RU2459239C2 - Distributed computer system of optimum solutions - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: distributed computer system of optimum solutions, based on distribution of computations PaaS, has a sever, a server accessibility verification unit, a master computer, n-slave computers, for example assume n=2, i.e. first and second, a task assignment unit, an information control unit, first and second two-way virtual buses for sending the task and the result, first and second two-way virtual buses for polling machines and a virtual initialisation bus with the following connections: server accessibility verification unit, through the master computer and the virtual initialisation bus, is connected to the control input of the server, the master computer is connected by the first and second two-way virtual buses for sending the task and the result to the first and second slave computers, respectively, the server is connected by the first and second two-way virtual polling buses to the first and second slave computers, respectively, the task assignment unit and the information control unit, which are connected by two-way virtual buses to corresponding inputs of the master computer.

EFFECT: high efficiency of a distributed computer system.

2 cl, 1 dwg

Description

The invention relates to computer technology and can be used to create distributed computing systems with optimal architecture for solving various problems: mathematical, administrative, applied, etc., requiring large computing power.

A common problem when creating computing systems or networks is to minimize the system while improving performance while simplifying software.

A significant factor, if not one of the main ones, is the reduction in energy consumption. So, the total energy costs of a modern supercomputer for 2-3 years exceed the cost of the supercomputer itself.

There is still no consensus on the optimal construction of a computing system. There is a Jaguar supercomputer, a massively parallel class supercomputer, hosted at the National Center for Computer Research in Oak Ridge, pc. Tennessee (National Center for Computational Sciences (NCCS)). The supercomputer has a mass-parallel architecture, that is, it consists of many autonomous cells (English nodes). All cells are divided into two sections (English partitions): XT5 and XT4 models Cray XT5 and XT4, respectively. The XT5 section contains 18688 compute cells, as well as auxiliary cells for user login and maintenance. Each computing cell contains 2 quad-core AMD Opteron 2356 processors (Barcelona) with an internal frequency of 2.3 GHz, 16 GB of DDR2-800 memory, and a SeaStar 2+ router. In total, the section contains 149504 cores, more than 300 TB of memory, more than 6 PB of disk space and a peak performance of 1.38 petaflops. The XT4 section contains 7832 computational cells plus auxiliary cells for user login and maintenance. The cell contains a 4-core AMD Opteron 1354 processor (Budapest) with an internal frequency of 2.1 GHz, 8 GB of DDR2-800 memory (in some cells - DDR2-667) and a SeaStar2 router. In total, the section contains 31,328 cores, more than 62 TB of memory, more than 600 TB of disk space and peak performance of 263 TFLOPS. The throughput of memory channels is 578 TB / s, and the I / O subsystem - the bottleneck of many high-performance systems - is capable of moving 284 GB of data every second.

Some of the tasks that Jaguar solves are related to climate change; another part is the modeling of combustion processes, see http://ru.wikipedia.org/wiki/jaguar/.

There is also a domestic Lomonosov supercomputer, the first hybrid supercomputer of this scale in Russia and Eastern Europe. It uses 3 types of computing nodes and processors with different architectures. The T-Blade2 blade platform is used as the main nodes providing over 90% of the system performance. It uses 3 types of computing nodes and processors with different architectures. The supercomputer uses a three-tier storage system with a total capacity of up to 1,350 TB with a parallel Luster file system. The data storage system provides simultaneous access to data for all computing nodes of the supercomputer with an aggregated data reading speed of 20 Gb / s and an aggregated writing speed of 16 Gb / s. It is supposed to use a supercomputer to solve resource-intensive computing problems within the framework of fundamental scientific research, as well as to conduct scientific work in the field of developing algorithms and software for powerful computing systems, see http: // www / t-platforms / ru / clusters / umgue / lomonosov.html. The prototype is taken by the GRID system (collider), see

http://www.gazeta.ru/science/2010/04/02_a_3346640.shtml

These supercomputers have two main drawbacks:

- relatively low performance / cost ratio;

- high energy consumption (energy costs may exceed the cost of a supercomputer in 2-4 years).

An object of the invention is to optimize the cost-effectiveness ratio.

To solve this problem, we propose a distributed computing system of optimal solutions based on the distribution of PaaS calculations, which differs from other systems in that it contains a server (a server is a computer that is selected from a group of personal computers to perform a service task without the direct involvement of a person. In some In some cases, the server and the workstation may have the same hardware configuration, see (“All About Servers” Computer Bild No. 22, 2008), server availability check block leading to computer, n-slave computers, for example, take n = 2, i.e., the first and second, task setting unit, information control unit, the first and second bidirectional virtual buses sending the task and the result, the first and second bidirectional virtual buses polling machines and virtual initialization bus with the following connections: server availability checker, through the host computer and a virtual initialization bus connected to the control input of the server, the host computer, the first and second bi-directional virtual buses sending task and the result is coupled to the first and second slave computers respectively server connected first and second bidirectional virtual tires survey with the driven first and second computers, respectively, the block statement of the problem and information control unit that bidirectional virtual tires are connected to respective inputs of the master computer.

The drawing shows a functional diagram of the system, which shows: 1 - the host computer, 2 and 3 - slave computers, 4 - the program unit for setting the task, 5 - server, 6 - unit for checking server availability, 7 - information control unit, 8 and 9 - the first and second bi-directional virtual buses for sending the task and the result, 10 - virtual buses for initialization, 11 and 12 - the bi-directional first and second virtual buses for polling machines, respectively, 13 - the virtual bus for information control.

The functional diagram has the following connections: the server availability check unit 6, through the host computer 1 and the virtual initialization bus 13 is connected to the control input of the server 5, the host computer 1 of the first and second bidirectional virtual buses sending the job and result 8 and 9 is connected to the first and second slaves by computers 2, respectively, server 5 is connected by the first and second bi-directional virtual poll buses 11 and 12 to the first and second slave computers 2, respectively, the task setting unit 4 and the information block Control 7 bi-directional virtual buses are connected to the corresponding inputs of the host computer 1.

The system operates as follows.

Server 5 dynamically monitors the status of the network via virtual buses 11 and 12.

When starting the client part, the availability of server 5, the access time to it, and the speed of the communication channel are checked. This task is handled by the server verification unit, which is a subprogram included in the client part. Then the power of the computer is checked and the possible amount of resources for distribution is calculated. This data is sent to the server, which registers the client on the network. Server 5 enters in the database the IP address (computer address on the Internet) of the client, connection parameters and machine power. After that, a signal is sent to the client machine, allowing further work.

System levels

In parallel with the processing of data from tasks that are applications of the usual level, the system operates at a low level (checks the executable code, mainly taking away system resources for a long time).

If a code is found during the viewing process, the execution of which can be optimized, the task setting unit is launched, and at the same time a request is sent to the server to search for available machines and a communication key. The task statement block is a subroutine that processes and prepares objects for further optimization. The communication key protects against theft of information by replacing machines. The server receives a signal from the client, selects a machine that is currently free, checks its availability on the network, sending a communication key, and sends connection data to the client.

The process of setting and solving new tasks is carried out constantly, regardless of the status of previous tasks.

Solving a problem on client slaves

Client machines are interconnected, verify the correctness of the communication key. If the keys match, one client machine (master) sends another (slave) task for execution. The second (slave) client machine solves this problem and sends a response.

Solving a problem on a client host

If the answer from their own computing power comes earlier, then the task is considered complete, otherwise a solution from the network is used. (With one or another method of obtaining the result, the acceptor of the result of the action on the host computer sends a signal to the server, and the server sends this signal to the slave computer, and that stops the task).

Note: the server does not store any information regarding the task being solved, except for its status (current state, task solved or not).

Possible disadvantages:

- the answer to the same task performed with the same variables may be different if random number generators are used in the solution process;

- simultaneous operation with antivirus software may not be possible (effect of simultaneous low-level access).

- the variability of time for solving problems due to the current parameters of the network load and slave client machines.

Claims (2)

1. Distributed computing system of optimal solutions based on the distribution of calculations in the PaaS format, which differs from other systems in that it contains a server, a server availability checker, a host computer, n-slave computers, for example, take n = 2, i.e. the first and second, task setting unit, information control unit, the first and second bi-directional virtual buses for sending the task and result, the first and second bi-directional virtual buses for polling machines and the virtual initialization bus, with the following connections: server availability check unit, through the host computer and virtual the initialization bus connected to the control input of the server, the host computer of the first and second bidirectional virtual buses sending the job and the result is connected to the first and second slave computers, respectively, the task setting unit and the information control unit, which are connected by bidirectional virtual buses to the corresponding inputs of the host computer, while the computing system itself is based on the technology of datagrams, i.e. on independent promotion of packets in packet networks without establishing logical channels, and in networks with the transmission of datagrams, packet routing is carried out on a packet basis, while packets are equipped with a destination address and they independently travel to destination nodes in such a way that a lot of packets that belong a single message can travel to the destination node by different routes. 2. The distributed computing system according to claim 1, characterized in that when solving problems using computers on the Internet, owned by private users, having less simultaneous workload of local tasks, especially at night, while any participant in the system - a private user - has the right to with other participants to request the required amount of resources required to solve the local problem.