WO2022150995A1

WO2022150995A1 - Supercomputer architecture implementation method

Info

Publication number: WO2022150995A1
Application number: PCT/CN2021/071347
Authority: WO
Inventors: 王志平
Original assignee: 王志平
Priority date: 2021-01-13
Filing date: 2021-01-13
Publication date: 2022-07-21

Abstract

A supercomputer architecture implementation method, which comprises steps of: constructing a process exchange unit connecting various nodes within a supercomputer (S1); establishing a routing path between a first node for sending a process and a second node for receiving a process in the supercomputer by means of the process exchange unit (S2); and running a process sent by the first node by means of the second node (S3). The present method does not need to rely on a TCP/IP protocol, is able to produce a more sensible and more efficient task load scheduling effect for a huge load, and thereby ensures the timeliness of a process.

Description

A kind of supercomputer architecture realization method

technical field

The present invention relates to the field of computer technology, in particular to a method for realizing supercomputer architecture.

Background technique

"Supercomputer" (or "mainframe computer") can be understood as a super large computer processing architecture formed by multiple servers through a network. The characteristic of "supercomputer" is that it can carry more tasks and can process, and has "super" powerful computing power is one of the important characteristics of "supercomputer". The topology of the "supercomputer" that physically constitutes the "supercomputer" and the task deployment strategy on the network are the key factors that determine the performance and efficiency of the "supercomputer". Different "supercomputer" manufacturers have different topological structures and task deployment strategies. There are some differences, but the interconnection between all computing nodes (ie, monolithic servers) that make up a "supercomputer" follows the TCP/IP protocol.

With the development of supercomputers, the load of the computer is increasing, and the existing network topology structure usually distributes the tasks of one area to the appropriate area according to the load situation of each area through the scheduling server.

However, the existing supercomputers rely on the TCP/IP protocol to build the entire system architecture and rely on software to implement task scheduling, which will inevitably lead to synchronous communication (especially real-time synchronization) between different processes in the entire "supercomputer" system. The software is implemented on the TCP/IP protocol socket (or other process communication protocol). For short messages (maybe most short messages with only a few or a dozen bytes), the synchronization efficiency will become very low; with At the same time, the synchronous communication based on socket (or other process communication protocol) will also cause the failure of synchronous communication to be too poor (ie, the delay is too large, which is not conducive to real-time control). Therefore, there is a need for a supercomputer architecture that does not need to rely on the TCP/IP protocol, can have a more reasonable and efficient task load scheduling effect when dealing with huge loads, and ensure the timeliness of processes (tasks).

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a supercomputer architecture implementation method, which does not need to rely on the TCP/IP protocol, and can have a more reasonable and efficient task load scheduling effect when dealing with huge loads, thereby ensuring the timeliness of the process. sex.

The technical scheme provided by the present invention is as follows:

The present invention provides a method for realizing a supercomputer architecture, comprising the steps of:

Build a process exchange unit that connects each node in the supercomputer;

Establish a routing path between a first node for sending a process and a second node for receiving a process in the supercomputer by using the process exchange unit;

The process sent by the first node is run through the second node.

Specifically, the process switching unit is abbreviated as PRU (Process Routing Unit), which has the function of "routing", but does not need too large and complex network communication protocols as support; from the perspective of hardware implementation, PRU no longer exists. The specific concept of "protocol" is to directly reduce the required functions to the various instructions in the PRU instruction set. As any node in the "supercomputer", when using the PRU to complete its role in the entire system, it only needs to send corresponding instructions to the PRU to complete it easily. Therefore, PRU is a special bus controller that realizes high throughput, high efficiency, high scalability, no software and no configuration, low power consumption, and low cost to connect all nodes in the "supercomputer" system. The first node may be referred to as node S, and the second node may be referred to as node T.

By building a process exchange unit that connects each node in the supercomputer, when a process in a node needs to be transferred to other nodes for operation, it can be transferred directly through the process exchange unit without relying on the TCP/IP protocol. Therefore, it can have a more reasonable and efficient task load scheduling effect when dealing with huge loads, which is beneficial to ensure the timeliness of the process.

Further, establishing the routing path between the first node for sending the process and the second node for receiving the process in the supercomputer through the process exchange unit specifically includes:

receiving a route establishment instruction sent by the first node;

Broadcast the route establishment instruction to other nodes of the supercomputer through the process exchange unit, and receive route establishment feedback instructions sent by other nodes correspondingly;

Selecting, by the process switching unit, the node with the smallest routing time as the second node among other nodes that send the establishment feedback instruction;

A routing path connecting the first node and the second node is established.

Specifically, the route establishment instruction is sent by "node S", and a route is established between the request and "node T", but the route establishment instruction does not specify which "node" is "node T" (in fact, as a "Node S", which does not have the ability to specify a routing target), but the PRU decides how to establish a route according to the specific load of the "node" connected to it and the value of the routing time.

In addition, after the routing path is successfully established, or after the establishment of the routing path fails, a "routing feedback" instruction will be sent back to "node S" to inform "node S" of the routing path establishment result, but the "routing feedback" instruction It will only be passed in the routing network, that is, after the command is passed to "node S", the corresponding port of "node S" will be fed back to the relevant driver process in the form of an interrupt.

The route establishment instruction includes the following parameters: the process number parameter (that is, the process number of the driver process), the RTF file number parameter (that is, the file number of the RTF file), and the kernel class number parameter (the kernel class number parameter indicates that the route to be established by the current command is Prepared for the process running on that type of kernel, the kernel type signal determines whether the relevant process can be executed in "node T", that is to say, the PRU needs to judge how to find the appropriate "node T" according to the "kernel type number" to establish a route), the kernel version number parameter (the kernel version number parameter refers to the version number of the kernel defined by the kernel category number. Only the kernel category number and the kernel version number meet the requirements to ensure that the relevant process can be correct in the "target system". run), code loading space parameter (code loading space parameter refers to the maximum storage space required when the execution code of the relevant process served by the route established by the current instruction is loaded (installed) into the "target system", "code loading space" "The meaning of the parameter is to judge whether the "target system" has enough storage resources to load the execution code of the relevant process), the data load space parameter (the data load space parameter refers to the needs of the relevant process served by the route established by the current instruction. The minimum storage amount used for variable storage, the meaning of the data loading space parameter is to judge whether the "target system" has enough storage resources for the lowest running scenario of the relevant process) and the TTR parameter (TTR is the English abbreviation of Time To Routing, that is "Route Time").

The route establishment feedback command is a feedback command for the route establishment command, and this command may be issued by the "node T" or the PRU. The route establishment feedback command includes the following parameters: the process number parameter (the process number of the driver process or 0 (if the command is issued by "node T", the parameter is the "driver process number", otherwise it must be 0)), RTF parameter ( RTF file number), TTR parameter (representing the number of PRUs experienced in the process of route establishment), route result parameter (representing the final result of route establishment, the possible results of route establishment include: "Route establishment is successful", " "Route establishment failed", if the parameter is "Route establishment failed", then you need to parse the "Error" parameter to know the reason for the "failure"), and the route property parameter (when the route result parameter shows "Route establishment succeeded" , "routing property" represents the load proportion of the "node T" directly or indirectly connected to the current port. At this time, the PRU can use this parameter to select the most suitable "node T" from many feedbacks as the best route. When the route result parameter shows "Route establishment failed", "Route property" indicates the reason for the failure of the current port routing. The specific possibilities are as follows: this port is not connected to "node T", and the hardware resources of node T are insufficient. , The load of "Node T" is too large, the hardware resources and load of "Node T" do not meet the requirements, the PRU memory is insufficient and cannot provide enough space to cache code and initial variables, the PRU hardware management resources are insufficient and cannot create RTF files, PRU memory and hardware management resources are insufficient).

Further, selecting the node with the smallest routing time as the second node among other nodes sending the establishment feedback instruction by the process switching unit further includes:

When there are multiple nodes with the smallest routing time, the node with the smallest load proportion is selected as the second node.

Further, when receiving the route establishment instruction sent by the first node, the method further includes:

Create a route and task description file for recording process information in the process exchange unit;

When establishing the routing path connecting the first node and the second node, the method further includes:

Write routing information in the routing and task description files;

The running of the process sent by the first node through the second node specifically includes:

The process sent by the first node is loaded into the second node through the routing and task description file, and the second node executes the process.

Specifically, the Routing & Task File (RTF for short) is an important file used by the PRU to establish and implement routing. This file records information related to the "process" and records the routing path information. The implementation of routing by the PRU will depend on the content (routing information) recorded in the RTF file. The RTF is only saved and managed in the PRU. For a routing path, the RTF is initially created by "node S", that is, in any routing path Before being established, "node S" needs to first create an RTF file in the PRU to which it is connected. In fact, the RTF file at this time is an empty file, and no content is recorded for the time being. After that, "node S" will send a message to the PRU to establish a route. For related instructions, the PRU will write the corresponding routing information and process information in the RTF file when establishing the routing path for the "node S".

The RTF file must contain the following main information: "Node S" or the port number connecting the PRU from the "Node S" direction to the current PRU; "Node T" or the port number connecting the PRU from the "Node S" direction to the current PRU; If the current PRU is PRU-S (that is, the process switching unit of the source end node), it needs to include the process ID of the process using this routing path in "node S"; if the current PRU is PRU-T (that is, the process switching unit of the target end end) When , it needs to include the process ID of the process using this routing path in "node T"; the total number of PRUs experienced by the current routing path, namely TTR (TTR is the English abbreviation of Time To Routing, that is, "routing time"); routing time Poke, ie RTS.

Further, after the process sent by the first node is loaded into the second node through the route and task description file, and after the second node performs the running of the process, it also includes:

When there is an instruction on the routing path, update the routing timestamp in the routing and task description files;

When the difference between the routing timestamp and the current time of the process switching unit exceeds a predetermined range, send a wake-up instruction to the first node through the process switching unit;

If the first node does not respond to the wake-up instruction, that is, the related process of "node S" has been terminated unexpectedly, the process in the second node is automatically uninstalled, and the current routing path is deleted;

If the first node responds to the wake-up command and communicates with the corresponding second node, but the second node does not respond, that is, the process related to "Node T" has been terminated unexpectedly, the first node Send uninstall process command.

Further, after the uninstalling process instruction is sent by the first node, the method further includes:

receiving the uninstallation process instruction sent by the first node;

The routing path between the first node and the second node is deleted.

Further, after the establishment of the routing path connecting the first node and the second node, the method further includes:

Receive the installation execution code sent by the first node and execute the instruction, the second node loads the process execution code, and starts running the process immediately after the loading is completed;

or;

After receiving the installation execution code and the initialization and execution instruction sent by the first node, the second node loads the process execution code and initializes variables, and starts running the process immediately after the loading is completed.

perform data interaction between the first node and the second node by means of short messages;

and / or;

The data interaction between the first node and the second node is performed by means of file messages.

Further, after running the process sent by the first node through the second node, the method further includes:

After the process finishes running, receive the no-message return instruction sent by the second node, and delete routing information in the process of transmitting the no-message return instruction;

or;

After the running of the process is completed, a message-returning instruction sent by the second node is received, and routing information is deleted in the process of transmitting the message-returning instruction.

Further, before the construction of the process exchange unit connecting each node in the supercomputer, it also includes the steps:

During software development, determine whether the user software supports dynamic deployment;

If supported, the communication between processes is encapsulated in the form of short messages and/or file messages by the compiler.

According to a method for realizing a supercomputer architecture provided by the present invention, by constructing a process exchange unit connecting each node in the supercomputer, when a process in a node needs to be transferred to other nodes for operation, the process exchange unit can be directly passed through the process exchange unit. The transfer does not need to rely on the TCP/IP protocol, so that it can have a more reasonable and efficient task load scheduling effect when dealing with huge loads, which is conducive to ensuring the timeliness of the process.

Description of drawings

The preferred embodiments will be described below in a clear and easy-to-understand manner with reference to the accompanying drawings, and the above-mentioned characteristics, technical features, advantages and implementations of the present solution will be further described.

Fig. 1 is the overall flow schematic diagram of the embodiment of the present invention;

2 is a schematic diagram of a connection between a process exchange unit and a node according to an embodiment of the present invention;

3 is another schematic diagram of a connection between a process exchange unit and a node according to an embodiment of the present invention;

4 is another schematic diagram of a connection between a process exchange unit and a node according to an embodiment of the present invention;

5 is a schematic structural diagram of a process exchange unit according to an embodiment of the present invention;

6 is a schematic structural diagram of a supercomputer architecture according to an embodiment of the present invention;

FIG. 7 is another schematic structural diagram of a supercomputer architecture according to an embodiment of the present invention.

Detailed ways

In order to more clearly describe the embodiments of the present invention or the technical solutions in the prior art, the specific embodiments of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative efforts, and obtain other implementations.

In order to keep the drawings concise, the drawings only schematically show the parts related to the present invention, and they do not represent its actual structure as a product. In addition, in order to make the drawings concise and easy to understand, in some drawings, only one of the components having the same structure or function is schematically shown, or only one of them is marked. As used herein, "one" not only means "only one", but also "more than one".

Example 1

An embodiment of the present invention, as shown in FIG. 1 , provides a method for implementing a supercomputer architecture, including the steps:

S1. Build a process exchange unit that connects each node in the supercomputer.

Specifically, the process switching unit is abbreviated as PRU (Process Routing Unit), which has the function of "routing", but does not need too large and complex network communication protocols as support; from the perspective of hardware implementation, PRU no longer exists. The specific concept of "protocol" is to directly reduce the required functions to the various instructions in the PRU instruction set. As any node in the "supercomputer", when using the PRU to complete its role in the entire system, it only needs to send the corresponding instructions to the PRU to complete it easily. Therefore, PRU is a special bus controller that realizes high throughput, high efficiency, high scalability, no software and no configuration, low power consumption, and low cost to connect all nodes in the "supercomputer" system.

S2. Establish a routing path between the first node used for sending the process and the second node used for receiving the process in the supercomputer through the process exchange unit.

The first node may be referred to as node S, and the second node may be referred to as node T.

Preferably, as shown in FIG. 2 to FIG. 4 , each port of the process exchange unit can be connected to each node of the corresponding supercomputer, so as to facilitate the transfer of processes in each node; and/or the port of the process exchange unit can be connected to other The port connection of the process exchange unit can facilitate the expansion of the supercomputer architecture by connecting the port of the process exchange unit with the ports of other process exchange units.

As shown in FIG. 5 , in this embodiment, the process switching unit has 8 ports from port 0 to port 7 inside, and these 8 ports outside the PRU can be connected to “computing nodes” or other PRUs, respectively. The three modules of PRE, EI and MI are included in the box. RRE is the abbreviation of Process Routing Engine, which is expressed as the "routing engine" independently owned by each port, which is used for PRU, that is, "process switching unit" command. Set execution/correspondence; EI is the English abbreviation of EFS Interface, which means that the module implements the function of the interface with the EFS module shown in the figure; MI is the English abbreviation of Matrix Interface, which means that the module implements the same function as the Matrix module shown in the figure. The function of the interface between them; the Matrix in the figure is a multi-layer bus controller that realizes the interconnection between all ports in the PRU.

S3. Run the process sent by the first node through the second node.

Specifically, the supercomputer architecture of this solution can be shown in Figure 6. Of course, it can also be continuously expanded on this basis. The expansion method can be shown in Figure 7, so that the architecture of the supercomputer is continuously increased.

Preferably, before constructing a process exchange unit connecting each node in the supercomputer, the steps further include:

During software development, it is judged whether the user software supports dynamic deployment; if so, the communication between processes is encapsulated in the form of short messages and/or file messages through the compiler.

This solution can achieve dynamic deployment during process deployment. The so-called "dynamic deployment" is relative to "static deployment", and "static deployment" refers to the need for user tasks to enter the internal network of the "supercomputer". Large tasks with large resources and applications with high concurrency perform task segmentation and possible network configuration for the hardware of the "supercomputer" and its internal network characteristics. Therefore, "static deployment" requires different adaptations for different "supercomputer" user programs, and the portability or compatibility is poor. "Static deployment" needs to be completed before the task is submitted. The development of very large tasks or large tasks with high concurrency becomes more complex and difficult.

The "dynamic deployment" that can be achieved in the present invention means that no matter how the user program consumes hardware resources or the concurrency nature, the user does not need to make a specific task for a specific "supercomputer" when submitting a task to the "supercomputer" The "adaptation operation" of the "supercomputer" does not require any configuration related to the internal network of the "supercomputer". After any user program is submitted to the "supercomputer", the system can automatically deploy the user program to the process accuracy. The so-called "deployment accurate to the process" refers to the sub-process generated at any time during the running process of the user program, and the system can dynamically deploy it according to the current load of the "node".

In addition, when the process deployment changes from "static deployment" to "dynamic deployment", the communication and synchronization methods between processes will change. Therefore, when the method of the present invention implements "dynamic deployment", it is required to determine whether the software supports "dynamic deployment" in the process of software development. When "dynamic deployment" is supported, the compiler will The method of inter-process communication and synchronization of "layout" is automatically encapsulated in the form of short messages and/or file messages in the PRU instruction set, so as to achieve the same process communication and synchronization effect as "static deployment" in function. Obviously, the "preparation" required by the present invention for the user program has nothing to do with the specific "supercomputer", so this "preparation" is completely independent of the "supercomputer" hardware and its internal network, and will not bring any portability issues.

Specifically, if the user program supports "dynamic deployment", then after the process is compiled, the compiler will insert a judgment branch for all programs related to process communication and process synchronization, that is, when judging the current program, the PRU-S ( That is, the source node process switching unit) or the PRU-T (target process switching unit), if it is running in the PRU-S, the same process communication and process synchronization procedures as "static deployment" are used. , and if it is run in PRU-T, the process communication and process synchronization program encapsulated by the PRU instruction is used. In the user process of "supercomputer", if the process is deployed in PRU-T to run, the process will be marked as "external process" (the process that is not deployed to other "nodes" is "local process", i.e. the default flag state).

Example 2

An embodiment of the present invention, on the basis of Embodiment 1, establishes a routing path between a first node for sending a process and a second node for receiving a process in a supercomputer by using a process exchange unit, which specifically includes:

S21. Receive a route establishment instruction sent by the first node.

S22: Broadcast the route establishment instruction to other nodes of the supercomputer through the process exchange unit, and receive the route establishment feedback instruction sent by other nodes correspondingly.

S23. The process switching unit selects the node with the smallest routing time as the second node among other nodes that send the establishment feedback instruction.

Preferably, the process switching unit selects the node with the smallest routing time as the second node among other nodes that send the establishment feedback instruction, further comprising:

S24. Establish a routing path connecting the first node and the second node.

Example 3

An embodiment of the present invention, on the basis of Embodiment 2, when receiving the route establishment instruction sent by the first node, further includes:

Create a routing and task description file for recording process information in the process exchange unit.

When establishing the routing path connecting the first node and the second node, it also includes:

Write routing information in routing and task description files.

Running the process sent by the first node through the second node specifically includes:

The process sent by the first node is loaded into the second node through the routing and task description file, and the process is run by the second node.

Example 4

In an embodiment of the present invention, on the basis of Embodiment 3, the process sent by the first node is loaded into the second node through the routing and task description file, and after the second node executes the process, the process further includes:

When there is an instruction on the routing path, update the routing timestamp in the routing and task description files; when the difference between the routing timestamp and the current time of the process switching unit exceeds a predetermined range, send a wake-up instruction to the first node through the process switching unit; If the first node does not respond to the wake-up command, that is, the related process of "node S" has been terminated unexpectedly, the process in the second node is automatically uninstalled, and the current routing path is deleted; if the first node responds to the wake-up command and corresponds to the corresponding second node The node communicates, but the second node does not respond, that is, the related process of "Node T" has been terminated unexpectedly, and the uninstallation process instruction is sent through the first node.

Preferably, after the uninstallation process instruction is sent by the first node, the method further includes: receiving the uninstallation process instruction sent by the first node; and deleting the routing path between the first node and the second node.

Example 5

An embodiment of the present invention, on the basis of any of the foregoing embodiments, after establishing a routing path connecting the first node and the second node, further includes: receiving the installation execution code sent by the first node and executing the instruction, the second node Load the process execution code, and start running the process immediately after the loading is completed; or receive the installation execution code and initialization and execution instructions sent by the first node, the second node loads the process execution code and initializes variables, and immediately after the loading is completed. Start the process running.

Preferably, after establishing the routing path connecting the first node and the second node, the method further includes: performing data interaction between the first node and the second node by means of short messages; and/or performing the first node and the second node by means of file messages Data interaction between the node and the second node.

Further preferably, after running the process sent by the first node through the second node, it also includes: after the process running ends, receiving a no-message return instruction sent by the second node, and deleting the routing information in the process of transmitting the no-message return instruction. ; or after the process ends, receive a message return instruction sent by the second node, and delete the routing information in the process of message return instruction delivery. The difference between a message return instruction and no message return instruction is in the process of instruction transmission. will carry a short message.

It should be noted that the above embodiments can be freely combined as required. The above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, several improvements and modifications can be made. It should be regarded as the protection scope of the present invention.

Claims

A method for realizing supercomputer architecture, comprising the steps of:

Build a process exchange unit that connects each node in the supercomputer;

Establish a routing path between a first node for sending a process and a second node for receiving a process in the supercomputer by using the process exchange unit;

The process sent by the first node is run through the second node.
The method for implementing a supercomputer architecture according to claim 1, wherein the process switching unit is used to establish a first node for sending a process and a second node for receiving a process in the supercomputer. Routing paths between nodes, including:

receiving a route establishment instruction sent by the first node;

Broadcast the route establishment instruction to other nodes of the supercomputer through the process exchange unit, and receive route establishment feedback instructions sent by other nodes correspondingly;

Selecting, by the process switching unit, the node with the smallest routing time as the second node among other nodes that send the establishment feedback instruction;

A routing path connecting the first node and the second node is established.
The method for implementing a supercomputer architecture according to claim 2, wherein the process switching unit selects a node with the smallest routing time among other nodes that send the establishment feedback instruction as the second node, which also includes:

When there are multiple nodes with the smallest routing time, the node with the smallest load proportion is selected as the second node.
The method for implementing a supercomputer architecture according to claim 2, wherein when receiving the route establishment instruction sent by the first node, the method further comprises:

Create a route and task description file for recording process information in the process exchange unit;

When establishing the routing path connecting the first node and the second node, the method further includes:

Write routing information in the routing and task description files;

The running of the process sent by the first node through the second node specifically includes:

The process sent by the first node is loaded into the second node through the routing and task description file, and the second node executes the process.
The method for implementing a supercomputer architecture according to claim 4, wherein the process sent by the first node is loaded into the second node through the routing and task description file, and the process is sent by the second node. After the second node runs the process, it further includes:

When there is an instruction on the routing path, update the routing timestamp in the routing and task description files;

When the difference between the routing timestamp and the current time of the process switching unit exceeds a predetermined range, send a wake-up instruction to the first node through the process switching unit;

If the first node does not respond to the wake-up instruction, automatically uninstall the process in the second node, and delete the current routing path;

If the first node responds to the wake-up instruction and communicates with the corresponding second node, but the second node does not respond, an uninstallation process instruction is sent through the first node.
The method for implementing a supercomputer architecture according to claim 5, characterized in that, after sending the uninstallation process instruction through the first node, the method further comprises:

receiving the uninstallation process instruction sent by the first node;

The routing path between the first node and the second node is deleted.
The method for implementing a supercomputer architecture according to claim 2, wherein, after the establishing a routing path connecting the first node and the second node, the method further comprises:

Receive the installation execution code sent by the first node and execute the instruction, the second node loads the process execution code, and starts running the process immediately after the loading is completed;

or;

After receiving the installation execution code and the initialization and execution instruction sent by the first node, the second node loads the process execution code and initializes variables, and starts running the process immediately after the loading is completed.
The method for implementing a supercomputer architecture according to claim 2, characterized in that after establishing the routing path connecting the first node and the second node, the method further comprises:

perform data interaction between the first node and the second node by means of short messages;

and / or;

The data interaction between the first node and the second node is performed by means of file messages.
The method for implementing a supercomputer architecture according to any one of claims 1-8, wherein after the process sent by the first node is executed by the second node, the method further comprises:

After the process finishes running, receive the no-message return instruction sent by the second node, and delete routing information in the process of transmitting the no-message return instruction;

or;

After the running of the process is completed, a message-returning instruction sent by the second node is received, and routing information is deleted in the process of transmitting the message-returning instruction.
The method for realizing a supercomputer architecture according to claim 8, characterized in that, before the process switching unit that connects each node in the supercomputer is constructed, the method further comprises the steps:

During software development, determine whether the user software supports dynamic deployment;

If supported, the communication between processes is encapsulated in the form of short messages and/or file messages by the compiler.