computing network, and a 1OOMbps Ethernet as a management network.

The 2-way HP Integrity rx2600 Kinetic server is equipped with the 1.5 GHz 2 processor with a 6MB three cache, using the HP ZX1 chipset, which enables the reduction of memory latency and increases the scalability of the memory and I/O subsystems. Enables the rx2600 kinetic energy Server to achieve industry-leading performance and memory scalability, processing more analog data, but less cost and complexity.

In the application of High-performance computing system, not only the high computational ability, but also the high requirements for storage, because high-performance computing applications often require a short time to read and write large quantities of data, which has a high demand for the performance of storage systems, especially the I/O performance of multiple parallel operations while accessing large amounts of data. At present, in the High-performance computing domain, when the number of nodes is relatively small, SAN storage technology can be used to build a separate SAN storage network. However, because the current cost of San construction is still relatively high, and the maximum number of ports supported is only 200-300 nodes (the specific number of different vendors), so when the number of nodes is relatively large, usually using networked storage technology. As a result, HP MSA1000, a low-cost, scalable, high-performance storage system, is a storage system with a storage capacity of 5TB in the High-performance computing systems designed for Hust.

The high performance computing system of the University of Chinese laborers using the Redhat Advance Server 2.1 os and the Linux Beowulf cluster technology, the biggest difference from the usual network workstation (network of Workstation) is two. First, the cluster system provides two independent network, has a separate computing network, independent of the Management network, customer and cluster system communication is through the management network to achieve; The second is that the whole cluster system shares a process ID number, simplifying communication between nodes.

Significant advantages

The scheme utilizes cluster technology to expand and upgrade the node computing capability or interconnect device, which protects the user's investment and ensures the system's processing ability to adapt to the increasing application demand.

The HP Integrity rx2600 Kinetic server based on the Intel Itanium® 2 architecture is used for cluster solutions, further enhancing the performance of the Itanium® processor and ZX1 chipset. Because each HP kinetic energy rx2600 Server is only 3.5 inches high (2U), up to 20 rx2600 kinetic servers can be clustered in an industrial standard cabinet-dense configuration delivers greater efficiency and high availability through the consolidation of system resources such as I/O, bandwidth, memory, mass storage, and computing capacity, Thus, it can further excavate and enhance the great development potential of rx2600 value and performance.

In the management aspect, the plan uses the HP to provide the cluster management software, the use effect is relatively good. Huazhong University of Hust originally also developed the management software, but HP this set of management software more stable and reliable, make the whole system installation and maintenance is also more convenient.

Summarize

Huazhong University of Hust also has some super computer systems, but the computational capacity is not up to the requirements, so will need to focus on the calculation capacity. The HP Integrity Kinetic Server uses the Intel Itanium 2 processor, 64-bit computing power is more powerful than the 32-bit server, and the Intel Itanium 2 processor has a strong floating-point computing capacity, a large storage capacity, and easy scalability based on the IA architecture.

At present, Huazhong University of Science and Technology has been operating on this supercomputer image processing applications, three-dimensional virtual human reconstruction, the operation is very good, and, fluid mechanics and physical applications are beginning to run on this supercomputer. The supercomputer has become the public computing platform of Huazhong University of Science and Technology, providing services for the education and research of the whole school. System stability is very good, so far there has been no problem. The application effect is better than expected.