Abstract
The chapter analyzes the ways of development of high-performance computing systems. It is shown that a real breakthrough in mastering parallel computing technologies can be achieved by developing an additional (actually basic) level in the hierarchy of hardware capacities of multiprocessor computing systems of MPP-architecture, the personal computing clusters. Thus, it is proposed to create the foundation of the hardware pyramid of parallel computing technology in the form of personal computing clusters. It is shown that on the basis of multiprocessor information systems processing and control, the control systems are implemented for many industries: space industry, aviation, air defense and anti-missile defense systems, and many others. However, the production of multiprocessor information processing and control systems is hampered by high cost at all its stages. As a result, the total cost of the system often makes it as an inaccessible tool. The use of modern multiprocessor cluster systems would reduce the costs of its production.
TopBackground
The paper analyzes the ways of development of high-performance multiprocessor computing systems. It is shown that a real breakthrough in mastering parallel computing technologies can be achieved by the development of an additional (actually, basic) level in the hardware hierarchy of multiprocessor computing systems of the MPP architecture - personal computing clusters. Thus, it is proposed to create the pyramid foundation of parallel computing technology hardware in the form of personal computing clusters. The scope of such systems is mastering parallel computing technologies, creation and debugging of parallel programs, including problem-oriented packages and libraries, as well as a simulation run of the software developed.
It is shown that on the basis of multiprocessor information and control systems, the latter are implemented for many industries: space industry, aviation, anti-aircraft and anti-missile defense systems, and many others. However, the production of multiprocessor information and control systems is hampered by high operational cost at all its stages. As a result, the overall cost of a system often makes it an inaccessible tool. Using modern multiprocessor clustered computing systems would reduce the cost of its production.
Currently, significant interest in the construction of parallel multiprocessor computing systems (MCS) is determined by the use of standard public technologies and components (Bashkov et al., 2011; The Blackford MultiCore cluster specification, http://www.mvs.icc.ru/cluster_info.html; Ivaschenko et al., 2013; Latsys, 2003). This is due to a number of factors. Let's emphasize the main ones. First, according to market needs, the performance growth of the standard network technologies such as GI(Gigabit Ethernet), FC (Fiber Channel) and IB (InfiniBand) allows them to be seen as a communication medium for multiprocessor computing systems made by the NUMA (Non-Uniform Memory Access) (Non-Uniform Memory Architecture (NUMA)). Secondly, an essential factor was the growing popularity of the Linux freely distributed operating system. At the first stage of its use, it was positioned as the UNIX platform for the platforms based on Intel architecture, but relatively fast versions appeared for other popular microprocessors, including the leader in performance over the past few years, the Alpha microprocessors.
Depending on the problems features and the budget, the project system can have all kinds of configuration options. The most affordable standard motherboard is the Core LGA1155 Z77 motherboard, which includes two free PCI Express 3.0 8 line slots on the Core i7-4790K 4 GHz platform and Gigabit Ethernet network adapters and FDR InfiniBand, which allows increasing performance, reducing latency and reducing network power boards consumption. The cluster nodes are interconnected by Gigabit Ethernet switch and InfiniBand, which is designed for the corresponding number of ports. The number of nodes in the system and their configuration depends on the requirements for computing resources in the form of specific problems, and on the user's financial possibilities.
In addition, the sharp competition among manufacturers of computer technology leads to the fact that the situation when components market prices change rather dynamically, especially with regard to the release of new products models. A wide range of modern electronic products allows to certainly state that by using standard components it is possible to build high-performance computing systems of general purpose in a very short time with the maximum full consideration of the needs and capabilities of different users.
Key Terms in this Chapter
Blade-Server Solutions: In this chapter, the authors consider multiprocessor systems where several similar motherboards are installed in the same package. Practice proves that “blade” systems are more compact and easy to maintain, and their implementation is not much more expensive compared to multiprocessor computing systems.
Supernumerary Devices: In fact, here we are talking about multiprocessor parallel computing systems of MPP architecture (massively parallel processing).
InfiniBand Technology: The exchange of data among computing nodes of a multiprocessor system is transferred to a separate network using the InfiniBand network interface, which increased the system's overall performance and significantly reduced the latency (channel download) of the network that connects nodes of the cluster computing system.
Vectorization of Computations: The creation of parallel computing systems required the development of a new approach for making of appropriate computational algorithms. For this reason, new ideas and approaches are emerging in computational mathematics, which is aimed towards the development of new methods for realizing numerical experiments based on the computations parallelization.
“Odd-Even” Reduction Algorithm: Among the known algorithms for recursive decomposition of solutions of systems of equations, only the algorithm of cyclic reduction of permutations allows the highest degree of its vectorization. In the algorithm of such an “odd-even” reduction, there are many possibilities for parallelization. In this connection, this approach is based on the method of parallelizing a mathematical problem.
Communication Environment: For the construction of supercomputers are taken serial microprocessors, each with its own local memory and connected through some communication environment.
Cluster Computing Systems: Multi-processor computing systems built on the basis of local networks, known as “cluster computing systems.”