This invention relates to a method of operating a hypercube network of processing devices. In one embodiment, the invention provides a mechanism to optimise data transmission within the hypercube network by calculating a central location for processing of data.

In many computing environments multiple processing devices are connected together in a network. For example, highly complex software products may be split across multiple physical machines that are connected together. Similarly, tasks that are performed on the Internet are quite often carried out by multiple servers that are connected together as a network. A further example of such multiple processing devices is within individual machines such as supercomputers, which provide enormous processing power via the provision of multiple processing devices within the single machine. When a number of processing devices are participating in a computation that requires data from multiple locations, the processing devices may need to pass a certain amount of data to a single point to be able to evaluate the computation. An example is a join query in a distributed database where data from two tables must be consolidated to a common node to evaluate the join query. The data may be present in differing volumes at different processing devices in the network.

Some computer networks, notably massively parallel supercomputers are configured in a hypercube topology. In this topology, each processing device has a small number of neighbours to which it is connected. To transmit data from one processing device to a known destination, it is passed between a succession of neighbours according to the logical addresses of the processing devices, which signify the position of the processing devices in the hypercube network, until the destination is reached. See, for example,

http://en.wikipedia.Org/wiki/MIMD#Hypercube_interconnecti on_network, for more information.

When data is required to be transmitted from multiple processing devices to a single location for processing, if the quantity of data is high and/or the bandwidth between processing devices in the network is low, it is desirable to choose to transmit all the data to a specific processing device so that the total amount of data transmitted is minimised, thereby using the minimum network bandwidth in performing the computation. There is a requirement for a method and system for determining the location of the ideal processing device in such a network in this situation.

It is therefore an object of the invention to improve upon the known art.

According to a first aspect of the present invention, there is provided a method of operating a hypercube network of processing devices comprising determining that a plurality of the processing devices are storing data to be processed at a single processing device, obtaining the addresses of the plurality of processing devices storing the data to be processed, determining the most common number for each digit of the addresses of the plurality of processing devices storing the data to be processed, generating a new address comprising the determined most common number for each digit, and transferring the data to be processed to the processing device with the generated new address.

According to a second aspect of the present invention, there is provided a system comprising a hypercube network of processing devices, the system operated to determine that a plurality of the processing devices are storing data to be processed at a single processing device, obtain the addresses of the plurality of processing devices storing the data to be processed, determine the most common number for each digit of the addresses of the plurality of processing devices storing the data to be processed, generate a new address comprising the determined most common number for each digit, and transfer the data to be processed to the processing device with the generated new address.

According to a third aspect of the present invention, there is provided a computer program product on a computer readable medium for operating a hypercube network of processing devices, the product comprising instructions for determining that a plurality of the processing devices are storing data to be processed at a single processing device, obtaining the addresses of the plurality of processing devices storing the data to be processed, determining the most common number for each digit of the addresses of the plurality of processing devices storing the data to be processed, generating a new address comprising the determined most common number for each digit, and transferring the data to be processed to the processing device with the generated new address. Thus it is possible to provide a mechanism to determine which processing device in a hypercube network would result in the minimum total amount of network traffic once the distribution of data to be retrieved is known. An advantage is that, it improves the throughput of data-intensive distributed computations across a hypercube network. For example, in a network of sixteen processing devices with four devices storing data that must be processed at a single location, the above method and system will determine which of the sixteen processing devices should do that work, while minimising the amount of data transmission that is required within the network. The selected processing device may be one of the four that are storing the data or may be a completely different processing device in the network.

The processing devices that make up the hypercube network could be individual processors within a supercomputer, for example, which have memory attached to each individual processor. The processing devices could be machines such as standard desktop computers that are connected together via Ethernet, for example, to form a network of devices.

Similarly, the processing devices could be geographically spread and connected together by a wide area network such as the Internet. The processing devices that make up the hypercube network need not be identical, different machines could be connected together as a network, for example a network of servers and client devices. A hypercube network is a network in which all nodes of the network have a unique address of the same length, with connections between nodes of the network determined by the address, so that two nodes with an address that differs by only one digit are connected together. The number of processing devices that make up the hypercube network is not material. In a theoretically mathematically pure hypercube network the number of nodes (processing devices) in the network is a power of (2 ^N ), with each node having a binary label (address) that is N digits in length. Each node is connected to those nodes that have an address that differs by only one digit from its own address. For example, if N=3, there are 8 nodes labelled 000, 001 , 010, 011 , 100, 101 , 110, and 111. Node 000 is connected to nodes 001, 010 and 100 etc. However in a practical implementation of the method and system, if the number of processing devices in the network is not a power of 2, then the network can still be configured as a hypercube network by labelling one or more processing devices with two binary addresses, as long as these two addresses are "adjacent" in hypercube topology terms and that the connections rules are still obeyed.

The method of determining the correct location of the processing device in the network for carrying out the processing of the data can be used on networks that can be categorised as "generalized hypercubes" (see Bhuyan, L.N. and Agrawal, D.P., "Generalized Hypercube and Hyperbus Structures for a Computer Network," IEEE Transactions on Computers, volume C-33, number 4, pages 323 to 333, April 1984 for more information). Generalized hypercubes expand on the standard concept of a hypercube by allowing the digits in the address label to be values other than just 0 or 1 used in binary addresses. For example all network labels could take digits 0,1 or 2, giving 00, 01, 02, 10, 11, 12, 20, 21 and 22 as the set of labels for a 2 dimensional network. All routing remains the same as in a standard hypercube, so nodes are neighbours if the labels differ by one and only one digit (02 will be connected to nodes 12, 22, 00 and 01). Preferably, the method further comprises determining the amount of data to be processed at each of the plurality of processing devices storing the data to be processed and wherein the step of determining the most common number for each digit of the addresses of the plurality of processing devices storing the data to be processed comprises weighting the determination of the most common number for each digit according to the determined amount of data. The method and system for deciding which of the processing devices to use to process the data can be performed either as an unweighted procedure or as a weighted procedure. In the weighted procedure, the amount of data that each processing device is going to transmit is used in the determination of the most common numbers in the addresses. The advantage of weighting the process is that the selection of the processing device to carry out the processing of the data will be the one that results in the least amount of data being transmitted around the network.

In a hypercube network that uses binary addresses each digit of the addresses is looked at to compare 0s and Is, which effectively splits the hypercube structure into two halves according to the dimension that is selected and then the number of contributing nodes that are in each half is determined, choosing the half with the most nodes. If the hypercube is split along all dimensions then a specific node is isolated as the centre. In a weighted case, the hypercube is split in half along each dimension to see which half contains the most total data, effectively finding the "centre of mass" of the data. Advantageously, the method further comprises, if the most common number for a digit of the binary addresses of the plurality of processing devices storing the data to be processed returns more than one result, selecting just one of the returned results. The method, whether used in the weighted or unweighted form, can return more than one address; thereby implying that more than one processing device within the network can carry out the processing at equal cost in terms of bandwidth consumption. In this case, some selection logic is required to select between the different processing devices.

Ideally, if a processing device of the plurality of processing devices storing the data to be processed is assigned more than one address, then the step of obtaining the addresses of the plurality of processing devices storing the data to be processed, in respect of the processing device that is assigned more than one address, obtains only one address for that processing device. It is conceivable that a processing device will have more than one address attached to it. In this case, in order for the procedure to function properly, only one of those addresses will be used in the calculation of the best processing device to carry out the processing of the data that is stored by the multiple devices.

Embodiments of the present invention will now be described, by way of example only, with reference to the accompanying drawings, in which: -

Figure 1 is a schematic diagram of a hypercube network, Figure 2 is a further schematic diagram of a hypercube network, Figure 3 is a table of binary addresses,

Figure 4 is a flowchart of a method of operating the hypercube network, Figure 5 is a table of binary addresses and data, and

Figure 6 is a yet further schematic diagram of a hypercube network. Figure 1 shows a hypercube network 10 comprises connected nodes 12. An n-dimensional hypercube network consists of N nodes where N is a power of 2 (2 ⁿ ). Each node 12 in the network 10 is allocated a node address which, in this example, is a binary number with n digits. In the example of Figure 1, there is a three dimensional hypercube network with eight nodes that have the three-digit binary addresses 000, 001, 010, 011, 100, 101, 110 and 111. The nodes are connected together if their binary addresses differ by one and only one digit. So the node labelled 000 is connected to nodes 001, 010 and 100. The node labelled 101 is connected to nodes 100, 111 and 001 and so on.

The shortest path between two nodes is determined by comparing the addresses of the two nodes one digit at a time and travelling to a neighbouring node when the address of the destination node and the address of the neighbouring are the same, and different from the current node, for the specific digit in the binary address. In the example of Figure 1, a message travelling from node 000 to node 101, will first travel to node 100 (the first digit changing from a 0 to 1) and then to node 101 (changing the final digit from a 0 to a 1). This defines the logical routing in a network that is configured as a hypercube network.

The same methodology applies to all hypercube networks of larger sizes. For example, if n=4 and if transmitting a network message from node 1100 to node 0111, the message can be transmitted from node 1100 to node 0100 (as the first digit of the address of the destination varies from the originator), then from node 0100 to node 0110 (the second digits of the destination and current node are the same, but the third digits are different), finally the message is transmitted from node 0110 to node 0111 (the destination). The message needs to be passed three steps to get from source to destination and this can be easily determined by counting the number of digits which are different in the source and destination addresses. In this example the addresses 1100 and 0111 differ by 3 digits so the message needs to travel three hops in the network to get between these nodes. The number of digits different between two addresses is called the Hamming Distance. There may be multiple paths of the same shortest length. In the example above the message could also have travelled 1100 to 1101 to 0101 to 0111, again using three hops.

Figure 2 shows the hypercube network 10 of Figure 1 which is comprised of a variety of processing devices 12. In this example, the processing devices 12 are servers and client devices, with the client devices made up of desktop computers and laptop computers. The processing devices are connected together via a suitable network such as a wide area network such as the Internet. The processing devices all have local storage available and store data for transmission around the network. The topology of the network is that of a hypercube network, with each processing device being assigned a unique binary address that reflects the connection topology of a hypercube network.

In general, a practical implementation of a hypercube network will have many more processing devices than the eight shown in Figure 2. For example, in a supercomputer which has a very large number of processing devices connected together in a hypercube network, then n could be 5, 6 or even larger with 2 ⁿ processing devices connected together. In such a network, each processing device is connected to n neighbours and the logical binary numbering represents the underlying physical structure of the network, in that, in general, processing devices that have binary addresses that differ by one (and so are logical neighbours) will be located physically close together within the network.

In such large networks, the transmission of data around the network is an important task, as for example if n=6, then two processing devices could theoretically be located six processing devices apart, and transmission of data between those two devices will pass through five other processing devices, following the routing discussed above. It is common in such systems that data is needed from multiple processing devices for processing at a single location. Since it is possible that any processing device within the network can carry out the processing, then the selection of the processing device for that task is very important in terms of bandwidth usage within the overall hypercube network.

The hypercube network can be configured to operate so that whenever data needs to be transmitted from multiple nodes to a single node for processing, the network will determine the centre of the set of nodes. For example, a five dimensional hypercube may have a set of nodes S which are the addresses of a subset of the nodes of the hypercube graph, such that S = (10100, 11111, 00110, 00010, 10000). In this case, the set S contains the binary addresses of five labels from a five dimensional hypercube.

The network analyses the digits in each of the binary addresses of the nodes in S, to determine how many addresses have a " 1 " for each digit and how many addresses have a "0". In this example, in their first digit, two addresses have a "0" and three addresses have a "1". In their second digit, four addresses have a "0" and one address has a "1", and so on. This can be represented as follows: DigitCount(O) = (2,4,2,2,4) and DigitCount(l) =

(3,1,3,3,1). For each digit is turn, there is then determined which digit count is the highest, and the address of a node C is set to the value of the digit with the highest count as follows: HighestDigitCount = (1,0,1,1,0) (in the first digit, digitcount(l)> digit count(O), in the second digit, digitcount(0)> digit count(l) etc.) so that C = 10110.

In a more generalised hypercube scheme, where the addresses are not necessarily binary addresses, the method to calculate the central processing device to reduce data transmission remains the same, only for each digit there is an increased number of DigitCounts. For example, in a generalised hypercube network that include network addresses with digits 0, 1 and 2, then there is a count up of three sets of digit count, DigitCount(O), DigitCount(l) and DigitCount(2) and the dimension with the highest digit count is chosen as the location of the central node, for each position in the address.

Figure 3 illustrates the calculation, in a system that uses binary addresses. The set S of binary addresses 14 is used to generate a new binary address 16 of node C. The node with C as the address is the centre node, relative to the members of the set S. This node has the minimum total distance from the set S of nodes. The total distance from node C to the nodes in S can be determined by summing the lowest digit count for each digit in the addresses as follows : DigitCount(O) = (2,4,2,2,4), DigitCount(l) = (3,1,3,3,1), LowestDigitCount = (2,1,2,2,1) and TotalLowestDigitCount = 8 (2+1+2+2+1) This method can be proved by considering the contribution of each digit in turn to the total Hamming Distance. If the node C has 0 as the first digit, there are three nodes which need to transmit data from their " 1 " dimension whereas if " 1 " is chosen as the first digit then there are only two nodes that need to transmit data (because two nodes have 0 in the first digit of their addresses). For each digit in the address, if the digit value is chosen with the highest count then the contribution is minimised to the overall number of legs that data needs to be transmitted, so minimising the overall distance that data needs to travel.

With an even number of nodes, there may be the same number of nodes with a 1 as with a 0 in a particular digit. In this case either 0 or 1 can be chosen for the digit of the central node, there will be multiple nodes each of which has equally minimum total distance to the set S of nodes.

This method of operating the hypercube network of processing devices is summarised in Figure 4. The first step in the method comprises determining that a plurality of the processing devices in the network is storing data to be processed at a single processing device. This may be as a result of a command being executed, that requires data from multiple locations to be processed as a single entity. The next step in the method is the step S2 of obtaining the addresses of the plurality of processing devices storing the data to be processed. As discussed above, the processing devices will each have a unique address (which may be a binary address) that identifies each device.

The next step is step S3, which comprises determining the most common number for each digit of the addresses of the plurality of processing devices storing the data to be processed. As illustrated above with reference to Figure 3, for example, the binary addresses of the processing devices in the set of devices that are storing the data to be processed are examined to identify the most common digit in each position of the binary addresses. The nature of the hypercube network topology means that the addresses will all be the same length, and a simple count operation can be used to determine the most common digit in each position. This step is followed by step S4, which comprises generating a new address comprising the determined most common number for each digit, and by step S5, which comprises transferring the data to be processed to the processing device with the generated new address. The most common digits determined in step S3 are used to generate the new address of the "centre node", shown as node C in the example of Figure 3. The data stored by all of the processing devices in the set of devices is then transferred to this new processing device with the new address comprised of the most common digits, where the data is then processed. The process described above can be modified to take into account the amount of data being stored by the different processing devices. If this information is available then it is possible to determine the centre of a set of nodes with different amounts of data at each node. For example, a set of nodes S which are the addresses of a subset of the nodes of a hypercube graph where each node has a known quantity of data that needs to be delivered to a common location such that D defines the quantity of data held at the nodes so that Di (the ith member of D) specifies the amount of data held at the node with address Si (the ith member of S).

Figure 5 illustrates such an example where S = (10100, 11111, 00110, 00010, 10000) and D = (100, 500, 20, 1, 500). In this case S contains the addresses of five labels from a five dimensional hypercube and D contains the amount of data held at the corresponding node in S (i.e. node 10100 holds 100 items of data, node 11111 holds 500 items of data etc.)

There is then performed an analysis of the corresponding digit in each of the addresses in S to determine the total amount of data held at addresses which have the value " 1 " for that digit and to determine the total amount of data held at addresses which have the value "0" for that digit. For example, in their first digits, two addresses have a "0", these addresses have 20 and 1 items of data so the total data at addresses with a "0" in the first digit is 21. Three addresses have a "1" in their first digit, these addresses have 100, 500 and 500 items of data so the total data at addresses with a "1" in the first digit is 1100. The analysis of each digit can be represented by maintaining a DataCount set for the 0 and 1 values as follows:

DataCount(O) = ( 21, 621, 501, 600, 621) and DataCount(l) = (1100, 500, 620, 521, 500). For each digit in turn, there is then determined which data count is the highest, and the address of the node C is set to the value of the digit with the highest count so that the HighestDataCount = (1,0,1,0,0) (in the first digit, DataCount(l) > DataCount(O), in the second digit, DataCount(O) > DataCount (1) etc. Therefore C = 10100 and the node with C as the address is the "centre" node for this set of nodes and data amounts. Sending all data D from nodes S to this node incurs the minimum total data transmission.

The total amount of data that is transmitted from the nodes in the set S to the node C can be determined by summing the lowest DataCount for each digit in the addresses as follows: DataCount(O) = (21, 621, 501, 600, 621), DataCount(l) = (1100, 500, 620, 521, 500), LowestDataCount = (21, 500, 501, 521, 621) and TotalLowestDataCount = 2043

(21+500+501+521+621).

This methodology can be proved by considering the contribution of each digit in turn to the total data transmitted distance. If the node C has a "0" as the first digit, there are 1100 units of data which need to be sent from the " 1 " dimension whereas if we choose " 1 " as the first digit then there are only 21 units of data that need to be transmitted. For each digit in the address, if the digit value with the most data is chosen then the contribution to the overall amount of data that needs to be transmitted is minimised from the other addresses. With any number of nodes, the total amount of data at addresses with "1" as with "0" in a particular digit may be the same. In this case either "0" or "1" can be chosen for the digit of the central node, there will be multiple nodes each of which has equally minimum total data

transmission from the set S of nodes. The methodology illustrated in Figure 5 and described above is effectively a weighted determination of the most common digit in the binary addresses of the processing devices that make up the set S. Figure 6 illustrates the hypercube network of Figure 1 showing that four of the nodes in the network are storing data for processing at a single location. The set of nodes S = (000, 001, 100, 111) and set of data D = (15, 6, 10, 20). Therefore,

DataCount(O) = (21, 31, 25) and DataCount(l) = (30, 20, 26) meaning that the binary address of the centre node C = 101. In this example, this would be the processing device of the hypercube network that carries out the processing of the data.