Добавил:
Upload Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:
11З.doc
Скачиваний:
4
Добавлен:
10.09.2019
Размер:
1.31 Mб
Скачать

2.Developing of parallel program for calculating of given task

For program paralleling critical is sharing between processes, which increases with increasing number of processors and exchange operations (Figure 3.1), so great attention should be given to autocommutation of systems and indicators.

Figure 3.1 – Calculation and changing time in MIMD’s systems

Nowadays the most common of these avtokomutatsiyi: InfiniBand, Ethernet, Myrinet, SCI, QsNet, RapidIO....InfiniBand

Infiniband - high-speed switched serial bus that is used both for internal (inward), and for mizhsystemnyh connections.

Like PCI Express Infiniband uses bidirectional serial bus. Basic rate - 2.5 Gbit / s in each direction, now used often, Single Data Rate (SDR) channels 1x, 4x and 12x. Supports also work with Double Data Rate (DDR, 20 Gbit / s) and Quad Data Rate (QDR, 40 Gbit / s).

The main purpose of Infiniband - mizhserverni connection, including the organization and RDMA (Remote Direct Memory Access).

Specifications:

- Manufacturers: the most common hardware solutions from companies Qlogic, Mellanox, Voltaire, Topspin.

- Tests show producers throughput at MPI about 800 MB / s and 1-7 ms delay.

- Topology: switched with Fat Tree for larger installations, the existing switches support a large number (> 256) ports (Example: SilverStorm).

- Software: Drivers from the manufacturers of hardware, different MPI libraries both commercial and free.

Myrinet

Myrinet - the most widely used for building clusters of the type of communication environment. About 10% of clusters Top500 list of most powerful computers built using peace Myrinet.

Specifications:

- Manufacturer: Company Myricom.

- Bandwidth: 250 Mbps, 1250 Mbps (Myri-10G).

- Time of delay - about 10 microseconds.

- Topology: switched, switching element is a 8x8 matrix. Switches on the basis of its support to 128 ports. To build a large network topology used different versions of Fat Tree, best efficiency is achieved using the Clos Network.

- Software: A low-level programming interface GM, MPICH / GM, PVM / GM, stack TCP / IP, as well as commercial products - MPIPro, Scali MPI Connect.

SCI

SCI (English Scalable Coherent Interface, coordinated interaction, scalable) - a network used for building clusters. SCI-clusters have the advantage in solving problems that require a large number of short messages peresylok because such problems delay time (latency) plays a crucial role.

Features

- Manufacturer: Company Dolphin.

- Capacity: physical transfer rate - 667 Mbps, depending on the hardware platform capacity at MPI - from 200 to 325 MB / sec.

- Time Delay: This type of communication among different record low delay time: 2-3 microseconds - hardware and about 4 ms - at MPI.

- Topology: ring, two - or three-dimensional torus, and switched ring. In connection with this topology, while increasing the size of torus is hardware saturation capacity, so impractical to build clusters of size torus over 6-8 for each measurement. Toroidal topology does not require the use of switches.

- Software: free (ie, distributed under the license GNU GPL) low-level software from Dolphin (drivers and library SISCI), SCI-MPICH from Aachen University, as well as commercial software from Scali.

QsNet

QsNet - communication environment from Quadrics, which provides the delay at the SCI and the capacity to 900 megabytes per second (QsNet II). Given the high cost of equipment QsNet, typically used to build large clusters especially teraflopnoho range.

Specifications:

- Capacity: MPI at 900 MB / sec.

- Time delay: 3 ms.

- Topology: Fat Tree. The architecture supports up to 1024 units (QsNet) and 4096 nodes (QsNet II).

- Software under GNU / Linux is distributed with the original text of the license GNU GPL and supports MPI (specialized version of MPICH) and TCP / IP.

General characteristics of systems autocommutation shown in Table 3.1

Table 3.1- Characteristics of autocommutation systems

Type

Capacity

Delay time

Transfer time of 1 bit on MPI level

Capacity on MPI level

InfiniBand

2.5Gb/s

1-7mcs

1.2ns

800Mb/s

RapidIO

800Mb/s

6mcs

?

?

Ethernet

10Mb/s

9.6mcs

95.4ns

10Mb/s

QsNET

900Mb/s

3mcs

1.1ns

900Mb/s

SCI

667Mb/s

4mcs

4.8ns

200Mb/s

Myrinet

250Mb/s

10mcs

?

?