WO2005031549A2

WO2005031549A2 - Parallel data processing device

Info

Publication number: WO2005031549A2
Application number: PCT/CH2004/000603
Authority: WO
Inventors: Ben Moore; Joachim Stadel
Original assignee: Universität Zürich
Priority date: 2003-09-29
Filing date: 2004-09-28
Publication date: 2005-04-07
Also published as: WO2005031549A3

Abstract

The invention relates to a parallel data processing device (1) consisting of a great number of data processing units (14) Each data processing unit is provided with a horizontally oriented plate (140) and at least one CPU or a store unit mounted thereon. Several processing units form a horizontal group (4), wherein data processing units of each horizontal group are disposed substentially on the same horizontal plane and surround a vertical opening. Several of said horizontal groups (4) of the date processing units (14) are superimposed in such away that said vertical openings form a vertical air guiding hole (16). Each date processing unit (14) is provided with means (142) for connecting at least one network cable (151, 152, 153, 154) disposed on the external surfaces of the data processing arrangement (1). A deflector (17) for deviating air is also disclosed.

Description

Parallel data processing device

Technical field

The invention relates to a parallel data processing device, i.e. a device with a multiplicity of similar data processing units.

Background of the Invention

Computer technology has experienced a dramatic development in recent years. High-performance computers, in particular workstation computers, are commercially available today for a relatively small amount. This has had an impact on the development of so-called supercomputers. A supercomputer is a computer whose performance is significantly higher than that of conventional workstations or workstations and which is optimized for handling very computing-intensive tasks.

Supercomputers are generally characterized by a special architecture, which differs from the architecture of conventional workstations. A construction principle that has become almost completely established in recent years is the use of a large number of central microprocessors (Central Processing Units or CPUs). The high computing power results from the simultaneous ("parallel") use of many CPUs. In the area of these so-called parallel There are a number of different types of computer, such as shared memory parallel computers (SMP), message passing parallel computers (MPP) or other architectures. In the context of this description, a parallel computer is to be understood as any computer that has a plurality of CPUs, for example more than four CPUs, and means for exchanging data between the CPUs.

One class that has become increasingly popular in recent years is the so-called Beowulf cluster. These contain a large number of commercially available computing units, as are used in conventional workstation computers. The computing units, which are usually mounted on a single motherboard (often referred to as a motherboard or mainboard), are usually installed in a specially manufactured frame. They are often cooled with a special ventilation arrangement. The individual computing units are connected to each other and to a central server via a fast network. Linux is often used as the operating system. An advantage of such a Beowulf cluster is that largely commercial components are used which are used in the mass market and are therefore available at low cost.

There are various problems when designing a Beowulf cluster, but also when designing any other parallel computer. One of these problems is related to the cooling of the heat emitting components, particularly the microprocessors. Because of the high number and density of such components in a parallel computer, a very high power density of the heat emission can result. The heat must therefore be dissipated efficiently. A large number of individual is often used for this Fans are used which blow or suck air through each individual computing unit. Such an arrangement is described, for example, in US Pat. No. 6,374,627 to Schumacher et al. proposed. A large number of computing units are arranged in several rows of straight frames. Between two rows of frames, cold air is supplied upwards from the floor. With the help of a large number of individually electrically operated fans, the cold air is redirected and led past the heat-generating parts. However, this leads to problems with a large number of computing units, because even the failure of a single fan can lead to malfunctions.

Various arrangements have therefore been proposed to avoid this and other problems. For example, U.S. Patent 5,150,279 to Collins et al. proposed an annular arrangement of computer boards. The cooling air is fed axially into the interior of the ring, where a structure with a plurality of cooling plates is attached, each of which deflects part of the cooling air flow in a radial direction towards horizontal plates. The horizontal boards only contain memory modules and modules for data transfer. The CPUs, on the other hand, are attached to their own, vertically arranged boards on the outer edge of the ring. This structure is relatively complex and cannot be transferred to the construction of a cluster of independent computing units.

Another problem concerns the networking of the individual computing units. In order to ensure fast data transmission, it is desirable to keep the connections between the computing units as short as possible. To solve this problem, for example, is in US Pat. No. 6,327,143 have been proposed to arrange a plurality of computing units along an arc. The connection between the computing units is made using cables on the inside of the circular arc. This ensures a shortening of the cable lengths compared to a linear arrangement. A disadvantage of such an arrangement is the poor accessibility of the network cables.

Similar problems also exist with other parallel data processing devices, in particular with parallel storage devices in which there are a large number of similar storage units, produce heat and have to be networked with one another.

Summary of the invention

It is therefore the task of creating a parallel data processing device with a plurality of data processing units, which enables efficient cooling without the disadvantages mentioned, which is simple in structure, has simple accessibility to the data processing units, and which enables the individual data processing units to be easily accessible for networking ,

The object is achieved by a parallel data processing device having the features of claim 1.

Accordingly, a plurality of independent data processing units are available, each of which has a (main) circuit board with at least one of its own CPUs mounted thereon or its own memory unit. As a rule, further components will be arranged on each of the main boards. The boards are aligned horizontally. You are so arranged that in each case a number of data processing units are located essentially in the same horizontal plane; these form a horizontal group. The arrangement of the data processing units within a horizontal group is essentially ring-shaped, ie the data processing units are arranged in such a way that they define, enclose and leave a central, vertical opening. Several horizontal groups of data processing units are arranged one above the other and are usually oriented in the same way. In this way, the central openings of the horizontal groups come to lie one above the other and form a vertical shaft. This is used to supply or remove air that can be used to cool the data processing units. In order to network the individual data processing units with one another, means for connecting at least one network cable are provided on each data processing unit. In order to ensure easy accessibility, these means, usually network cards, are arranged on the outside of the parallel data processing device.

Due to the construction according to the invention, a cooling air flow can be fed centrally and can be divided equally among the horizontal groups of data processing units. The arrangement of the network cards on the outside enables easy access to the network connections.

Furthermore, the invention comprises a parallel data processing unit with the features of claim 8 and a deflector with the features of claim 9. By providing a deflector with a continuously expanding shape, ie a deflector whose horizontal one Cross-sectional area increases steadily in a vertical direction, this enables a uniform deflection of air supplied from the vertical shaft into the horizontal groups of data processing units or a uniform suction of air from the data processing units into the shaft.

The invention further comprises a method with the features of claim 10 for operating a parallel data processing unit with a vertical shaft. In this process, cooling air is fed into the vertical shaft from below, which enables particularly efficient cooling.

A parallel data processing device is to be understood here to mean any data processing device in which a plurality of similar data processing units are used, which are connected to one another in such a way that they can exchange data with one another. In particular, this term should be understood to mean: parallel computers and parallel storage devices.

Advantageous embodiments result from the dependent claims.

Brief description of the drawings

The invention is explained in more detail below on the basis of preferred embodiments with reference to the drawings, in which

Fig. 1 shows a schematic side sectional view of a parallel computer; Fig. 2 is a schematic perspective view of a Shows part of a parallel computer;

3 shows a schematic perspective view of a parallel computer;

4 shows a temperature profile through the parallel computer; such as

Fig. 5 shows three possible profiles of a deflector.

Detailed description of the invention

1 shows a parallel data processing device in the form of a parallel computer 1 in a schematic sectional side view. The parallel computer 1 is set up on a floor 2 with floor plates 21. The floor 2 is a blind floor, ie there is a cavity under the floor 2 through which cables and air inlets can run. The parallel computer 1 comprises a frame 11 on which a plurality of horizontally arranged frames 12 are attached as holding means. In the present embodiment, there are forty-eight such frames 12. A base plate 13 can be inserted into each frame 12. Each base plate 13 carries three computing units (data processing units) 14. A total of one hundred and forty-four computing units are therefore present in the parallel computer 1. Each computing unit 14 comprises a main circuit board 140, as is also used in a similar form in commercially available workstation computers. Various components of the computing unit 14 are mounted on each main board 140, including two CPUs, memory modules, a power supply unit 141 and a network card 142. A total of two hundred and eighty-eight CPUs are thus present. The parallel computer 1 further comprises a housing 18, which is arranged at least on the sides of the computer for protection purposes. The housing 18 is essentially formed by perforated sheet metal plates 181, which vertical supports 183 are attached from angle profiles and allow air to flow through from the inside of the housing into the surrounding outside space, and to which handles 182 for attaching and removing the sheet metal plates 181 are attached. However, the housing is not critical for the function of the parallel computer, and so any other material can be used instead of perforated sheet metal plates, which ensures the outflow of air. Perforated sheet metal has the advantage that it is suitable for shielding electromagnetic interference by acting as a Faraday cage.

An inner part of the frame 11 extends into an opening 22 of the floor 2. The inner part is essentially formed from four vertical square tubes 111 on the vertical edges of a tower and a plurality of horizontal square tubes 112 which connect the vertical square tubes 111. The interior of the frame 11 leaves a vertical shaft 16 free. The shaft is open at the bottom and has a plurality of openings 113 on the sides, which are delimited by the vertical and horizontal square tubes 111, 112. In this shaft 16, a deflector 17 is arranged, which has the shape of a straight, regular, four-sided pyramid with a square base standing on top. The deflector 17 closes the shaft 16 towards the top. It serves to divide a cooling air flow 3 fed to the shaft 16 from below and to divert it in the direction of the computing units 14.

FIG. 2 shows a part of the parallel computer 1 in a schematic perspective illustration, which enables the structure to be recognized more clearly. Parts of the frame 11, the housing 18, the deflector 17 and a number of frames 12 with base plates 13 and computing units 14 have been omitted for the sake of clarity. The same parts are provided with the same reference numbers as in FIG. 1. In particular, the frame 11 and the shaft 16 running in its interior can be seen, in which the deflector (not shown in FIG. 2) is inserted. Furthermore, the individual frames 12 can be seen, each holding a base plate 13 with three computing units 14 each. A network card 142 is mounted in each computing unit 14 on the outward-facing side of the computing unit 14.

The concept of network connections between the computing units is also shown by way of example in FIG. 2. In this application, a network connection is understood to mean any connection that is suitable for data transmission between the computing units. The same applies to the term network cable and other terms related to a network. In particular, a network connection can be a connection in which electrical signals are transmitted or a connection in which optical signals are transmitted. Accordingly, a network cable can be a suitable electrical cable or a suitable optical fiber cable. Hybrid forms of data transmission are also conceivable. Each computing unit is connected to the next two computing units, which lie in the same horizontal plane, by a network cable 151 or 152, which is connected to the network card of the computing unit. As a result, there is a closed, horizontal ring of network cables 151, 152 in each horizontal plane, via which each computing unit 14 is connected to every other computing unit 14 in the same horizontal plane. In addition, all computing units are ten, which are vertically one above the other, connected to one another by a second, vertical ring of network cables 153, 154. This vertical ring follows the following scheme: The lowest computing unit of a stack of computing units lying vertically one above the other is connected to the second lowest computing unit via a short vertical network cable 153. The bottom computing unit is also connected to the third bottom computing unit via a longer vertical network cable 154. Similarly, the top computing unit is connected to the second and third uppermost computing units via two vertical network cables. All other computing units are connected to the computing unit after next in the vertical direction both downwards and upwards. Overall, this results in a closed vertical ring. Since there are twelve vertical stacks of twelve computing units each in the arrangement shown, the vertical rings of network cables together form a torus of network cables. Together with the twelve horizontal rings, one speaks of a double toroidal wiring diagram. It goes without saying that the illustration in FIG. 2 is only schematic, in particular that network connections exist between all and not only between the computing units shown. Other network topologies are also possible, such as communication via a single central switch or networking of several independent double toroidal arrangements. Three-dimensional toroidal networks have also been described. In the present case, however, the double toroidal scheme of data exchange described here reflects the physical construction principle of the parallel computer particularly well.

Each network card provides an interface between the internal data bus (eg PCI bus) of the computing unit and one or several network cables. A preferred protocol for the operation of the network connections is the so-called SCI (scalable coherent interface) protocol. This has been available since approx. 1992 and is standardized under the ANSl / lEEE standard 1596.

In the present embodiment, a single network card per computing unit allows the connection of both the horizontal and the vertical network cables, i.e. there are at least four connections for network cables available per network card; however, there may also be several network cards, in particular two network cards, e.g. one network card is responsible for communication with the computing units in the same horizontal plane and another network card for communication with the computing units within the same vertical stack.

The network connections between the computing units must be able to transmit the largest possible amounts of data per unit of time. Various systems are available on the market for fast data transmission. All systems have in common that their latency increases with the distance over which data is transmitted. This is due to physical reasons, since the data can be transmitted at the maximum at the respective group speed of the transmission medium, at the maximum at the vacuum light speed. As far as systems are concerned in which electrical signals are transmitted, these systems also have the problem in common that the distances over which data can be transmitted become shorter with increasing transmission rate or bandwidth. To ensure a short latency and high bandwidth of data transmission, the connections must be as short as possible being held. This requirement is well met in the present parallel computer. The connection lengths between the computing units which are mounted on the same base plate (here three each) are essentially identical to the width of a computing unit. The shorter horizontal network cables 151 are used for connection. Longer horizontal network cables 152 are required for connections to the next base plate. The shorter and longer horizontal cables can be of the same type or of a different type. A major advantage of the present structure is that no network cables are required that are considerably longer than the average cable length. However, this cannot be avoided with many other types of parallel computers.

Another network connection, not shown in the drawings, exists between each computing unit and a central server or switch (likewise not shown). These connections are preferably commercially available Ethernet connections, for example 100 Mbit or 1 Gbit Ethernet connections. The components required for this (especially network cards and cables) are also used in the mass market and are therefore available at low cost. The performance of these connections is not of critical importance, since during the calculations, as a rule, a large part of the communication, as far as this is necessary, takes place between the computing units via the fast double-toroidal network of the parallel computer. The central server preferably comprises a so-called Ethernet switch, which is able to control the one hundred and forty-four individual network connections to the computing units and thus to enable communication between the server and each computing unit. For the network connection of every computing unit with the central server, a second network card, not shown in the drawings, is present on each computing unit. This is also preferably arranged on the outside in order to ensure easy access.

3 shows the entire parallel computer again in a highly schematic, perspective representation. In order to make the individual parts of the computer visible, the sheet metal plates 181 of the housing 18 are shown transparently. In particular, further vertical supports 113 of the frame 11 can be seen on the outer edges of the parallel computer 1, which increase the stability and stability.

An important feature of the present parallel computer is the vertical shaft 16. This serves for the supply of cooled air, which is supplied to the computing units 14 for cooling through the side openings of the frame 11. For this purpose, for example, a circulating air cooler can be provided, which is installed in the same or an adjacent room as the parallel computer. Cold air is led through the blind floor 2 from the circulating air cooler to the shaft 16 and blown into it as a primary air stream 3. In order to ensure a uniform distribution of the air from the shaft into the computing units of the different horizontal levels, the deflector 17 is provided. In Fig. 1, the deflector has the shape of a straight, regular, four-sided pyramid standing on the top. An important feature of the deflector is its continuous shape. There are no protruding parts, eg baffles, on the deflector. This enables a simple and elegant construction. The primary air flow 3 is divided by the deflector and deflected evenly onto the horizontal levels of computing units. Each arithmetic unit 14 is thereby affected by a partial air flow 31 flowed through and cooled. The exhaust air radially leaves the computing units 14 and passes through the permeable housing 18 into the exterior of the parallel computer 1. Here it can be sucked in again by the circulating air cooler, filtered and cooled again. The deflector 17 can be made of any suitable material, for example plastic, metal sheet, wood or cardboard. However, the material preferably has a low thermal conductivity in order to avoid heat losses from the outside.

The cooling air flow directed through the shaft onto the computing units has a number of advantages. The most important advantage is that no active devices are required to divert air from the primary air flow 3 into the individual computing units. In many parallel computers of the prior art, for example, electrically driven fans are provided for this purpose, but they can often fail and thus cause local overheating. In addition, the maintenance of such computers cooled with active devices is relatively complex. In contrast, the cooling components inside the present parallel computer do not require any maintenance. Another advantage is that clean, filtered air can be supplied to the shaft 16 from the cooler. The uniform air flow from the inside to the outside, which flows through the parallel computer 1, prevents the ingress of dust from the outside, so that no dust can be deposited in the computer that could impair the operation of the computer. Dust protection measures in the housing 18, as are common in other computers of the prior art, are thus unnecessary. Furthermore, the cooling capacity in the present arrangement of air flows within the surrounding space as well as the temperature in the surrounding space is going independent.

Of course, active fans can nevertheless also be provided in the individual computing units. As a rule, each CPU will have its own CPU fan, which generates a directed air flow to the CPU. In the event of such a fan failing, as well as in the case of other errors, the present parallel computer offers further advantages. For example, due to the construction with a frame on which frames are arranged, in each of which a base plate is inserted, due to the external network cables, and due to the double toroidal cabling, a single base plate 13 with three computing units 14 is made for maintenance purposes to remove the parallel computer. For this purpose, even other parts of the parallel computer can remain in operation almost unaffected.

The parallel computer shown was set up and tested in the laboratory. One hundred and forty-four computing units were used. The width of each computing unit was little more than the standard width of a motherboard, i.e. less than 35 cm. The vertical distance between two computing units was approx. 3 standard rack units, ie around 14 cm. Due to the relatively large vertical distance, standard power supplies and fans could be used; In addition, the network cards could be mounted directly on the boards of the computing units without an adapter. Furthermore, the relatively large distance ensures a good flow of cooling air. The square cross section of the vertical shaft 16 had an edge length of approximately 35 cm. This turned out to be perfectly adequate to achieve sufficient airflow. The deflector 17 had the pyramidal shape shown in Fig. 1 and was made of cardboard. Overall pointed the parallel computer has dimensions of approx. 175 2 175 cm ² at a height of approx. 190 cm. Each processing unit had two CPUs. A total of two hundred and eighty-eight CPUs with a clock frequency of 1.8 GHz were thus used. Each computing unit had 1 GB of RAM space. In total there was an additional 11.5 TB hard disk capacity. The network between the computing units reached a transfer rate of 4 GBit per second with a latency of 5 microseconds. Overall, a computing power of 0.57 Tflops (5.7 x 10 ¹¹ floating point operations per second) was achieved in the so-called Linpack benchmark (a widely used performance test). The parallel computer generated a thermal output of approx. 45 IkW during operation. Two circulating air coolers were used for cooling. The heat generated was easily dissipated, despite the ambient room temperature of up to 32 ° C.

4 shows a graphical representation of the temperatures T (in ° C.) in the CPUs of the individual horizontal groups of computing units. For this purpose, the temperature of each CPU was measured, and the temperatures determined in this way were in each case averaged over a horizontal group (i.e. plane). The illustration shows this average temperature (horizontal axis) depending on the level N (vertical axis). It can be seen that the temperature distribution is relatively uniform, with the highest temperatures occurring in the third and fourth lowest levels. None of the temperatures exceed 60 ° C. The CPUs remain functional up to approx. 75 ° C.

In order to further improve the uniformity of the temperature distribution, the profile (ie the shape of a central longitudinal section) of the deflector can be changed. 5 shows three possible profiles 501, 502 and 503. By Such a profile, whose slope (ie, the increase in horizontal width in relation to a height unit) is greater in the lower area than in the upper area, ensures that a larger amount of air is deflected into the lower levels by computing units and these are thereby cooled more. The optimal shape can easily be adjusted empirically. All deflectors with such profiles have in common with the pyramid shape that the width of the deflector increases steadily in a vertical direction.

Instead of a deflector with a square base, a deflector is also conceivable that has a different type, e.g. has a round or polygonal base, e.g. a deflector in the form of a cone. It is important that this does not affect the ability of the deflector to evenly distribute the air supplied to the levels of the parallel computer. It is also not absolutely necessary for the deflector to have a tip at one end. A shape e.g. of a truncated pyramid can be particularly advantageous if the horizontal dimensions of the parallel computer are larger than in the case described here, in particular if the cross-sectional width of the shaft increases, e.g. goes beyond 40 cm or 50 cm. In this case, it may be advantageous to divide the primary air flow into a number of partial air flows before the supply to the deflector, which preferably corresponds to the number of side surfaces of the deflector.

Of course, various changes to the structure of the parallel computer are possible without leaving the underlying principles. The number of computing units and processors can be easily scaled up or down without having to change the construction principle. Instead of the frame 12, other suitable holding means for the base plates can also be used 13 are used. Instead of arranging three computing units 14 on a base plate 13, these can also be more or fewer computing units, for example two or four. The base plates 13 can also be omitted entirely, and instead one computing unit 14 can be held individually by a holding means. Instead of an arrangement with a square plan, an arrangement with a triangular, pentagonal or polygonal plan is also conceivable. Accordingly, instead of twelve, a horizontal group of computing units can also include more or fewer computing units. The deflector will then be adjusted accordingly. The frame can also be made of other materials instead of square tubes, as long as it allows a radial flow from the shaft to the computing units.

In one variant, the flow of cooling air can take place in the opposite direction. In this case, a deflector is used, which widens from top to bottom and closes the shaft 16 at the lower end. Air is actively drawn off at the top of the shaft. This creates a flow through the computing units radially from the outside inwards. The radial air flow is deflected upwards by the deflector and discharged. This variant is suitable e.g. when there is no blind floor available and therefore an air supply from below is difficult. However, with such a design principle, measures may be necessary to prevent dirt or dust from entering the computing units. Furthermore, with this variant it is necessary to cool at least a limited part of the outside space of the parallel computer, which can be associated with higher losses.

Finally, it is for the operation of the deflector 17th it is not absolutely necessary that the computing units close the vertical shaft u. It is therefore also possible for the computing units 14 to be mounted in longitudinally aligned racks, as is the case in normal so-called rack-mount systems (as shown, for example, in principle in US Pat. No. 6,374,627). Two such racks are arranged in parallel, and the space between the racks is closed at the ends of the racks with suitable means (for example, by a vertical wall). The two frames together with the two end means form a shaft. In this a deflector is attached, which widens horizontally along a vertical direction. In the simplest case, the deflector has the shape of a straight prism, that is to say a rectangular plan and a cross section in the form of an isosceles triangle when cut along a plane perpendicular to the longitudinal direction of the frames. Analogous to the description in connection with FIG. 5, the deflector can also have the cross sections 501, 502 or 503 shown there or other suitable shapes.

In the previous explanations, reference was mostly made to a parallel computer. However, another type of parallel data processing device is also covered, in particular also such a device in which the primary task of the individual data processing units is not to execute algorithms, but rather to efficiently store and make data available , Storage can be volatile in the event of a power failure, such as in RAM modules (RAM = Random Access Memory), or non-volatile, such as in particular on hard disks. The efficient storage of large amounts of data on a large number of hard drives is of particular importance. This creates similar requirements to a parallel computer: Fast, parallel data access to several hard disks and data exchange between them must be guaranteed, ie the individual storage units on which the hard disks are mounted must be networkable. In addition, the individual hard disk drives and power supplies produce heat that must be dissipated efficiently. For example, a parallel computer of the Beowulf class can also be understood and used as such a "parallel memory", since a standard hard disk is normally available on each main board. With a hard disk capacity of 80 Gbytes per main board, such a parallel memory achieves a storage capacity of 11.5 TB (Terabyte) A multiple of these is easily conceivable, especially if several hard disk drives are installed per motherboard.

Claims

claims

1. Parallel data processing device (1), comprising a plurality of data processing units. (14), each data processing unit (14) having a horizontally oriented circuit board (140) and at least one CPU or a memory unit mounted thereon, - in each case a plurality of data processing units (14) forming a horizontal group (4), wherein the data processing units of each horizontal group (4) are arranged essentially in the same horizontal plane and close a vertical opening, - a plurality of such horizontal groups (4) of data processing units (14) being arranged one above the other in such a way that the said vertical openings are arranged one above the other and form a vertical duct (16) for guiding air (3), and wherein each data processing unit (14) comprises means (142) for connecting at least one network cable (151, 152, 153, 154), and wherein these means (142) are arranged on the outside of the data processing device (1).

2. Parallel data processing device (1) according to claim 1, characterized in that the data processing device (1) further comprises a deflector (17) for deflecting air, which is arranged in the interior of the vertical shaft (16) and has a continuous shape which expands horizontally along a vertical direction.

3. Parallel data processing device (1) according to claim 2, characterized in that the deflector (17) widens horizontally from bottom to top and closes the vertical shaft (16) at the upper end.

4. Parallel data processing device (1) according to claim 2 or 3, characterized in that the vertical shaft (16) has a square plan and the deflector (17) has the shape of a straight four-sided pyramid with a square plan.

5. Parallel data processing device (1) according to one of claims 1 to 4, characterized in that the data processing device (1) furthermore has horizontal and vertical network cables (151, 152, 153, 154) which are completely on the outside of the data processing device (1) that the horizontal cables (151, 152) each connect all data processing units (14) in a horizontal group (4) in a ring, and that the vertical network cables (153, 154) each have a plurality of data processing units (14) from different horizontal groups (4 ) connect in a ring.

6. Parallel data processing device (1) according to one of claims 1 to 5, characterized in that the data processing device (1) further comprises a frame (11) which encloses the vertical shaft (16) and lateral openings (113) for the flow of air owns, and that on the frame (11) a plurality of horizontally arranged holding means (12) is fastened, at least one data processing unit (14) being held by each holding means (12).

7. Parallel data processing device (1) according to claim 6, characterized in that the data processing device (1) further comprises a plurality of base plates (13), that each holding means (12) is a frame, that each frame holds a base plate (13), and that at least two data processing units (14) are arranged on each base plate (13).

8. Parallel data processing device (1), comprising - a plurality of data processing units (14); - a vertical shaft (16) to which a plurality of the data processing units (14) adjoin; and - a deflector (17) for deflecting air, which is arranged in the interior of the vertical shaft (16) and has a continuous shape which expands horizontally along a vertical direction.

9. deflector (17) for deflecting air, which is designed for use in a vertical shaft (16) of a parallel data processing device (1), the deflector (17) having a continuous shape which expands horizontally along a vertical direction ,

10. A method for operating a parallel data processing device (1) according to one of claims 1 to 8, characterized in that cooling air (3) is fed to the vertical shaft (16) from below.