What is a supercomputer?

Supercomputer is a very large electronic computer. It has a strong ability to calculate and process data, and its main features are high speed, large capacity, various external and peripheral devices and rich high-function software systems.

Supercomputer is actually a huge computer system, which is mainly used to undertake major scientific research, cutting-edge national defense technology, large-scale computing topics and data processing tasks in the field of national economy. For example, large-scale weather forecasting, sorting out satellite photos, exploring nuclear materials, studying intercontinental missiles and spaceships, and making national economic development plans are all numerous and time-consuming. It is necessary to comprehensively consider various factors and rely on a supercomputer to successfully complete it.

Some economists stipulate the indicators of supercomputers: First, the average computing speed of computers is greater than 65.438+0 billion times per second; Secondly, the storage capacity is greater than 6.5438+million bits. For example, the "Galaxy" computer successfully developed by China belongs to the giant computer. The development of supercomputers is an important development direction of electronic computers. Its development level marks a country's scientific, technological and industrial development level, and reflects the strength of national economic development. Some developed countries are investing a lot of financial, human and material resources to develop super-large computers with operation speed of tens of billions of times.

The computer system with the fastest speed, the highest performance, the largest volume and the highest price in a certain period. Supercomputer is a relative concept. Supercomputers in one period may become general-purpose computers in the next. Supercomputer technology in one period may become general computer technology in the next period. Modern supercomputers are used in nuclear physics research, nuclear weapon design, space vehicle design, national economy prediction and decision-making, energy development, medium and long-term weather forecast, satellite image processing, information analysis and various scientific research. It is a powerful simulation tool, which is of great value to the national economy and national defense construction.

According to statistics, the performance of a computer is directly proportional to the square of its use value, which is the so-called square law. According to this statistical law, the higher the computer performance, the cheaper the relative price. Therefore, with the increasing demand for computer performance in large-scale scientific projects, ultra-high performance supercomputers will get more and more economic benefits.

First, the development of supercomputers.

In the mid-1950s, there were LARC machines of UNIVAC and rally machines of IBM. These two computers adopt parallel processing technologies such as instruction priority control, multi-arithmetic unit, memory interleaving, multi-program and time-sharing system. Supercomputers in 1960s include CDC6600 and CDC 7600, both of which are equipped with multiple peripheral processors, and the central processing unit of the host computer contains multiple independent and parallel processing units. Modern supercomputers appeared in the 1970s, and the execution speed of instructions has reached more than 50 million times per second, or floating-point results can be obtained more than 20 million times per second.

Modern supercomputers have gone through three stages of development. In the first stage, there are American supercomputers such as Ilium -IV (1973), STAR- 100( 1974) and ASC( 1972). Illiac-Ⅳ computer is an array computer controlled by 64 processing units, and the last two are vector pipelined computers. CRAY- 1 successfully developed in 976 marked the second stage of modern supercomputers. This computer is equipped with general registers such as vector, scalar and address, and has 12 pipeline components. Instruction control and data access are also pipelined. The main frequency of the machine is 80 MHz, and 80 million floating-point operation results can be obtained per second; Main memory capacity100 ~ 4 million words (64 bits per word) and external memory capacity109 ~10/word; The main cabinet is cylindrical and consumes hundreds of kilowatts of electricity. Freon cooling is adopted. The picture shows the logical structure of this machine. China's "Yinhe" Billion Supercomputer (1983) is also a multi-purpose register and full pipeline supercomputer. There are 65,438+08 pipeline components, which adopt bidirectional vector array structure. The main memory has a capacity of 2-4 million words (64 bits per word) and is equipped with a large-capacity disk memory. The architectures of these supercomputers all belong to SIMD structure. Since 1980s, the third-stage higher performance supercomputers with multi-processor (multi-instruction stream and multi-data stream MIMD) structure and multi-vector array structure have come out one after another. For example, Clay -XMP, CDCCYBER205 in the United States, S8110 and 20, VP/ 100 and 200, S× 1 and S×2 in Japan all adopt the micro-assembly process of sintering ultra-high-speed gate array chips onto multilayer ceramic chips. The maximum speed can reach 500- 10 billion floating-point results per second, the main memory capacity is 4-32 million words (64 bits per word), and the external memory capacity is greater than 10 12 words.

There is also a supercomputer with strong specificity. For example, the giant parallel processor MPP of American Godel Aerospace Company is composed of 16384 processors, which is dedicated to the high-speed processing of satellite image information. The processing speed of 8-bit integer addition can reach 6 billion times per second, and the processing speed of 32-bit floating-point addition can reach160 million times per second. DAP is a special distributed array processor system developed by ICL Company in Britain. It consists of 4096 one-bit microprocessors and a large-scale serial computer 2900, and its highest speed can reach 64-bit floating-point results per second 1 100 million.

Second, the composition of the supercomputer

The host consists of high-speed computing components and large-capacity fast main memory. Due to the large throughput of giant processing data, only main memory is not enough, which is generally supported by semiconductor rapid expansion memory and mass storage subsystem. For users of large-scale data processing systems, large-scale online tape subsystem or optical disk subsystem is often needed as a medium for input/output of a large amount of information and data. Generally, the host computer does not directly manage the slow input/output (I/O) equipment, but connects to the front-end computer through the I/O interface channel, and the front-end computer does I/O work, including the writing of user programs and data, the printing and drawing output of running results, etc. Front-end machines generally use minicomputers. Another way of input and output is through the network. Users on the network use supercomputers through the network with the help of their end computers (microcomputers, workstations and small mainframes), and all I/O is completed by client computers. Network mode can greatly improve the utilization of supercomputers.

Third, supercomputer technology.

Parallel processing is the basis of supercomputer technology. In order to improve the system performance, modern supercomputers adopt various technologies supporting parallel processing in system structure, hardware, software, technology and circuit.

Data types In order to facilitate high-speed parallel processing, the data types of the central processor are all added with vector or array types in addition to the traditional scalar. The essence of vector or array operations is to perform a batch of the same operations continuously or simultaneously, while scalar operations only process one or a pair of operands, so the speed of vector operations is generally much faster than scalar operations.

Hardware structure Modern supercomputer hardware mostly adopts various technologies such as pipeline, multi-functional components, array structure or multiprocessor. Pipeline divides the whole component into several segments, so that a large amount of data can be calculated in each segment overlap, which is especially suitable for vector operation, with high cost performance and universal application. Multifunctional components can perform different operations at the same time, and pipeline technology is often used inside each component, which is suitable for both vector operation and scalar operation. China's Galaxy computer and Japanese VP/200 and S8 10/20 computers further doubled each vector pipeline component or vector processor to form a bidirectional vector array, which doubled the speed of vector operation. The vector processor of American CYBER-205 computer can form one, two or four array pipelines according to the needs of users, and the technology has been developed. Multiprocessor system improves the processing capacity of the system by working in parallel with multiple processors. Each processor can cooperate to complete a job or independently complete its own job. Each processor can also adopt various suitable parallel processing technologies. In the multiprocessor system, there are still many problems to be solved in the division and distribution of tasks, synchronization and communication between multiprocessors and the benefits of interconnection network. Modern supercomputers mainly use dual-processor systems (such as Cray -XMP) and quad-processor systems (such as HEP).

Vector Register In order to reduce the requirements of storage flow and bandwidth and solve the problem of low speed of short vector operation, the second-stage supercomputer adopts vector register technology. CRAY- 1 has eight vector registers, and all vector operation instructions are oriented to vector registers and other general registers. In order to effectively support each pipeline to execute their respective vector operations in parallel, Japanese third-level supercomputers such as VP/ 100 and S8 10 have huge vector registers with a total capacity of 64 kbytes.

Scalar operation Scalar operation speed has great influence on the comprehensive speed of supercomputer system. Therefore, in addition to adding scalar registers, scalar backup registers or scalar cache memories, advanced scalar control technologies (such as look-ahead control) can be adopted, and technologies such as functional components and scalar processors specially used for scalar operations can also be adopted. For example, among the multi-functional components of CRAY- 1, six are dedicated to scalar and address operations, and three are also used for scalar floating-point operations, and the scalar operation speed can reach more than 20 million times per second; CYBER205 has specially designed a scalar processor, which contains five operational components, and the scalar operation speed can reach more than 50 million times per second. It has become an important research topic to improve the performance of supercomputer system by increasing the speed of vector operation, further increasing the speed of scalar operation and narrowing the gap between them as much as possible.

Main memory In order to make 3D processing of complex systems possible, it is required that the main memory can accommodate a large amount of data. The capacity of supercomputers in the 1980s has reached 256 megabytes. In order to match the speed of the operation unit, the main memory must greatly increase the information flow. Therefore, the main measures are as follows: ① adopting mature multi-module cross access technology, the number of modules is generally 2n, and some supercomputers adopt prime module new technology to avoid vector access conflict as much as possible; (2) Continuously reduce the access cycle of each module. For example, the access period of Cray -XMP machine is 38 nanoseconds, and the static MOS memory of S8 10 machine is only 40 nanoseconds, which is equivalent to bipolar memory. ③ Increase the access port of the main memory. For example, compared with CRAY- 1, CRAY-XMP has four access ports per processor, which solves the bottleneck problem of storage access.

I/O channel supercomputer is not only equipped with a large number of I/O channels, such as 16 ~ 32, but also has a high channel transmission rate. For example, Cray -XMP has two channels with a transmission rate of 65,438+000 megabytes per second and one channel with a transmission rate of 65,438+0,250 megabytes per second.

Solid-state mass storage In order to adapt to the frequent scheduling of large amounts of data between main memory and external memory, a new supercomputer uses solid-state mass storage as ultra-high-speed external memory. Solid-state memory of Cray -XMP adopts MOS technology, with a capacity of 64 ~ 256 megabytes and a transmission rate 50 ~ 100 times faster than that of disk. The solid-state memory capacity of S8 10 is 256 ~ 1024 megabytes, and the transmission rate reaches 1000 megabytes per second.

The logic circuits of large-scale integrated circuit supercomputers all adopt ultra-high-speed ECL circuits, the gate delay is about 0.25 ~ 0.5 nanosecond, and the number of chip gates is tens to 1000. In 1984, Japan successfully developed a 4K gate array room temperature GaAs chip with a stage delay of about 50 picoseconds. The access time of ultra-high speed bipolar random access memory to vector register is 3.5 ~ 5.5 nanoseconds.

Assembly technology is the basis of improving the speed of supercomputer by shortening the wiring length in the machine and improving the main frequency of the machine. The main frequency of modern supercomputers has reached more than 250 MHz. Therefore, in addition to improving the integration and speed of the chip, the high-density multi-layer assembly process such as micro-assembly is also adopted. The resulting heat dissipation problem is very prominent, and special heat dissipation measures need to be taken.

Parallel Algorithms and Software Technology In order to give full play to the system performance of supercomputers, it is necessary to study various parallel algorithms and develop parallel software systems. According to the characteristics of super-large scientific computing, supercomputers are usually equipped with the following software: batch distributed operating system with multiple processing capabilities, efficient assembly language, vector FORTRAN or PASCAL, ADA language and vector recognizer, parallel standard subroutine library, scientific subroutine library and application library, system utility program, diagnostic program, etc.