A server customization system and method based on super-computing and AI cluster technology

Abstract

The application discloses a kind of server customization system and method based on supercomputer and AI cluster technology, it is related to server customization technical field, the application includes information acquisition module, hardware selection module, topology customization module and database, first according to the type of each hardware currently owned, the flowability of data under the combination of different types of each hardware and different connection relations, obtain the required type of each hardware of server, then analyze the best physical topology of different supercomputer and AI cluster load ratio, build physical topology matching table, then customize the physical topology of server, and real-time monitoring the load ratio of supercomputer and AI cluster in server during running, according to monitoring result and physical topology matching table, re-customize physical topology and adjust, guarantee the adjustability of physical topology, also guarantee the core performance and the stability of operation of server.

Description

Technical Field

[0001] This invention relates to the field of server customization technology, and specifically to a server customization system and method based on supercomputing and AI cluster technology. Background Technology

[0002] With the rapid development of AI large model technology and the widespread application of artificial intelligence in scientific research, the demand for integrated computing power for supercomputing scientific computing and AI large model training has become a core development trend in the field of high-performance computing. This has placed new demands on the computing power density, interconnection capabilities, scheduling efficiency, and scenario adaptability of server clusters.

[0003] In traditional server customization systems and methods, servers adopt standardized designs, and the physical topology of the servers is set before leaving the factory. During operation, the servers run supercomputing and AI cluster tasks according to the fixed physical topology. Obviously, this kind of server customization system and method has at least the following shortcomings: 1. When setting the physical topology, traditional server customization systems and methods ignore the connection relationship between different hardware models, and cannot guarantee the adjustability of the physical topology. During the operation of the server, the physical topology structure cannot be adjusted according to the server's operating information.

[0004] 2. Traditional server customization systems and methods neglect the compatibility between the physical topology of the server and the load ratio of the supercomputing and AI clusters when designing the physical topology of the server. Furthermore, they cannot adjust the physical topology according to changes in the load of the supercomputing and AI clusters during server operation, resulting in a mismatch between the physical topology and the load ratio, increasing operation and maintenance costs, and consequently failing to guarantee the core performance and operational stability of the server. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the present invention aims to provide a server customization system and method based on supercomputing and AI cluster technology.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In the first aspect, the present invention provides a server customization system based on supercomputing and AI cluster technology, including: an information acquisition module, a hardware selection module, a topology customization module, and a database.

[0007] The information acquisition module is used to retrieve the model numbers of all currently owned hardware from the database and refer to them as the model numbers of each hardware.

[0008] The hardware selection module is used to determine the required model of each hardware component in the server based on the identifiable model number and physical topology adjustment degree of each hardware component.

[0009] The topology customization module is used to customize the physical topology of the server and monitor the supercomputing load and AI cluster load of the server in real time during server operation. It determines whether the physical topology of the server needs to be re-customized at each time and adjusts the physical topology through a programmable silicon photonics switching chip.

[0010] The database is used to store the models of all currently owned hardware and the real-time load ratio of servers during each historical supercomputing and AI cluster task.

[0011] Secondly, the present invention provides a server customization method based on supercomputing and AI cluster technology, including: S1, obtaining information: obtaining the model of each hardware currently owned from the database and referring to them as the model number of each hardware.

[0012] S2. Select Hardware: Based on the identifiable model number and physical topology adjustment of each hardware component, determine the required model number for each hardware component in the server.

[0013] S3. Custom Topology: Customize the physical topology of the server and monitor the supercomputing load and AI cluster load of the server in real time during server operation. Determine whether the physical topology of the server needs to be re-customized at each time and adjust the physical topology through a programmable silicon photonics switching chip.

[0014] The beneficial effects of this invention are as follows: 1. This invention provides a server customization system and method based on supercomputing and AI cluster technology. First, based on the models of the currently available hardware, the data flow under different combinations of hardware models and different connection relationships, the required models of each hardware component of the server are obtained. Then, the optimal physical topology for the load ratio of different supercomputing and AI clusters is analyzed, and a physical topology matching table is constructed. After that, the physical topology of the server is customized, and the load ratio of the supercomputing and AI clusters in the server is monitored in real time during operation. Based on the monitoring results and the physical topology matching table, the physical topology is re-customized and adjusted, ensuring the adjustability of the physical topology and also ensuring the core performance and operational stability of the server.

[0015] 2. This invention randomly combines the model numbers of each hardware component to obtain various hardware combinations. Connection simulation is performed on each hardware combination. Within a given hardware combination, the data flow of the server under different connection relationships is obtained. If the server exhibits good data flow under a certain connection relationship within that hardware combination, this connection relationship is designated as a "marked connection relationship." This method is used to obtain the marked connection relationships for each hardware combination. The total number of marked connection relationships for each hardware combination is counted and used as the physical topology adjustability of each hardware combination. The hardware combination with the highest physical topology adjustability is obtained, and the model numbers used by each hardware component in that combination are used as the required model numbers for each hardware structure, ensuring the adjustability of the physical topology.

[0016] 3. Based on the required models of various hardware structures, this invention obtains various available physical topologies and sets up several supercomputing-AI cluster load groups. Each supercomputing-AI cluster load group is applied sequentially to each available physical topology to simulate server operation. The matching degree between each supercomputing-AI cluster load group and each available physical topology is obtained. Among each supercomputing-AI cluster load group, the available physical topology with the highest matching degree is selected as the optimal physical topology. Based on the analysis results and the load ratio of the supercomputing and AI clusters during server operation, the physical topology is customized and adjusted to ensure the core performance and operational stability of the server. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the system structure connection of the present invention.

[0019] Figure 2 This is a schematic diagram of the implementation steps of the method of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 As shown, the present invention provides a server customization system based on supercomputing and AI cluster technology, including: an information acquisition module, a hardware selection module, a topology customization module, and a database.

[0022] The information acquisition module is connected to the hardware selection module, the hardware selection module is connected to the topology customization module, and the database is connected to both the information acquisition module and the topology customization module.

[0023] The information acquisition module is used to retrieve the model numbers of all currently owned hardware from the database and refer to them as the model numbers of each hardware.

[0024] It should be noted that each piece of hardware refers to the hardware components that make up the server, including the computing server, leaf switch, spine switch, core switch, and programmable silicon photonics switching chip, etc.

[0025] Different hardware models have different specifications. For example, leaf switches include models such as MQM9790-NS2F, MQM9700-NS2F, and QM8700-HS2R. The core specifications of different models are different. For example, the core specifications of the leaf switch model MQM9790-NS2F are 64-port 400G NDR InfiniBand, 32 OSFP ports (which can be split into 128×200G), 1U / 2U models, 25.6Tb / s switching capacity, and end-to-end latency <90ns. This example is for illustrative purposes only and is not the only limitation.

[0026] The hardware selection module is used to determine the required model of each hardware component in the server based on the identifiable model number and physical topology adjustment degree of each hardware component.

[0027] In a specific embodiment, the process of determining the required model of each hardware component in the server is as follows: randomly combine the model numbers of each hardware component to obtain each hardware combination; perform connection simulation on each hardware combination; in a certain hardware combination, obtain the data flow of the server under different connection relationships of that hardware combination; if the data flow of the server is good under a certain connection relationship of that hardware combination, then that connection relationship is called a marked connection relationship; obtain the marked connection relationships of each hardware combination in this way.

[0028] It should be noted that the hardware combination includes all the hardware components that make up the server.

[0029] It should also be noted that the connection relationship of a hardware combination refers to the connection relationship between the individual hardware components in the hardware combination.

[0030] The total number of marked connections for each hardware combination is counted and used as the physical topology adjustment degree of each hardware combination. The hardware combination with the largest physical topology adjustment degree is obtained, and the model number used by each hardware component in the hardware combination is used as the required model number for each hardware structure.

[0031] The specific process for obtaining the data flow of the server under different connection relationships of the hardware combination described above is as follows: Under a certain connection relationship of the hardware combination, the physical topology of the hardware combination is obtained according to the connection relationship of the hardware combination, and data is input into the server. The data reception status of each node in the physical topology is monitored in real time. Each node that does not receive data is called a disconnected node. The total number of disconnected nodes is counted. The data flow value of the server under the connection relationship of the hardware combination is determined based on the total number of disconnected nodes. The data flow value includes 1 and 0, which represent good data flow and poor data flow, respectively. The data flow of the server under different connection relationships of the hardware combination is obtained by this method.

[0032] It should be noted that the data reception status of a node includes both receiving data and not receiving data.

[0033] It should also be noted that each node in the physical topology represents a piece of hardware, and the data reception status of each node in the physical topology is monitored by the listening component.

[0034] Specifically, when the total number of disconnected nodes is 0, the data flow value of the server under this connection relationship of the hardware combination is 1; otherwise, the data flow value of the server under this connection relationship of the hardware combination is 0.

[0035] The topology customization module is used to customize the physical topology of the server and monitor the supercomputing load and AI cluster load of the server in real time during server operation. It determines whether the physical topology of the server needs to be re-customized at each time and adjusts the physical topology through a programmable silicon photonics switching chip.

[0036] It should be noted that the supercomputing load and AI cluster load of the server are monitored through the MPIProfiling interface and TensorFlow monitoring hooks.

[0037] It should also be noted that by replacing the traditional electrical switch core with a programmable silicon photonics switching chip, the silicon photonics chip integrates a massive number of micron-level optical switch units. Through electrical signals, it can control the pass-through / disconnection of any input optical port and any output optical port, realizing the full cross-connection reconstruction of the physical layer.

[0038] In one specific embodiment, the topology customization module includes a physical topology matching table and a customization adjustment unit.

[0039] The physical topology matching table is used to store the optimal physical topology for different load ratios.

[0040] In a specific embodiment, the construction process of the physical topology matching table is as follows: A11. According to the required model of each hardware, obtain each available physical topology and set up several supercomputing-AI cluster load groups.

[0041] It should be noted that the hardware selection module determines the hardware combination corresponding to the required model of each hardware structure, and obtains each available physical topology based on the connection relationships with good data flow in the hardware combination.

[0042] It should also be noted that the supercomputing-AI cluster load group stores the supercomputing load and the AI ​​cluster load. The load ratio of each supercomputing-AI cluster load group is different. The load ratio of the supercomputing-AI cluster load group refers to the ratio of the supercomputing load to the AI ​​cluster load in the supercomputing-AI cluster load group.

[0043] A12. In a given available physical topology, each supercomputing-AI cluster load group is sequentially applied to the available physical topology for server simulation. In a given supercomputing-AI cluster load group, during operation, the communication latency of each node in the available physical topology, the training cycle of AI cluster tasks, and the server congestion are monitored in real time. Based on the monitoring results, the matching degree between the available physical topology and the supercomputing-AI cluster load group is calculated. This method is used to obtain the matching degree between the available physical topology and each supercomputing-AI cluster load group, and these are compared. The supercomputing-AI cluster load group with the highest matching degree with the available physical topology is then identified as the optimal physical topology for the corresponding load ratio of the supercomputing-AI cluster load group.

[0044] It should be noted that the communication latency of each node in the physical topology, the training cycle of the AI ​​cluster task, and the server congestion are monitored through a PTP clock server, the server's built-in BMC chip, and a network card hardware congestion monitoring instrument.

[0045] It should also be noted that server congestion status includes "good" and "poor". "Good" congestion status means that the server is not congested, while "poor" congestion status means that the server is congested.

[0046] A13. Obtain the optimal physical topology for each load ratio according to the method in step A12, and store each load ratio and its optimal physical topology one-to-one to form a physical topology matching table.

[0047] The specific process for calculating the matching degree between the available physical topology and the supercomputing-AI cluster load group is as follows: at each time point, the node communication latency difference and the maximum communication latency are obtained based on the communication latency of each node, and the first return value of the server is determined based on the node communication latency difference and the maximum communication latency at each time point.

[0048] It should be noted that at a certain moment, the communication delays of each node are compared to obtain the maximum and minimum communication delays. The difference between the maximum and minimum communication delays is taken as the node communication delay difference. The node communication delay difference at each moment is obtained in this way.

[0049] Based on the training period and training period limit of the AI ​​cluster task, determine the second return value of the server, and based on the server congestion at each time, determine the third return value of the server.

[0050] It should be noted that the training period of the current AI cluster task on a high-performance and stable server will be obtained and used as the training period limit.

[0051] When the first, second, or third return value of the server is 0, the matching degree between the available physical topology and the supercomputing-AI cluster load group is 0. Otherwise, based on the node communication latency difference and the maximum communication latency at each time point, the average node communication latency difference and the average maximum communication latency are calculated. The average node communication latency difference, the average maximum communication latency, and the training period of the AI ​​cluster task are called server parameters. Based on the server parameters and their limits, the matching coefficient between the available physical topology and the supercomputing-AI cluster load group is obtained. The matching coefficients are sorted from largest to smallest as the matching degree is sorted from largest to smallest.

[0052] It should be noted that the difference between the server parameters and their limits is calculated and used as the matching coefficient. It is particularly important to note that when the difference is 0, the matching coefficient is the smallest value greater than 0. A difference of 0 does not mean a matching degree of 0, but only that the matching degree is the lowest. When the difference is 0, the server parameters have reached their limits. At this point, the available physical topology has the lowest matching degree with the supercomputing-AI cluster load group, not a mismatch. A matching degree of 0 represents a mismatch.

[0053] It should also be noted that the node communication latency difference limit and the maximum communication latency limit are both set by the staff according to the needs of the supercomputing task.

[0054] The process for determining the first, second, and third return values ​​of the server is as follows: The node communication delay difference and the maximum communication delay at each time point are compared with the node communication delay difference limit and the maximum communication delay limit. If there is a time when the node communication delay difference is greater than the node communication delay difference limit, or the maximum communication delay is greater than the maximum communication delay limit, then the first return value of the server is determined to be 0; otherwise, the first return value of the server is 1.

[0055] If the training period of the AI ​​cluster task is greater than the training period limit, the second return value of the server is 0; otherwise, the second return value of the server is 1.

[0056] If the server congestion is good at all times, the server's third return value is 0; otherwise, the server's third return value is 1.

[0057] The customized adjustment unit is used to customize the physical topology of the server and apply it to the server. During the operation of the server, based on the real-time monitoring of the server's supercomputing load and AI cluster load, it determines whether the server needs to re-customize the physical topology at various times and adjusts the physical topology accordingly.

[0058] In a specific embodiment, the physical topology of the customized server is determined as follows: real-time load ratios during each historical supercomputing and AI cluster task are obtained from the database; clustering is performed on time periods with the same load ratio to obtain the time periods for each load ratio; the time periods for each load ratio during each historical supercomputing and AI cluster task are obtained in this way, and these time periods are compared to obtain the time periods with the same load ratio in a group; and each group is obtained in this way.

[0059] Within each group, the durations of each time period are summed to obtain the duration of the corresponding load ratio for each group. The load ratio with the longest duration is taken as the marked load ratio. The marked load ratio is then matched with the physical topology matching table to obtain the physical topology of the server.

[0060] In another specific embodiment, the process of determining whether the server needs to re-customize the physical topology at each moment is as follows: obtain the supercomputing load and AI cluster load of the server at each moment; calculate the load ratio of the server at each moment based on the supercomputing load and AI cluster load of the server at each moment; compare the load ratio of the server at each moment; if the load ratio at a certain moment is the same as the load at the previous moment, it means that the server does not need to re-customize the physical topology at that moment; otherwise, it means that the server needs to re-customize the physical topology at that moment. This method is used to determine whether the server needs to re-customize the physical topology at each moment.

[0061] The database is used to store the models of all currently owned hardware and the real-time load ratio of servers during each historical supercomputing and AI cluster task.

[0062] Please see Figure 2 As shown, the present invention provides a server customization method based on supercomputing and AI cluster technology, including: S1, obtaining information: obtaining the model of each hardware currently owned from the database and referring to them as the model of each hardware.

[0063] S2. Select Hardware: Based on the identifiable model number and physical topology adjustment of each hardware component, determine the required model number for each hardware component in the server.

[0064] S3. Custom Topology: Customize the physical topology of the server and monitor the supercomputing load and AI cluster load of the server in real time during server operation. Determine whether the physical topology of the server needs to be re-customized at each time and adjust the physical topology through a programmable silicon photonics switching chip.

[0065] This invention first obtains the required hardware models for the server based on the available hardware models, combinations of different hardware models, and data flow under different connection relationships. Then, it analyzes the optimal physical topology for different load ratios of supercomputing and AI clusters, constructs a physical topology matching table, customizes the server's physical topology, and monitors the load ratio of supercomputing and AI clusters in the server in real time during operation. Based on the monitoring results and the physical topology matching table, the physical topology is re-customized and adjusted, ensuring the adjustability of the physical topology and also ensuring the core performance and operational stability of the server.

[0066] The examples described in this invention are not limited to the specific embodiments listed above. The examples are merely illustrative to facilitate understanding of the invention and do not constitute a limitation on the scope of protection of this invention. Any modifications, equivalent substitutions, etc., made within the spirit and principles of this invention should be included within the scope of protection.

[0067] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in this specification, they should all fall within the protection scope of the present invention.

Claims

1. A server customization system based on supercomputing and AI cluster technology, characterized in that, Includes the following modules: The information acquisition module is used to retrieve the model numbers of all currently owned hardware from the database and refer to them as the model numbers of each hardware. The hardware selection module is used to determine the required model of each piece of hardware in the server based on the identifiable model and physical topology adjustment of each piece of hardware. The topology customization module is used to customize the physical topology of the server and monitor the supercomputing load and AI cluster load of the server in real time during server operation. It determines whether the physical topology of the server needs to be re-customized at various times and adjusts the physical topology through a programmable silicon photonics switching chip. The database is used to store the models of all currently owned hardware and the real-time load ratios of servers during each historical supercomputing and AI cluster task.

2. The server customization system based on supercomputing and AI cluster technology according to claim 1, characterized in that, The specific process for determining the required models of each hardware component in the server is as follows: The various hardware models are randomly combined to obtain various hardware combinations. Connection simulation is performed on each hardware combination. In a certain hardware combination, the data flow of the server under different connection relationships of the hardware combination is obtained. If the data flow of the server is good under a certain connection relationship of the hardware combination, the connection relationship is called the marked connection relationship. The marked connection relationships of each hardware combination are obtained in this way. The total number of marked connections for each hardware combination is counted and used as the physical topology adjustment degree of each hardware combination. The hardware combination with the largest physical topology adjustment degree is obtained, and the model number used by each hardware component in the hardware combination is used as the required model number for each hardware structure.

3. A server customization system based on supercomputing and AI cluster technology according to claim 2, characterized in that, The specific process for obtaining the data flow of the server under different connection relationships of this hardware combination is as follows: Under a certain connection relationship of the hardware combination, the physical topology of the hardware combination is obtained according to the connection relationship, and data is input into the server. The data reception status of each node in the physical topology is monitored in real time. Nodes that do not receive data are called disconnected nodes. The total number of disconnected nodes is counted. Based on the total number of disconnected nodes, the data flow value of the server under the connection relationship of the hardware combination is determined. The data flow value includes 1 and 0, which represent good data flow and poor data flow, respectively. This method is used to obtain the data flow of the server under different connection relationships of the hardware combination.

4. A server customization system based on supercomputing and AI cluster technology according to claim 2, characterized in that, The topology customization module includes a physical topology matching table and a customization adjustment unit; The physical topology matching table is used to store the optimal physical topology for different load ratios; The customized adjustment unit is used to customize the physical topology of the server and apply it to the server. During the operation of the server, based on the real-time monitoring of the server's supercomputing load and AI cluster load, it determines whether the server needs to re-customize the physical topology at various times and adjusts the physical topology accordingly.

5. A server customization system based on supercomputing and AI cluster technology according to claim 4, characterized in that, The process of constructing the physical topology matching table is as follows: A11. Based on the required models of each hardware component, obtain the available physical topologies and set up several supercomputing-AI cluster load groups. A12. In a certain available physical topology, each supercomputing-AI cluster load group is applied sequentially to the available physical topology to simulate server operation. In a certain supercomputing-AI cluster load group, the communication latency of each node in the available physical topology, the training cycle of AI cluster tasks, and the congestion of the server are monitored in real time during the operation. Based on the monitoring results, the matching degree between the available physical topology and the supercomputing-AI cluster load group is calculated. The matching degree between the available physical topology and each supercomputing-AI cluster load group is obtained by this method, and they are compared. The supercomputing-AI cluster load group with the highest matching degree with the available physical topology is obtained. Then, the available physical topology is the optimal physical topology for the corresponding load ratio of the supercomputing-AI cluster load group. A13. Obtain the optimal physical topology for each load ratio according to the method in step A12, and store each load ratio and its optimal physical topology one-to-one to form a physical topology matching table.

6. A server customization system based on supercomputing and AI cluster technology according to claim 5, characterized in that, The specific process for calculating the matching degree between the available physical topology and the supercomputing-AI cluster load group is as follows: At each moment, the communication delay difference and maximum communication delay of each node are obtained based on the communication delay of each node. Based on the communication delay difference and maximum communication delay of each node at each moment, the first return value of the server is determined. Based on the training period and training period limit of the AI ​​cluster task, determine the second return value of the server, and based on the server congestion at each time, determine the third return value of the server. When the first, second, or third return value of the server is 0, the matching degree between the available physical topology and the supercomputing-AI cluster load group is 0. Otherwise, based on the node communication latency difference and the maximum communication latency at each time point, the average node communication latency difference and the average maximum communication latency are calculated. The average node communication latency difference, the average maximum communication latency, and the training period of the AI ​​cluster task are called server parameters. Based on the server parameters and their limits, the matching coefficient between the available physical topology and the supercomputing-AI cluster load group is obtained. The matching coefficients are sorted from largest to smallest as the matching degree is sorted from largest to smallest.

7. A server customization system based on supercomputing and AI cluster technology according to claim 6, characterized in that, The process for determining the server's first, second, and third return values ​​is as follows: The node communication delay difference and maximum communication delay at each time point are compared with the node communication delay difference limit and maximum communication delay limit. If there is a time when the node communication delay difference is greater than the node communication delay difference limit or the maximum communication delay is greater than the maximum communication delay limit, then the server's first return value is determined to be 0; otherwise, the server's first return value is 1. If the training period of the AI ​​cluster task is greater than the training period limit, the second return value of the server is 0; otherwise, the second return value of the server is 1. If the server congestion is good at all times, the server's third return value is 0; otherwise, the server's third return value is 1.

8. A server customization system based on supercomputing and AI cluster technology according to claim 4, characterized in that, The physical topology of the customized server is defined as follows: The system retrieves the real-time load ratio of the server during each historical supercomputing and AI cluster task from the database. It then clusters the time periods of each load ratio based on the time periods of the same load ratio. This method is used to obtain the time periods of each load ratio during each historical supercomputing and AI cluster task, and the data is compared. The time periods of the same load ratio are grouped together, and each group is obtained using this method. Within each group, the durations of each time period are summed to obtain the duration of the corresponding load ratio for each group. The load ratio with the longest duration is taken as the marked load ratio. The marked load ratio is then matched with the physical topology matching table to obtain the physical topology of the server.

9. A server customization system based on supercomputing and AI cluster technology according to claim 4, characterized in that, The specific process for determining whether the server needs to re-customize its physical topology at each time point is as follows: Obtain the supercomputing load and AI cluster load of the server at each time point. Based on the supercomputing load and AI cluster load of the server at each time point, calculate the load ratio of the server at each time point. Compare the load ratio of the server at each time point. If the load ratio at a certain time point is the same as the load at the previous time point, it means that the server does not need to re-customize the physical topology at that time point. Conversely, it means that the server needs to re-customize the physical topology at that time point. Use this method to determine whether the server needs to re-customize the physical topology at each time point.

10. A server customization method using the server customization system based on supercomputing and AI cluster technology as described in any one of claims 1-9, characterized in that, include: S1. Obtain information: Retrieve the model numbers of all currently owned hardware from the database and refer to them as the model numbers of each hardware. S2. Hardware Selection: Based on the identifiable model and physical topology adjustment of each hardware component, determine the required model of each hardware component in the server. S3. Custom Topology: Customize the physical topology of the server and monitor the supercomputing load and AI cluster load of the server in real time during server operation. Determine whether the physical topology of the server needs to be re-customized at each time and adjust the physical topology through a programmable silicon photonics switching chip.

Images