WO2022007587A1 - Switch and data processing system - Google Patents

Abstract

A switch. The switch is connected to at least two data nodes. The switch is used for respectively receiving result data, which is sent by the at least two data nodes, of a first operation of a distributed computing task, executing a second operation of the distributed computing task according to the received result data of the first operation, so as to obtain result data of the second operation, and distributing the result data of the second operation. Thus, data nodes and a switch jointly execute operation processes of a distributed computing task, thereby improving the data processing efficiency and reducing the processing delay.

Description

Switches and data processing systems technical field

The present application relates to the field of communications, and in particular, to a switch and a data processing system.

Background technique

Artificial intelligence (AI) parameter training that simulates human thought processes and intelligent behaviors (such as training, reasoning), and high performance computing (HPC) that utilizes aggregate computing power to handle computationally intensive computing tasks In such scenarios, it is often necessary to aggregate the same data of multiple data nodes (data nodes are usually servers). For example, the computer program all_reduce() for aggregation processing used in artificial intelligence parameter training, and the computer program MPI_all_reduce() for aggregation processing of message passing interface (MPI) used in high-performance computing. The above aggregation processing is performed by an independent aggregation node, where the aggregation node can be an independent server. However, as the number of data nodes increases, the amount of data to be aggregated generated by the data nodes also increases. The amount of data that the data node needs to transmit to the aggregation node also increases, and the aggregation node needs to process more data, so that the data transmission bandwidth in the entire system cannot meet the requirements, resulting in an increase in the delay of data aggregation processing. problem. Therefore, how to reduce the processing delay of data aggregation has become an urgent technical problem to be solved.

SUMMARY OF THE INVENTION

The present application provides a switch, device and system for data processing, so as to provide a low-latency data processing method and improve the efficiency of data processing.

In a first aspect, the present application provides a switch, the switch is connected to at least two data nodes, and the at least two data nodes respectively perform a first operation of a distributed computing task; the switch is used for receiving data sent by the at least two data nodes. the result data of the first operation, perform the second operation of the distributed computing task according to the received result data of the first operation, obtain the result data of the second operation, and distribute the result data of the second operation to the above at least two data nodes . It can be seen from the above description that the switch and the data node can jointly complete the distributed computing task, that is, the operation of the distributed computing task can be performed during the data transmission process of the switch, so as to avoid the problem of low efficiency caused by the second operation performed by a separate node , thereby improving the efficiency of data processing. In addition, since the switch completes the second operation of the distributed computing task during data transmission, there is no need to deploy a separate node, which reduces the cost of the system.

As a possible implementation manner, the switch includes a processing unit and at least two ports, each port is connected to a data node, and each port is used to receive the result data of the first operation sent by the connected data node, and send the first operation The result data of the operation is forwarded to the processing unit. Thereby, the second operation of executing the distributed computing task by the processing unit of the switch is implemented, thereby reducing the delay of data processing.

As another possible implementation manner, before forwarding the result data of the first operation to the processing unit, each port is further configured to perform the third operation of the distributed computing task on the result data of the first operation. That is to say, each port can also perform the third operation of the distributed computing task before forwarding the result data of the first operation to the processing unit, thereby accelerating the speed of data processing.

As another possible implementation manner, the distributed computing tasks include distributed artificial intelligence computing tasks or distributed high-performance computing tasks or distributed graphics computing tasks or distributed cloud computing tasks.

As another possible implementation manner, the second operation or the third operation of the distributed computing task includes an operation of aggregating data of the same type.

As another possible implementation manner, the switch is an access switch or an aggregation switch.

As another possible implementation manner, the processing unit is further configured to send an operation command to the at least two ports, where the operation command is used to instruct the at least two ports to respectively perform the third operation of the distributed computing task. The processing unit can instruct the ports connected to the data nodes that perform the distributed computing task to perform the third operation of the distributed computing task respectively through the operation command, so as to realize the purpose of the operation of the switch performing the distributed computing task during the data transmission process, improving the performance of the distributed computing task. Efficiency of data processing.

As another possible implementation manner, the processing unit sends an operation command to the at least two ports connected to the data nodes performing distributed computing tasks through a first loop, where the first loop includes at least For two ports, the order of the first loop indicates the order in which the above at least two ports receive or execute operation commands. The transmission of the operation command and the result data of the operation command is realized through the first loop, so that the influence of the data aggregation processing on other types of data processing processes can be avoided. Moreover, the bandwidth of the first loop can be configured according to service requirements, thereby ensuring the performance of data processing.

As another possible implementation manner, after the first port in the first loop performs the third operation according to the operation command, the result data of the third operation and the operation command are forwarded to the adjacent subsequent port, until The sequentially last port in the first loop sends all the result data of the third operation to the processing unit. Through the data communication path of the first loop, the operation commands can be executed by each port in the first loop in turn, and the result data of the first operation is sent to the adjacent subsequent ports until the first loop The last port in the above completes the processing process of the operation command, so that each port completes the operation of the distributed computing task according to the operation command, and accelerates the process of data processing.

As another possible implementation manner, the processing unit is further configured to, before sending the operation command, receive the packet headers respectively sent by the at least two ports connected to the data node executing the distributed computing task, each packet header Including data type and message serial number; establishing operation table entry according to the message header, wherein the operation table entry records the data to be processed and processed data in each type of data; sending operation command according to the operation table entry. In this way, the processing unit instructs the operation commands of the at least two ports according to the processing conditions of each type of data, and then each port completes the processing process of the distributed computing task according to the operation commands.

As another possible implementation manner, the packet header further includes a port identifier, and the operation table entry is further used to record the port identifier corresponding to each data to be processed in each type of data.

As another possible implementation manner, the switch includes at least one of the first loops.

As another possible implementation manner, the switch is further configured to establish a second loop, and distribute result data of the second operation through the second loop. The switch may further include a second loop for distributing the result data of the second operation, so as to avoid the influence of distributing the result data of the second operation on other types of operations. In the second loop, each port can sequentially acquire the result data of the second operation. For example, in the distributed computing task of aggregated data, the result data of the second operation refers to the aggregated results of all the same type of data, and the switch can send the above-mentioned aggregated results of all the same type of data to the execution distribution through the second loop. The data nodes of the distributed computing task are connected to the port, and then the aggregation results of all the same type of data are sent to the data nodes that execute the distributed computing task through each port, so as to realize the operation of executing the distributed computing task during the data transmission process and accelerate the data. processing efficiency.

As another possible implementation, the first loop and the second loop may be the same.

As another possible implementation manner, the first loop and the second loop may also be different.

In a second aspect, the present application provides a method for data processing. The method is executed by a switch, the switch is connected to at least two data nodes, and each data node is used to perform a first operation of a distributed computing task, and the specific data processing process Including: the switch respectively receiving the result data of the first operation sent by the at least two data nodes; performing the second operation of the distributed computing task according to the received result data of the first operation, obtaining the result data of the second operation, and distributing the second operation of the distributed computing task. The result data of the second operation. It can be seen from the above content that the switch performs the operation of distributed computing tasks during the data transmission process, which improves the efficiency of data processing.

In a possible implementation manner, the switch includes a processing unit and at least two ports, each port is connected to a data node, and each port is used to receive the result data of the first operation sent by the connected data node, and send the first operation The result data of an operation is forwarded to the processing unit. Thereby, the second operation of executing the distributed computing task by the processing unit of the switch is implemented, thereby reducing the delay of data processing.

As another possible implementation manner, before forwarding the result data of the first operation to the processing unit, each port is further configured to perform the third operation of the distributed computing task on the result data of the first operation. That is to say, each port can also perform the third operation of the distributed computing task before forwarding the result data of the first operation to the processing unit, thereby accelerating the speed of data processing.

As another possible implementation manner, the distributed computing tasks include distributed artificial intelligence computing tasks or distributed high-performance computing tasks or distributed graphics computing tasks or distributed cloud computing tasks.

As another possible implementation manner, the second operation or the third operation of the distributed computing task includes an operation of aggregating data of the same type.

As another possible implementation manner, the switch is an access switch or an aggregation switch.

As another possible implementation manner, the processing unit is further configured to send an operation command to the at least two ports, where the operation command is used to instruct the at least two ports to respectively perform the third operation of the distributed computing task. The processing unit can instruct the ports connected to the data nodes that perform the distributed computing task to perform the third operation of the distributed computing task respectively through the operation command, so as to realize the purpose of the operation of the switch performing the distributed computing task during the data transmission process, improving the performance of the distributed computing task. Efficiency of data processing.

As another possible implementation manner, the processing unit sends an operation command to the at least two ports connected to the data nodes performing distributed computing tasks through a first loop, where the first loop includes at least For two ports, the order of the first loop indicates the order in which the above at least two ports receive or execute operation commands. The transmission of the operation command and the result data of the operation command is realized through the first loop, so that the influence of the data aggregation processing on other types of data processing processes can be avoided. Moreover, the bandwidth of the first loop can be configured according to service requirements, thereby ensuring the performance of data processing.

As another possible implementation manner, after the first port in the first loop performs the third operation according to the operation command, the result data of the third operation and the operation command are forwarded to the adjacent subsequent port, until The sequentially last port in the first loop sends all the result data of the third operation to the processing unit. Through the data communication path of the first loop, the operation commands can be executed by each port in the first loop in turn, and the result data of the first operation is sent to the adjacent subsequent ports until the first loop The last port in the above completes the processing process of the operation command, so that each port completes the operation of the distributed computing task according to the operation command, and accelerates the process of data processing.

As another possible implementation manner, the processing unit is further configured to, before sending the operation command, receive the packet headers respectively sent by the at least two ports connected to the data node executing the distributed computing task, each packet header Including data type and message serial number; establishing operation table entry according to the message header, wherein the operation table entry records the data to be processed and processed data in each type of data; sending operation command according to the operation table entry. In this way, the processing unit instructs the operation commands of the at least two ports according to the processing conditions of each type of data, and then each port completes the processing process of the distributed computing task according to the operation commands.

As another possible implementation manner, the packet header further includes a port identifier, and the operation table entry is further used to record the port identifier corresponding to each data to be processed in each type of data.

As another possible implementation manner, the switch includes at least one of the first loops.

As another possible implementation manner, the switch is further configured to establish a second loop, and distribute result data of the second operation through the second loop. The switch may further include a second loop for distributing the result data of the second operation, so as to avoid the influence of distributing the result data of the second operation on other types of operations. In the second loop, each port can sequentially acquire the result data of the second operation. For example, in the distributed computing task of aggregated data, the result data of the second operation refers to the aggregated results of all the same type of data, and the switch can send the above-mentioned aggregated results of all the same type of data to the execution distribution through the second loop. The data nodes of the distributed computing task are connected to the port, and then the aggregation results of all the same type of data are sent to the data nodes that execute the distributed computing task through each port, so as to realize the operation of executing the distributed computing task during the data transmission process and accelerate the data. processing efficiency.

As another possible implementation, the first loop and the second loop may be the same.

As another possible implementation manner, the first loop and the second loop may also be different.

In a third aspect, the present application provides an apparatus for data processing, the apparatus comprising various modules for executing the data processing method in the second aspect or any possible implementation manner of the second aspect.

In a fourth aspect, the present application provides a system for data processing, the system includes a switching network and at least two data nodes connected to the switching network, wherein the at least two data nodes are used to perform the first step of the distributed computing task respectively. An operation: The switching network includes at least one switch, and the at least one switch is configured to respectively receive the result data of the first operation sent by the at least two data nodes, and perform the second operation of the distributed computing task according to the received result data of the first operation. operation, obtain the result data of the second operation, and distribute the result data of the second operation.

As a possible implementation manner, each switch includes a first processor and at least two ports, each port is respectively used for connecting with a data node that performs distributed computing tasks, and the first processor and each port are respectively used for The operation steps of the method described in any possible implementation manner of the second aspect are performed.

In a fifth aspect, the present application provides a computer-readable storage medium, where a command is stored in the computer-readable storage medium, which, when executed on a computer, causes the computer to execute the methods described in the above aspects.

In a sixth aspect, the present application provides a computer program product comprising commands that, when run on a computer, cause the computer to perform the methods described in the above aspects.

On the basis of the implementation manners provided by the above aspects, the present application may further combine to provide more implementation manners.

Description of drawings

1 is a schematic diagram of a polymerization process provided by an embodiment of the present application;

FIG. 2 is a schematic structural diagram of a data processing system 100 provided by the present application;

FIG. 3 is a schematic structural diagram of a switch according to an embodiment of the present application;

4 is a schematic flowchart of a data processing method provided by an embodiment of the present application;

FIG. 5 is a schematic structural diagram of a switch 500 according to an embodiment of the present application;

FIG. 6 is a schematic structural diagram of another switch 600 according to an embodiment of the present application.

detailed description

In order to solve the problem of high data processing delay in the traditional technology, this application proposes a data processing method. The computing tasks are jointly performed by a data node and a switch connected to the data node, so as to improve the efficiency of data processing. For the convenience of description, also The computing tasks performed jointly by the data nodes and switches can be called distributed computing tasks. The data node may be a node in the form of a computing device (for example, a server), or a node in a virtualized form such as a virtual machine or a container. In this case, the virtual machine or container may be deployed on at least one computing device (for example, a server). ), each computing device is connected to a switch. Distributed computing tasks include distributed artificial intelligence (AI) computing tasks or distributed high-performance computing (HPC) tasks or distributed graphics computing (graphic computing) tasks or distributed cloud computing tasks or other Computational tasks that can be processed by distributed computing. Among them, the distributed cloud computing task refers to that in artificial intelligence or high-performance computing or image computing or other scenarios, the computing tasks in the scenario are jointly performed by data nodes and switches in the form of virtual machines or containers. Optionally, in addition to data nodes that can be implemented in a virtualized form, in a distributed cloud computing task, switches can also be implemented in a virtualized form.

For ease of description, the operation of the distributed computing task performed by the data node may also be referred to as the first operation, and the operation of the distributed computing task performed by the switch may be referred to as the second operation.

For example, in artificial intelligence, high-performance computing, graphics computing and other application scenarios, some computing tasks can use data nodes and switches connecting data nodes to jointly participate in the process of data processing. Taking the distributed computing task of data aggregation as an example, aggregation Refers to the operation of accumulating similar data. For example, in an artificial intelligence or high-performance computing scenario, a data node can generate the same kind of data. In this case, the operation of the data node generating the same kind of data can also be called the first operation of a distributed computing task. Similar data includes similar parameters in algorithms used in artificial intelligence scenarios, or similar data generated by computing-intensive tasks. Specific similar data can be set according to application scenarios and business requirements; switches can perform data aggregation operations on similar data generated by data nodes. Then, an aggregation result of the same type of data is obtained. At this time, the operation of performing data aggregation on the same type of data generated by the data node by the switch may also be referred to as a second operation.

1 is a schematic diagram of an aggregation operation provided by this application. As shown in the figure, data node A, data node B and data node C generate three types of parameters, for example, data node A generates A0, A1 and A2, data node A generates A0, A1 and A2 B generates B0, B1 and B2, and data node C generates C0, C1 and C2. Assuming that the data with the same mantissa is the same type of data, the aggregation results obtained by the switch after the aggregation operation include A0+B0+C0, A1+B1+C1, and A2+B2+C2.

Next, taking the distributed computing task as data aggregation processing as an example, the technical solutions to be protected by the present application will be described in detail with reference to the accompanying drawings.

2 is a schematic structural diagram of a data processing system 100 provided by an embodiment of the present application. As shown in the figure, the system 100 includes a switching network 10 and a data node 20 , and the data node 20 is connected to the switching network 10 .

The data node 20 is used to generate data to be processed by distributed computing tasks, for example, artificial intelligence parameter training and/or data of the same type to be aggregated for data-intensive computing tasks in high-performance computing scenarios, and generate the data in the form of packets. The form sends the same kind of data to be aggregated to the switch connected to it. For example, as shown in FIG. 2 , the data node 20 includes six data nodes, wherein the data nodes 201 to 203 are connected to the switch 102 , the data nodes 204 to 206 are connected to the switch 103 , and the switch 102 and the switch 103 are connected through the switch 101 , Further, the communication connection between the data nodes is realized, and each data node can send the generated data of the same type to be aggregated to the switching network 10, and the switch in the switching network 10 performs the data aggregation operation.

The switching network 10 is used for implementing data transmission and performing operations of distributed computing tasks in the system 100 (eg, performing data aggregation operations in a distributed computing task of aggregating data). The switching network 10 includes at least one switch. As shown in the figure, the present application takes the switch network 10 including three switches as an example, the switch 101 may also be referred to as an aggregation switch, and the switches 102 and 103 may also be referred to as access switches. The access switch is used to connect the data nodes 20 and perform the operation of distributed computing tasks of the data nodes 20 connected to it; the aggregation switch is used to realize the data transmission and distributed computing tasks of the data nodes connected by different access switches. operate.

Optionally, in the system 100 shown in FIG. 2 , only one switch may be set in the switching network 10 to implement data transmission and aggregation processing of the data nodes 201 to 206 .

It is worth noting that due to the limited number of ports in the switch, as the number of switching nodes in the system increases, a single switch may not be able to meet the system networking requirements. In this case, the accessible data in the system can be expanded by increasing the number of switches. the number of nodes. Therefore, in specific implementation, the structure of the switching network and the number of switches can be set according to service requirements, and the present application does not limit the number and networking mode of the switches in the switching network 10 . For ease of description, the following embodiments of the present application take the switching network 10 shown in FIG. 2 as an example for description.

Each switch in FIG. 2 can implement data aggregation processing in the process of transmitting data of data nodes. Further, referring to FIG. 3, FIG. 3 is a schematic structural diagram of a switch according to an embodiment of the application. As shown in the figure, the switch includes a processing unit 110, a plurality of ports (for example, ports 1201 to 1222), and a crossover network (crossbar)130.

Among them, the processing unit 110 is configured to perform a data aggregation operation. Optionally, the processing unit 110 is further configured to instruct the port to perform data aggregation processing. Further, the processing unit 110 further includes an aggregation result cache 111 , a calculation unit 112 , a command generation module 113 and a packet header management module 114 . The aggregation result cache 111 is used to store the aggregation results of the same type of data. In specific implementation, the memory in the switch or the cache of the processor in the switch can be used to realize the function of the aggregation result cache 111 . The computing unit 112 is configured to perform an aggregation operation on the aggregation results sent by each port, and manage operation table items (for aggregation operations, it may specifically be an aggregation table item), including generating, updating, and deleting operation table items, wherein the operation table items It is used to record the data to be processed and the processed data in each type of data. The command generating unit 113 is configured to determine the same type of data to be aggregated according to the operation table entry, generate an operation command (for a data aggregation operation, it may specifically be an aggregation command), and send the aggregation command to the first port of the aggregation loop, The port searches whether there is the same type of data to be aggregated according to the data stored in the input buffer of the port, and performs the aggregation operation. The message header management unit 114 is used to parse the message header sent by each port, so as to send the message sequence number in the message header to the command generation unit 114, and the command generation unit 114 generates the aggregation command according to the message sequence number and the aggregation table entry. . In addition, the packet header management unit 114 is also used for connecting to the cross-connect network 130, which may also be referred to as a cross-connect matrix, and is used for implementing the transmission of the packet header between the processing unit 110 and each port.

Multiple ports in the switch are respectively used to connect data nodes, each port can be connected to a data node, and each port includes a computing unit and a memory, for example, port 1201 includes a computing unit 12011 and a memory 12012 . The computing unit is configured to parse the message sent by the data node to obtain message header and payload data. Optionally, the computing unit is further configured to perform the aggregation operation according to the aggregation command sent by the processing unit 110 . In addition, the memory of each port can be further divided into an input buffer and an output buffer (not shown in the figure) according to different types of stored data. The input buffer is used to store the message sent by the data node connected to the port. The message includes a message header and payload data, and the payload data includes the same type of data to be aggregated. The output cache is used to receive and store the aggregation results of all the same type of data sent by the processing unit 110 after the system 100 completes the aggregation operation of all the same type of data to be processed, so as to send all the same type of data to the data node connected to the port. Aggregate result of class data.

It should be noted that the number of ports in the switch varies according to different products produced by manufacturers, and the present application does not limit the number of ports included in the switch.

FIG. 3 also shows two data transmission loops: an aggregation loop and a distribution loop. For convenience of description, the aggregation loop can also be referred to as the first loop, and the distribution loop can be referred to as the second loop.

The aggregation loop is a set of ports connected to data nodes participating in distributed computing tasks, including at least two ports sorted according to preset rules, and the sorting of ports in the aggregation loop is used to indicate that the at least two ports receive Or the order in which aggregate commands are executed. The aggregation command generated by the processing unit 110 can be sequentially transmitted from the first port of the aggregation loop to the last port, and each port can perform a corresponding aggregation operation according to the aggregation command and whether the same type of data to be aggregated is stored in the port, and aggregate the data. The result and the aggregation command are sent to the next port in the loop adjacent to the port, ..., and so on, until the last port in the aggregation loop completes the processing of the aggregation command, and the aggregation result of the aggregation command is sent to The processing unit further completes the processing of an aggregation command in the aggregation loop. As shown in FIG. 3, the aggregation loop includes: processing unit 110-port 1201-port 1202-...-port 1211-port 1212-...-port 1221-port 1222-processing unit 110, forming a closed loop, aggregating commands It can be executed by each port of the aggregation ring, and the final result data is transmitted to the processing unit 110 by the last port in the aggregation ring. The port 1201 connected to the processing unit 110 may also be called the first port of the aggregation loop, and is used to receive the aggregation command sent by the processing unit 110 . Aggregate commands are processed by port 1201, port 1202, port 1211, port 1212, . Among them, port 1202, port 1211, port 1212, and port 1221 can also be referred to as link ports of the aggregation ring, and each port can receive the aggregation command sent by its adjacent previous port in the aggregation ring and the corresponding The adjacent preceding port performs the aggregation operation according to the aggregation result of the aggregation command, and based on the received aggregation command and the adjacent preceding port according to the aggregation result of the aggregation command. For example, port 1202 can receive the aggregation command sent by port 1201 and the aggregation result of the aggregation command by port 1201. When port 1201 does not include the same type of data to be aggregated, port 1201 directly sends the aggregation command to port 1202. It can be understood that the aggregation result of the aggregation command by port 1201 is empty or none; when the aggregation port 1201 includes the same kind of data to be aggregated, in addition to sending the aggregation command to port 1202, port 1201 will also perform the operation on port 1201. The aggregation results are sent to ports 1202, ..., and so on. In the aggregation loop, each port receives the aggregation command in turn, performs aggregation operations according to the aggregation command, and combines the aggregation command and the aggregation performed by the port according to the aggregation command. The result is sent to the next port in the aggregation ring that is adjacent to it. Finally, the final aggregation result of the aggregation command is transmitted to the processing unit 110 by the last port in the aggregation loop (also referred to as a tail port). For example, the port 1222 is the last port in the order of the aggregation loop, and the port 1222 transmits the final aggregation result to the processing unit 110, and then the processing unit 110 determines whether the aggregation processing of all similar data is completed.

Optionally, the switch may include at least one aggregation loop, and each aggregation loop includes at least two ports sorted according to preset rules.

A distribution loop is a collection of ports connected to data nodes participating in distributed computing tasks, including at least two ports sorted according to preset rules, and the sorting of ports is used to indicate ports connected to data nodes participating in distributed computing tasks The order of receiving the aggregated results of all the same kind of data, so that the data nodes connected to the above ports can obtain the aggregated results of all the same kind of data, and then complete other operations of the distributed computing task. For example, as shown in FIG. 3, the distribution loop includes processing unit 110-port 1201-port 1202-...-port 1211, then the processing unit 110 can send the aggregation results of all the same data to the distribution loop through the above-mentioned distribution loop each port.

Specifically, similar to the aggregation loop, the central unit 110 can also send the aggregation results of all similar data to the first port (for example, port 1201) of the distribution loop, and then the first port sends the aggregation results of all similar data To subsequent ports adjacent to the first port (eg, port 1202), . . . , and so on, each port in the distribution loop can obtain the Aggregate results. In addition, after each port receives the aggregated results of all the same data, it can store the aggregated results of all the same types of data in the memory of the port, specifically in the output cache of the memory.

Optionally, after receiving the aggregation results of all similar data, the last port in the distribution loop may also send a notification message to the central unit, where the notification message is used to instruct the ports in the distribution loop to obtain the aggregation results of all similar data. condition.

Optionally, the switch includes at least one distribution loop, and each distribution loop includes at least two ports sorted according to preset rules.

Optionally, the aggregation loop and the distribution loop can be the same loop, that is, the aggregation loop is used to transmit aggregation commands and port aggregation results in the data aggregation process, and is also used to transmit the aggregation results of the aggregation commands and ports involved in distributed computing. The port to which the data node is connected sends the aggregated result of all the first type of data.

Optionally, the aggregation loop and the distribution loop can also be different loops.

As a possible embodiment, the adjacent ports in the aggregation loop and the distribution loop, as well as the connection between the ports and the processing units, may be physical connections (eg, conductive traces) in a printed circuit board (PCB). connect.

It is worth noting that the number of ports included in the aggregation loop and the distribution loop can be configured according to service requirements, and the data transmission paths are transmitted one by one along the ports included in the aggregation loop or the distribution loop.

The data processing method provided by the application is described in detail below with reference to FIG. 4 . As shown in the figure, the method is described by taking a distributed computing task as the processing process of data aggregation as an example. In addition, for the convenience of description, the data to be aggregated is referred to as For the first type of data, the first port, the second port and the third port are respectively the ports connected to the data nodes performing distributed computing tasks, and the first port, the second port and the third port constitute an aggregation loop, and the first port, the second port and the third port constitute an aggregation loop. One port is the first port of the aggregation loop, and the third port is the last port of the aggregation loop. Specifically, the method includes:

S301. The processing unit receives the first packet header sent by the first port.

S302. The processing unit receives the second packet header sent by the second port.

S303. The processing unit receives the third packet header sent by the third port.

In a distributed computing task, the same type of data to be aggregated has an associated packet column number, and each data can be sent from the data node to the switch using one packet. Each packet includes a packet header and static payload data. Specifically, after the port parses the packet to obtain the packet header, it can send the packet header to the processing unit through the cross-connect network. Each packet header includes the packet sequence number, and the packet sequence number is used to instruct the data node connected to the port to send the packet. The sequence number of the packet, each packet carries at least one piece of data to be aggregated. Optionally, the packet header further includes a data type, where the data type is used to indicate the type of data to be processed. In specific implementation, the generation rule of the packet header may be determined by the data node and then notified to the switch, or may be determined by the switch and then notified to the data node, which is not limited in this application.

For example, a fixed identification bit can be set in the specified field of the message sequence number. For example, in Table 1, the first field is the sequence number, and the field 2 is used to indicate the data type. When the port receives the message sequence number of the message, the second field is 1. , it means that the message associated with the message sequence number includes the first type of data with the data type 1, and the aggregation operation can be performed on the data with the data type 1 in the data processing process.

Optionally, the packet header further includes a third field for indicating the offset bit. Offset bit, used to indicate the total number of data of the same type to be aggregated on the same port. For example, when the sequence number of the received packet is 3 in the third field, it means that the total number of data to be aggregated on the port is 3.

Optionally, the packet header may further include field 4 for indicating the port identifier. The port identifier is used to indicate the identifier of the port that sends the packet header to the processing unit, and the port identifier may be represented by numbers and/or letters. Optionally, when receiving the packet header sent by the port, the processing unit may separately record the identifier of the port that sends the packet header.

Table 1 is an example of a packet header

field 1	field 2	field 3	field 4
serial number	data category	offset bit	Port ID

S304 (optionally): The processing unit checks the reliability of the packet sequence numbers in the respective packet headers respectively.

After the processing unit receives the packet headers sent by the port, it can perform reliability verification on the packet serial numbers included in the respective packet headers, and the reliability verification method can be any one of the following methods:

Manner 1: The processing unit may check the reliability of the packet serial number according to a preset rule.

The data node and the processing unit may pre-agreed a generation rule for the message sequence number, which may also be called a preset rule, and each message sequence number is a globally unique identifier. Optionally, each packet carries a first type of data, that is, the packet sequence number can uniquely identify a first type of data. The processing unit can check the validity of each packet serial number according to the preset rule. Specifically, the processing unit may pre-store a preset message sequence number table, where the preset message sequence number table is used to record the set of all message sequence numbers generated according to the preset rules, and the processing unit may store the preset message sequence number table in the preset message sequence number table. If there is a message sequence number to be queried in the preset message sequence number table, the result of the reliability check of the message sequence number is considered to be passed; otherwise, the message sequence number is considered to be reliable. The result of the sex check is failed.

In a second manner, the processing unit may calculate the validity of the packet sequence number according to a preset rule.

The packet sequence number may be a random number or an identifier generated according to a preset rule, and is used to globally uniquely identify the sequence number of a packet. For example, when the packet sequence number is a random number generated by a hash algorithm and obtained by encryption using an encryption algorithm, the processing unit can decrypt the algorithm, determine the decrypted packet sequence number according to the hash algorithm, and determine the decrypted packet sequence number. Whether the message sequence number of the message is within the pre-agreed range of message sequence numbers, if it is within the pre-agreed range of message sequence numbers, it is considered that the reliability check result of the message sequence number is passed; if it is not within the pre-agreed range of message sequence numbers within, it is considered that the reliability check result of the packet number 1 is not passed. Optionally, the packet sequence number may also be generated by using a custom algorithm or a general algorithm other than the hash algorithm, which is not limited in this application.

By verifying the reliability of the packet sequence number, the validity of the packet header can be determined before the aggregation operation is performed, thereby avoiding the problem of data errors caused by aggregating non-typed data, and improving the accuracy of distributed computing tasks.

S305. The processing unit generates an aggregation entry according to the sequence numbers of each packet.

The processing unit can generate an aggregation entry according to the packet sequence number in the received packet header, and the aggregation entry records the data to be processed and the processed data in each type of data, that is, in the data aggregation processing, the aggregation table The item is used to indicate the aggregation status of the first type of data, including the packet sequence number and the aggregation status of each packet sequence number, where the aggregation status is used to indicate the aggregation status of the first type of data associated with each packet sequence number, and the aggregation status It includes any one of "not aggregated", "aggregated", and "not aggregated, and the header has not been received". Optionally, the aggregation entry may also include a port identifier associated with the packet sequence number. Optionally, the aggregation entry may further include the data type and offset bit associated with the packet sequence number.

Exemplarily, Table 2 is a summary result of a processing unit receiving packet headers provided by an embodiment of the present application. As shown in the table, the processing unit can know from the packet headers received by each port that the ports whose port identifier is 1 are to be aggregated. There are 3 data with the data type 1, and the received packets with the serial numbers 1 and 2; the port with the port ID of 2 has a total of 4 data with the data type 1 to be aggregated, and the received packets have the serial numbers 1 and 2. 3 packets; the port with the port ID of 3 has a total of 2 data types of 1 data to be aggregated, and received packets with packet sequence numbers 1 and 2.

Table 2 A summary result of the header received by a processing unit

The processing unit can determine the message sequence numbers of all the first-type data to be aggregated according to the offset bits of the message headers in Table 2, and the aggregation state of the first-type data corresponding to each message sequence number, and then generate according to the above determination results. Aggregate entry indicating the aggregation of the first type of data. For example, according to Table 2, it can be known that the data nodes connected to the port with the port identifier 1 generate a total of 3 message sequence numbers with the data category 1, and the processing unit has received the message sequence numbers sent by the port with the port identifier 1 as 1 and 1. The packet header of 2, the processing unit has not obtained the packet header with the packet sequence number of 3; the data node connected with the port ID of 2 generates a total of 4 packet sequence numbers with the data category of 1, and the processing unit has received the port ID of 2. The packet headers with the packet sequence numbers 1 and 3 sent by the port, the processing unit does not obtain the packet headers with the packet sequence numbers 2 and 4; the data node connected to the port with the port ID of 3 generates a total of 2 data types of 1 The processing unit has obtained the packet headers with the packet sequence numbers 1 and 2. At this time, as shown in Table 3, the processing unit can first determine the packet sequence numbers of all the first type of data to be aggregated and the port identifiers associated with each packet sequence number according to the above situation, and further identify the aggregation of each packet sequence number. state. For example, the port ID is 1 and the aggregation status of the packet sequence number 1 is not "un-aggregated", and the port ID is 1 and the aggregation status of the packet sequence number 1 is not "un-aggregated and the packet header has not been received".

Table 3 An example of an aggregate table entry

Optionally, in addition to identifying the aggregation state in a literal form as shown in Table 3, the aggregation state may also be identified in any form such as numbers or letters or a combination of data and letters.

The processing unit can learn the aggregation state of the first type of data to be aggregated by generating the aggregation entry as shown in Table 3. Further, the processing unit can generate an aggregation command according to the aggregation entry, and the aggregation command is used to instruct the port according to the aggregation command. Perform data aggregation operations.

S306. The processing unit determines the packet sequence number of the first type of data to be aggregated according to the aggregation entry, and generates an aggregation command.

After determining the packet sequence numbers of all the first-type data to be aggregated and the aggregation state of each packet sequence number, the processing unit can generate an aggregation command based on the packet sequence numbers and port identifiers associated with the unaggregated data. It includes at least one packet sequence number of the first type of data to be aggregated.

Specifically, the processing unit may generate an aggregation command according to a filtering rule, and the filtering rule is used to filter the packet sequence numbers of the first type of data to be aggregated included in the aggregation command, which specifically includes any one of the following methods:

Manner 1: In a polling manner, the packet sequence number of at least one type of data to be aggregated is determined according to the size of the packet sequence number.

Specifically, one or more packet sequence numbers of the first type of data to be aggregated may be selected from all the packet sequence numbers of the first type of data to be aggregated in a polling manner and according to the size of the packet sequence numbers.

Manner 2: Determine at least one packet sequence number of the first type of data to be aggregated according to the priority mode.

The first type of data to be aggregated may also carry a priority identifier, which is carried in the message, and the priority is used to identify the priority of the first type of data associated with it. The first type of data generated by each data node is important data in the aggregated data, and the priority of the first data can be marked as high. Correspondingly, the message sent by the data node also carries information indicating the priority. The processing unit may select one or more pieces of data to be aggregated from the packet sequence numbers of all the first type of data to be aggregated according to the priority of the first type of data to be aggregated.

Manner 3: Select at least one packet sequence number of the first type of data to be aggregated according to the status of the received packet headers.

In addition to filtering the packet sequence numbers in the packet sequence numbers of all the first type of data to be aggregated in the above two methods, the processing unit can also determine the received packet sequence numbers first, and then use the method in the received packet sequence numbers. The method of the first or the second mode selects at least one packet sequence number of the first data.

Further, the processing unit can generate only one aggregation command according to the aggregation table entry, and the aggregation command includes the packet sequence numbers of all the first-type data to be aggregated filtered in any of the above methods; it can also generate multiple aggregation commands, Each aggregation command includes a packet sequence number of the first type of data to be aggregated; multiple aggregation commands can also be generated, and each aggregation command includes a packet sequence number of part of the first type of data to be aggregated. For ease of description, the processing unit only generates one aggregation command, and the command includes the packet sequence numbers of all the first type of data to be aggregated filtered in any of the foregoing manners for description.

Optionally, the aggregation command further includes a port identifier associated with the packet sequence number of the first type of data to be aggregated.

S307. The processing unit sends an aggregation command to the first port.

The processing unit may use the aggregation loop to send the aggregation command, and if the first port is the first port of the aggregation loop, the processing unit sends the aggregation command to the first port. That is to say, the processing unit directly sends the aggregation command to the first port of the aggregation ring. After the first port completes the processing of the aggregation command, it sends the aggregation command to the subsequent ports adjacent to the first port in the aggregation ring. command, and then the port completes the aggregation processing of the port according to the aggregation result of the first port and the aggregation command. After the previous port performs the aggregation operation according to the aggregation command, the result data of the aggregation operation and the aggregation command are forwarded to the adjacent port. The last port in the aggregation loop sends all the result data of the aggregation command to the processing unit until the last port in the order in the aggregation loop. For the specific process, refer to step S308 to step S310.

S308. When the first port includes the packet sequence number of the first type of data to be aggregated, the first port performs an aggregation operation, and sends an aggregation command and an aggregation result of the first port to the second port.

After the first port receives the message from the data node connected to it, it parses the message to obtain the message header and static payload data, sends the message header to the processing unit through the cross-connect network, and combines the message header with the static payload data. The dead load data is stored to the memory of the first port. When the first port receives the aggregation command, the processing of the aggregation command is performed according to the packet sequence number of the first type of data to be aggregated in the aggregation command. Specifically, the first port can first determine whether the memory of the first port includes the packet sequence number of the first type of data to be aggregated; then, determine the static load data associated with the packet sequence number according to the packet sequence number; Aggregate operations are performed on the data.

Optionally, the first port may also first determine whether to include the identifier of the first port according to the port identifier associated with the packet sequence number in the aggregation command; then, determine whether the memory of the first port includes the packet of the first type of data to be aggregated. sequence number; then determine the payload data associated with the message sequence number according to the message sequence number; finally, perform an aggregation operation according to the payload data. For the convenience of description, the operation performed by each port in the aggregation ring according to the aggregation command may also be referred to as the third operation of the distributed computing task. Correspondingly, each port performs the third operation to obtain result data, which may also be It is called the result data of the third operation.

After the first port executes the aggregation command, it will send the aggregation command and the aggregation result of the first port to the port (for example, the second port) adjacent to the first port in the aggregation loop, and the second port will continue according to the aggregation. The aggregation operation is performed on the command and the aggregation result of the first port.

Illustratively, the aggregation command generated with reference to the aggregation table entry shown in Table 3 includes the packet sequence numbers 1 and 2 in the port with the port identifier 1, and the packet sequence numbers 1 and 2 in the port with the port identifier 2. The packet sequence number of 3, and the packet sequence numbers of 1 and 2 in the port with the port ID of 3 as an example, when the first port receives the aggregation command, it will aggregate the packet sequence numbers associated with 1 and 2. The static load data obtained by the first port can also be called the aggregation result obtained by the first port according to the aggregation command, or the aggregation result of the first port, or the result data of the first port executing the aggregation command. Or referred to as the result data of the third operation performed by the first port.

Optionally, when the second port does not include the packet sequence number of the first type of data to be aggregated in the aggregation command, the first port can directly send the aggregation command to the second port. The aggregate result is zero or empty.

S309. When the second port includes the packet sequence number of the first type of data to be aggregated, the second port performs an aggregation operation, and sends an aggregation command and an aggregation result of the second port to the third port.

Similar to step S308, after receiving the aggregation command, the second port may also search for matching static payload data in the memory of the second port according to the packet sequence number of the first type of data to be aggregated in the aggregation command. Specifically, the second port may first determine whether the memory of the second port includes the message sequence number of the first type of data to be aggregated; then, determine the first data to be aggregated according to the message sequence number, that is, the message sequence number is associated with the static load data; perform aggregation operations based on the static load data. Wherein, when the second port performs the aggregation operation, it needs to first determine whether the first port sends the aggregation result generated by the first port according to the aggregation command. The aggregation operation is performed on the first type of data to be aggregated stored in the storage device to obtain the aggregation result of the second port, that is, when the second port needs to perform the aggregation operation on the basis of the aggregation result of the first port.

Illustratively, the aggregation command generated with reference to the aggregation table entry shown in Table 3 includes the packet sequence numbers 1 and 2 in the port whose port identifier is 1, and the packet sequence number in the port whose port identifier is 2 is 1. and 3, and the packet sequence numbers 1 and 2 in the port with the port ID 3 as an example, when the first port aggregates the first type of data associated with the packet numbers 1 and 2 according to the aggregation command When the aggregation result of the first port is obtained, the above-mentioned aggregation result and the aggregation command are sent to the second port. Accordingly, the second port will execute the above-mentioned aggregation result according to the above-mentioned aggregation result and the aggregation command. One type of data aggregation obtains the aggregation result of the second port. At this time, the aggregation result of the second port includes the first type of data associated with the packet sequence numbers 1 and 2 in port 1, and the packet sequence numbers 1 and 3 in port 2. The aggregated result of the associated first-class data. For ease of description, the aggregation result obtained by the second port according to the aggregation command may also be referred to as the aggregation result of the second port, or the aggregation result obtained by the second port executing the aggregation command and the result data obtained by the second port executing the third operation.

Optionally, when the second port does not include the packet sequence number of the first type of data to be aggregated by the aggregation command, the second port can directly send the aggregation command and the aggregation result of the first port to the third port. The aggregated result of the second port can be considered to be zero or empty.

It is worth noting that, if the first port does not include the packet sequence number of the first type of data to be aggregated by the aggregation command, the second port can directly aggregate the command to the third port, that is, in this case, the first Neither the one port nor the second port includes the packet sequence number of the first type of data to be aggregated by the aggregation command, and the aggregation result of the first port and the aggregation result of the second port are both zero or empty.

S310. When the third port includes the packet sequence number of the first type of data to be aggregated, perform an aggregation operation, and send the aggregation result of the third port to the processing unit.

The third port is the last port in the aggregation ring, that is, the third port is the last port in the aggregation ring in sequence. Illustratively, port 1222 as shown in FIG. 3 . Similar to the above step S309, the third port will also determine whether there is a packet sequence number of the first type of data to be aggregated indicated in the aggregation command in the memory of the third port, and whether the data in the aggregation loop adjacent to the third port exists. The aggregation result of the previous port performs the aggregation operation, and sends the result data of the aggregation operation to the processing unit.

It is worth noting that the third port performs the aggregation operation according to the first type of data to be aggregated stored in the third port and the aggregation result of the second loop to obtain the aggregation result, which can be called the aggregation of the third port. As a result, either the aggregation result obtained by the second port executing the aggregation command, or the result data of the third port executing the aggregation command.

Optionally, each port in the aggregation ring does not store the aggregation result of the first type of data that the port performs aggregation according to the aggregation command. When the processing of the aggregation command is completed, the aggregation result of the aggregation command is sent to the aggregation ring and the The port is adjacent to the rear port.

S311. When the processing unit determines that the aggregation operation of all the first type of data has not been completed, the processing unit generates a new aggregation command according to the aggregation table item, repeats the operations of steps S306 to S310, and determines according to the aggregation results of at least two aggregation commands Aggregate result of all first-class data.

Further, when the processing unit obtains the aggregation result of the last port in the aggregation loop in step S310, it can update the aggregation state of the message sequence number in Table 3, and judge whether the aggregation of all the first type data has been completed according to the updated result. Aggregation operation, if the aggregation operation of all the first type of data has not been completed, you can refer to the above steps S306 to S310 to generate a new aggregation command again, and each port in the aggregation loop performs the aggregation operation according to the new aggregation command, and then by The third port sends the aggregation result of the new aggregation command to the processing unit, and then the processing unit obtains the aggregation result of all the first type of data according to the aggregation results of the multiple aggregation commands. That is, when there are aggregation results obtained by multiple aggregation commands, the processing unit may perform the aggregation operation again on the aggregation results of the multiple aggregation commands, thereby obtaining the aggregation results of all the first-type data.

The above steps S301 to S311 can also be referred to as a data aggregation process. After completing the aggregation operation of all the first type of data, the processing unit can send the aggregation results of all data through the distribution loop to and participate in distributed computing through the data distribution process. The port to which the data nodes of the task are connected, and then all the first type of data is sent to the data nodes participating in the distributed computing task through the above port, so that the data nodes participating in the distributed computing task continue to complete other operations of the distributed computing. For the specific process, please refer to the description of steps S312 to S313.

S312. When the aggregation of all the first type data has been completed, the processing unit sends the aggregation result of all the first type data to the first port through the distribution loop.

S313. The first port sends the aggregation result of all the first type of data to the second port.

S314. The second port sends the aggregation result of all the first type of data to the third port.

The distribution loop is a path in the switch for sending the aggregated results of all the first type of data, and is a data transmission loop formed by at least two ports sorted according to preset rules.

For example, the switch shown in FIG. 3 includes two distribution loops. Distribution loop 1 is processing unit 110-port 1201-port 1202...-port 1211, and distribution loop 2 is processing unit 110-port 1222-port 1221- ... - port 1212. When the processing unit determines that the aggregation of all the data of the first type has been completed, the results of all the data of the first type can be distributed to the ports connected to the data nodes participating in the distributed computing through the distribution loop, and then transmitted to the data nodes participating in the distributed computing. data node.

S315. (Optionally) the processing unit clears the aggregation command table entry and the aggregation result cache.

After the processing unit completes the distribution of the aggregated results including all the first type of data, the processing unit may clear the aggregated result cache and delete the aggregated command entry, thereby freeing the storage space of the processing unit.

It can be seen from the above description of the data aggregation process and the aggregation data distribution process that the aggregation method provided by the present application can directly perform the data aggregation operation by the switch during the data transmission process, avoiding the occupation caused by the aggregation operation performed by the dedicated aggregation node in the traditional technology. Problems such as network resources, low transmission rate, and prolonged processing time have improved the efficiency of aggregation processing. In addition, the processing unit and each port in the switch can perform aggregation operations on the data to be aggregated in a distributed manner, avoiding the performance bottleneck problem caused by the aggregation operation performed by a single subject, and further reducing the latency of aggregation processing. Moreover, since the use of independent equipment to perform aggregation operations is avoided, the number of nodes in the system is reduced, and the system cost is reduced. On the other hand, by performing aggregation processing and distribution processing through the aggregation loop and the distribution loop, respectively, to avoid occupying the transmission bandwidth of other types of data, the transmission bandwidth of distributed computing can be greatly improved. In addition, in the data processing process, the aggregation results of the first type of data are only stored in the aggregation result cache of the processing unit, and the port does not need to cache the aggregation results of some types of data during the data processing process, and only completes the aggregation of all the same types of data. After the operation, it is necessary to store the aggregated results of all similar data, which greatly reduces the capacity requirement of cached data in the port.

As a possible embodiment, in the data processing method shown in FIG. 4, in addition to using the aggregation loop and the distribution loop to transmit data, the crossover network 130 can also be used to directly implement data transmission between the processing unit and each port. In this case, The cross-connect network 130 is not only used for realizing the transmission of packet headers between the processing unit and each port, but also for transmitting the aggregation command generated by the processing unit and the processing result of each port executing the aggregation command. The above implementation manner can also realize the process of data aggregation realized by the switch during data transmission, thereby avoiding the problems of long time and low efficiency caused by the aggregation operation performed by a single aggregation node in the traditional technology.

As a possible embodiment, in addition to the method shown in FIG. 4 , in the data processing method provided by the present application, the distributed computing task can also be performed only by the processing unit of the switch, that is, each port obtains data when After the node sends the packet carrying the first type of data to be aggregated and the packet header, the packet is sent to the processing unit, which parses the packet, obtains the packet sequence number in the packet header, and executes the execution according to the packet sequence number. Aggregate operation. Optionally, each port can also complete the packet parsing process, and send the packet header and the static payload data to the processing unit respectively, and then the processing unit performs the aggregation operation. The above process can also achieve the purpose of performing aggregation operations by the switch during the data transmission process, thereby improving the efficiency of data processing.

Through the above description, the switch 500 provided by the present application can complete the distributed computing task together with the data nodes, so that the switch can complete the operation of the distributed computing task during the data transmission process, which improves the efficiency and speed of data processing. Moreover, aggregation processing and distribution processing are respectively performed through the aggregation loop and the distribution loop, so as to avoid occupying the transmission bandwidth of other types of data, which can greatly improve the transmission bandwidth of distributed computing.

It is worth noting that, for the purpose of simple description, the above method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the described action sequence, and secondly, Those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the present application.

Other reasonable step combinations that those skilled in the art can think of based on the above description also fall within the protection scope of the present application. Secondly, those skilled in the art should also be familiar with that, the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the present application.

The data processing method provided according to the present application is described in detail above with reference to FIGS. 1 to 4 , and the data processing switch provided according to the present application will be described below with reference to FIGS. 5 to 6 .

FIG. 5 is a schematic structural diagram of a switch 500 provided by the present application. As shown in the figure, the switch 500 is used to connect at least two data nodes, and the at least two data nodes are used to respectively perform a first operation of a distributed computing task. , the switch 500 includes a first processing unit 501, wherein

a first processing unit 501, configured to receive the result data of the first operation sent by the at least two data nodes; perform the second operation of the distributed computing task according to the received result data of the first operation, Obtaining result data of the second operation; distributing the result data of the second operation.

Optionally, the shown switch 500 further includes at least two ports, each port is connected to a data node, and each port includes a receiving unit 502 and a sending unit 503, wherein,

a receiving unit 502, configured to receive the result data of the first operation sent by the connected data node;

The sending unit 503 is configured to forward the result data of the first operation to the first processing unit 501 .

Optionally, each port further includes a second processing unit 504, configured to perform all operations on the result data of the first operation before the sending unit 503 forwards the result data of the first operation to the first processing unit 501. The third operation of the distributed computing task is described.

Optionally, the distributed computing tasks include distributed artificial intelligence computing tasks or distributed high-performance computing tasks or distributed graphics computing tasks.

Optionally, the second operation or the third operation of the distributed computing task includes an operation of aggregating data of the same type.

Optionally, the switch is an access switch or an aggregation switch.

Optionally, the first processing unit 501 is further configured to send an operation command to the at least two ports, where the operation command is used to instruct the second processor 504 of the at least two ports to perform the third operation respectively .

Optionally, the at least two ports are sorted according to a preset rule to form a first loop, and the order of the first loop indicates an order in which the at least two ports receive or execute the operation command.

Optionally, after the port in the first loop performs the third operation according to the operation command, the result data of the third operation and the operation command are forwarded to the adjacent subsequent port. port until the last port in the first loop in sequence sends all the result data of the third operation to the first processing unit 501 .

Optionally, the first processing unit 501 is further configured to receive packet headers respectively sent by the at least two ports before sending the operation command, and each packet header includes a data type and a packet sequence number; The message header establishes an operation table entry, and the operation table entry records the data to be processed and the processed data in each type of data; the operation command is sent according to the operation table entry.

Optionally, the packet header further includes a port identifier, and the operation table entry is further used to record the port identifier corresponding to each data to be processed in each type of data.

Optionally, the switch includes at least one of the first loops.

Optionally, the switch is further configured to establish a second loop, and the first processing unit 501 is further configured to distribute result data of the second operation through the second loop.

Optionally, the switch includes at least one of the second loops.

It should be understood that the first processing unit 501 and the second processing unit 504 in this embodiment of the present application may be implemented by an application-specific integrated circuit (ASIC), respectively, or a programmable logic device (PLD) To achieve, the above-mentioned PLD can be a complex program logic device (complex programmable logical device, CPLD), field-programmable gate array (field-programmable gate array, FPGA), general array logic (generic array logic, GAL) or any combination thereof. When the data processing method shown in FIG. 4 can also be implemented by software, the first processing unit 501 , the second processing unit 504 and their respective modules can also be software modules.

The switch 500 according to the embodiments of the present application may correspond to executing the methods described in the embodiments of the present application, and the above and other operations and/or functions of the various units in the switch 500 are respectively in order to implement the corresponding processes of the respective methods in FIG. 4 , For brevity, details are not repeated here.

Through the above description, the switch 500 provided by the present application can complete the distributed computing task together with the data nodes, so that the switch can complete the operation of the distributed computing task during the data transmission process, which improves the efficiency and speed of data processing. Moreover, aggregation processing and distribution processing are respectively performed through the aggregation loop and the distribution loop, so as to avoid occupying the transmission bandwidth of other types of data, which can greatly improve the transmission bandwidth of distributed computing.

FIG. 6 is a schematic structural diagram of another switch 600 provided by the present application. As shown in the figure, the switch 600 includes a first processor 601 and at least two ports 602, wherein each port 602 is respectively used for participating in distributed The data nodes of the computing tasks are connected through the network 603, wherein,

A first processor 601, configured to respectively receive the result data of the first operation sent by the at least two data nodes, and perform the second operation of the distributed computing task according to the received result data of the first operation , obtain the result data of the second operation, and distribute the result data of the second operation.

It should be understood that in this embodiment of the present application, the first processor 601 may be a CPU, and the processor 601 may also be other general-purpose processors, digital signal processors (digital signal processing, DSP), application specific integrated circuits (ASICs) , Field Programmable Gate Array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general purpose processor may be a microprocessor or any conventional processor or the like.

The network 603 may be a bus, and the bus may include a power bus, a control bus, a status signal bus, and the like in addition to a data bus.

Optionally, the first processor 601 may be configured to implement the functions of the computing unit 112 , the command generating unit 13 , and the message header management unit 114 in the processing unit 110 shown in FIG. 2 , which will not be repeated here for brevity.

Optionally, the first processor 601 further includes a memory (not shown in the figure), and the memory is used to provide commands and data to the first processor 601, so that the first processor can perform the operations of the method shown in FIG. 4 . step. The memory may include read-only memory and random access memory, and the memory may also include non-volatile random access memory.

Optionally, a memory may also be included outside the first processor 601 to provide commands and data to the first processor 601, so that the first processor may execute the operation steps of the method shown in FIG. 4 .

Optionally, each port 602 includes a second processor 6021 and a memory 6022, wherein the second processor 6021 may also be a CPU, and the processor 6021 may also be other general-purpose processors, digital signal processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general purpose processor may be a microprocessor or any conventional processor or the like.

Each port 602 may be used to implement the operation steps of the method performed by the first port or the second port or the third port in the method shown in FIG. 4 , which will not be repeated here for brevity.

It should be understood that the switch 600 according to the embodiment of the present application may correspond to the switch 500 in the embodiment of the present application, and may correspond to the corresponding subject in executing the method shown in FIG. 4 in the embodiment of the present application, and the switch 600 The above-mentioned and other operations and/or functions of each module are respectively to implement the corresponding flow of each method in FIG. 4 , and are not repeated here for brevity.

Through the above description, the switch 600 provided by the present application can complete the distributed computing task together with the data node, so that the switch can complete the operation of the distributed computing task during the data transmission process, which improves the efficiency and speed of data processing. Moreover, aggregation processing and distribution processing are respectively performed through the aggregation loop and the distribution loop, so as to avoid occupying the transmission bandwidth of other types of data, which can greatly improve the transmission bandwidth of distributed computing.

The present application also provides a data processing system, the system includes a switching network and at least two data nodes connected to the switching network that respectively perform a first operation of a distributed computing task, the switching network includes at least one switch, and each switch includes The first processor and at least two ports shown in FIG. 6 are used to implement the functions of the corresponding execution body in the method shown in FIG. 4 , which are not described here for brevity. The system can realize distributed computing tasks. In the process of data transmission, switches perform operations of distributed computing tasks, thereby improving the efficiency of data processing and reducing the delay of data processing.

The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented in software, the above-described embodiments may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer commands. When the computer program commands are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer may be a general purpose computer, special purpose computer, computer network, or other programmable device. The computer commands may be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer commands may be downloaded from a website site, computer, server, or data center Transmission to another website site, computer, server, or data center is by wire (eg, coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (eg, infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, a data center, or the like that contains one or more sets of available media. The usable media may be magnetic media (eg, floppy disks, hard disks, magnetic tapes), optical media (eg, DVDs), or semiconductor media. The semiconductor medium may be a solid state drive (SSD).

The above descriptions are merely specific embodiments of the present application. Those skilled in the art can think of changes or substitutions based on the specific embodiments provided by the present application, which should all fall within the protection scope of the present application.

Claims (25)

1. A switch, characterized in that the switch is connected to at least two data nodes;

   The switch is configured to respectively receive the result data of the first operation of the distributed computing task sent by the at least two data nodes, and execute the second operation of the distributed computing task according to the received result data of the first operation. operation, obtain the result data of the second operation, and distribute the result data of the second operation.
2. The switch of claim 1, wherein the switch comprises a processing unit and at least two ports, each port being connected to a data node;

   Each port is configured to receive the result data of the first operation sent by the connected data node, and forward the result data of the first operation to the processing unit.
3. The switch according to claim 2, wherein, before forwarding the result data of the first operation to the processing unit, each port is further configured to perform all the required operations on the result data of the first operation. The third operation of the distributed computing task is described.
4. The switch according to any one of claims 1-3, wherein the distributed computing tasks include distributed artificial intelligence computing tasks, distributed high-performance computing tasks, distributed graphics computing tasks, or distributed cloud computing tasks .
5. The switch according to claim 4, wherein the second operation or the third operation of the distributed computing task comprises an operation of aggregating data of the same type.
6. The switch according to any one of claims 1-5, wherein the switch is an access switch or an aggregation switch.
7. The switch according to any one of claims 3-6, wherein the processing unit is further configured to send an operation command to the at least two ports, where the operation command is used to instruct the at least two ports to respectively The third operation is performed.
8. The switch according to claim 7, wherein the at least two ports are sorted according to a preset rule to form a first loop, and the order of the first loop instructs the at least two ports to receive or execute the The sequence of action commands.
9. The switch according to claim 8, wherein after the port in the first loop performs the third operation according to the operation command, the result data of the third operation is compared with the The operation command is forwarded to the adjacent subsequent ports until the last port in the first loop in sequence sends all the result data of the third operation to the processing unit.
10. The switch according to claim 7, wherein,

    The processing unit is further configured to receive message headers respectively sent by the at least two ports, and each message header includes a message sequence number; an operation table entry is established according to the message header, and the operation table entry records each message header. The data to be processed and the processed data in a class of data; the operation command is sent according to the operation table entry.
11. The switch according to claim 10, wherein the message header further includes a port identifier, and the operation table entry is further used to record the port identifier corresponding to each data to be processed in each type of data.
12. The switch according to any one of claims 1-11, wherein the switch is further configured to establish a second loop, and distribute result data of the second operation through the second loop.
13. A system for data processing, characterized in that the system includes a switching network and at least two data nodes connected to the switching network;

    The at least two data nodes are used to respectively perform the first operation of the distributed computing task;

    The switching network includes at least one switch, and the at least one switch is configured to respectively receive result data of the first operation sent by the at least two data nodes, and execute the result data of the first operation according to the received result data of the first operation. The second operation of the distributed computing task obtains the result data of the second operation, and distributes the result data of the second operation.
14. The system of claim 13, wherein each of the at least one switch includes a processing unit and at least two ports, each port being connected to a data node;

    Each port is configured to receive the result data of the first operation sent by the connected data node, and forward the result data of the first operation to the processing unit.
15. 15. The system according to claim 14, wherein each port is further configured to perform all operations on the result data of the first operation before forwarding the result data of the first operation to the processing unit. The third operation of the distributed computing task is described.
16. The system according to any one of claims 13 to 15, wherein the distributed computing tasks include distributed artificial intelligence computing tasks or distributed high-performance computing tasks or distributed graphics computing tasks.
17. The system according to claim 16, wherein the second operation or the third operation of the distributed computing task comprises an operation of aggregating data of the same type.
18. The system according to any one of claims 13 to 16, wherein the switch is an access switch or an aggregation switch.
19. The system according to any one of claims 15 to 18, wherein the processing unit is further configured to send an operation command to the at least two ports, where the operation command is used to instruct the at least two ports to respectively The third operation is performed.
20. The system according to claim 19, wherein the at least two ports are ordered according to a preset rule to form a first loop, and the order of the first loop indicates that the at least two ports receive or execute the Describe the sequence of operation commands.
21. The system according to claim 20, wherein after the port in the first loop performs the third operation according to the operation command, the result data of the third operation is compared with the result data of the third operation. The operation command is forwarded to the adjacent subsequent ports until the last port in the first loop in sequence sends all the result data of the third operation to the processing unit.
22. The system according to claim 19, wherein the processing unit is further configured to receive packet headers respectively sent by the at least two ports, and each packet header includes a packet sequence number; The header establishes an operation table entry, and the operation table entry records the data to be processed and the processed data in each type of data, and the processing unit sends the operation command according to the operation table entry.
23. The switch according to claim 22, wherein the packet header further includes a port identifier, and the operation table entry is further used to record the port identifier corresponding to each data to be processed in each type of data.
24. The system according to any one of claims 13 to 23, wherein the switch is further configured to establish a second loop, and distribute the result data of the second operation through the second loop.
25. A non-volatile computer-readable storage medium, characterized in that the non-volatile computer-readable storage medium includes instructions that, when executed on a computer, cause the computer to execute any one of claims 1 to 12 The operation steps performed by the switch.

Images