Method and device for evaluating power design peak of server cluster, equipment and medium

Abstract

The application discloses a kind of server cluster power design peak value evaluation method, device, equipment and medium.The method is executed by target server and power analysis system, including: by target server, when determining to enter cluster power design peak value evaluation stage, carry out server cluster time synchronization;According to the delay time, trigger period and starting time, calculate the next trigger time of GCU thread;Timing reading server clock, when determining to reach next trigger time, synchronously trigger all GCU threads in target server, to align with the GCU thread trigger time of the rest of each server in server cluster;Through power analysis system, multi-dimensional power data of each server in the process of GCU thread execution is collected, and cluster power design peak value evaluation result is obtained by analysis.Using the above technical scheme, the GCU thread of server cluster high-precision synchronous trigger can be realized, and the real power design peak value is accurately obtained.

Description

Technical Field

[0001] This invention relates to the field of server performance testing technology, and in particular to methods, apparatus, equipment and media for evaluating peak power design of server clusters. Background Technology

[0002] With the rapid development of large-scale AI (Artificial Intelligence) model training and inference services, the scale of AI server clusters continues to expand. AI server clusters typically contain dozens or even hundreds of servers, with each server usually equipped with multiple GCUs (Graphics Computing Units). Accurately assessing the Electrical Design Power Peak (EDPP) of AI server clusters under full load is a core requirement for data center power distribution planning, thermal design, and stability assurance.

[0003] In existing technologies, the GCU threads in an AI server cluster can generally be triggered through high-precision dedicated hardware synchronization, secure shell protocol broadcast command synchronization, or message queue-triggered synchronization.

[0004] However, the dedicated hardware used for high-precision hardware synchronization is expensive, requires professional maintenance personnel for configuration, and takes a long time to deploy, making it unsuitable for use in large-scale AI server clusters. Existing software synchronization methods are limited by network round-trip latency jitter, and cross-server synchronization can have a time error of 50 to 500 milliseconds, far exceeding the duration of GCU power peaks. This results in the power consumption peaks of each GCU being completely misaligned, making it impossible to assess the true power design peak of the AI ​​server cluster. Summary of the Invention

[0005] This invention provides a method, apparatus, device, and medium for evaluating the power design peak of a server cluster. It can achieve synchronous triggering of GCU threads within millisecond-level precision in a server cluster with multiple servers and multiple GCUs without relying on dedicated hardware, thereby accurately obtaining the true power design peak.

[0006] According to one aspect of the present invention, a method for evaluating peak power design in a server cluster is provided, which is executed in cooperation with a power analysis system in an AI server cluster, comprising: When the target server determines that the cluster power design peak assessment phase has been entered, server cluster time synchronization is performed. Using the target server, at the start time of the current triggering round, the next triggering time of the GCU thread is calculated based on the preset delay time, triggering period, and the start time. The target server periodically reads the server clock, determines whether the next trigger time has been reached based on the currently read server clock, and when it is determined that the next trigger time has been reached, synchronously triggers all GCU threads in the target server to align with the trigger times of the GCU threads of the other servers in the server cluster. The power analysis system collects multi-dimensional power data of each server in the server cluster during the execution of GCU threads, and analyzes the data to obtain the peak power design evaluation results of the cluster.

[0007] According to another aspect of the present invention, a power design peak assessment device for a server cluster is provided, which is executed in cooperation with a power analysis system in an AI server cluster, comprising: The time synchronization module is used to perform server cluster time synchronization through the target server when it is determined that the cluster power design peak assessment stage has been entered. The trigger time calculation module is used to calculate the next trigger time of the GCU thread at the start time of the current trigger round, based on the preset delay time, trigger period and the start time, through the target server. The synchronization triggering module is used to periodically read the server clock through the target server, determine whether the next triggering time has been reached based on the currently read server clock, and when it is determined that the next triggering time has been reached, synchronously trigger all GCU threads in the target server to align with the triggering time of the GCU threads of the other servers in the server cluster. The cluster power data acquisition and analysis module is used to collect multi-dimensional power data of each server in the server cluster during the execution of the GCU thread through the power analysis system, and analyze the data to obtain the peak power design evaluation results of the cluster.

[0008] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the power design peak assessment method for server clusters according to any embodiment of the present invention.

[0009] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the power design peak assessment method for a server cluster according to any embodiment of the present invention.

[0010] The technical solution of this invention adopts the following approach: When the target server determines that it has entered the peak power design assessment stage, it performs server cluster time synchronization; at the start of the current triggering round, it calculates the next triggering time of the GCU thread based on a preset delay time, triggering cycle, and start time; it periodically reads the server clock, determines whether the next triggering time has been reached based on the currently read server clock, and when the next triggering time is determined, it synchronously triggers all GCU threads in the target server to align with the triggering times of the GCU threads of the other servers in the server cluster; and it uses a power analysis system to collect multi-dimensional power data from each server in the server cluster during the execution of the GCU threads and analyzes the data to obtain the peak power design assessment results of the cluster. This approach enables high-precision time synchronization of the server cluster without relying on dedicated hardware, ensuring high alignment of the triggering times of the GCU threads of all servers in the cluster. This allows the peak power consumption of each GCU to be completely superimposed on the time axis, truly reflecting the peak power design of the cluster, and thus enabling accurate assessment of the peak power design of the server cluster.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0013] Figure 1 This is a flowchart of a peak power design evaluation method for a server cluster according to Embodiment 1 of the present invention; Figure 2 This is a time synchronization relationship topology diagram provided by an embodiment of the present invention; Figure 3 This is a schematic diagram of the timeline for GCU synchronous triggering provided by an embodiment of the present invention; Figure 4 This is a flowchart of another peak power design evaluation method for server clusters provided in Embodiment 2 of the present invention; Figure 5 This is a flowchart of another method for evaluating the peak power design of a server cluster according to Embodiment 3 of the present invention; Figure 6 This is a schematic diagram of synchronous wake-up of a single server GCU thread according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of a power design peak assessment device for a server cluster according to Embodiment 4 of the present invention; Figure 8 This is a schematic diagram of the structure of an electronic device that implements the peak power design assessment method for server clusters according to embodiments of the present invention. Detailed Implementation

[0014] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0015] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0016] Example 1 Figure 1 This is a flowchart of a method for evaluating the peak power design of a server cluster according to Embodiment 1 of the present invention. This embodiment is applicable to AI server clusters with multiple servers and multiple GCUs, where full-load stress tests are triggered synchronously to obtain the actual peak power design. This method can be executed by a peak power design evaluation device for the server cluster. This peak power design evaluation device for the server cluster can be implemented in hardware and / or software, and can generally be configured in a server or power analysis system with data processing capabilities.

[0017] Optionally, the server cluster mentioned in this invention may refer to an AI server cluster. A single server cluster may include multiple servers, and a single server may be equipped with at least one GCU. When collecting the power design peak of the server cluster, it is necessary to synchronously trigger the GCU threads in each GCU at the same time, so that the instantaneous power consumption peaks of each GCU are completely superimposed on the time axis to form the real power design peak, so that the data center can further realize power distribution planning, heat dissipation design and stability assurance based on the real power design peak.

[0018] Optionally, both the server cluster and the power analysis system are deployed inside the data center. The data center serves as the carrier for the server cluster and the power analysis system. The server cluster can provide large-scale data computing support when handling large model training and inference tasks in the data center.

[0019] like Figure 1 As shown, the method includes: S110. When the target server determines that the cluster power design peak assessment stage has been entered, server cluster time synchronization is performed.

[0020] It is understandable that only when all GCUs in the server cluster start full-load stress testing at the same time can the instantaneous power consumption peaks of each GCU be fully superimposed on the time axis. If there is a deviation in the start time of each GCU, the power consumption peaks will be staggered on the time axis, resulting in a severely underestimation of the evaluation results. This will further lead to insufficient design margin in the power supply and distribution system, posing a safety risk.

[0021] Optionally, the target server can be any server in the server cluster. During the peak power design assessment phase of the cluster, all servers in the server cluster execute the peak power design assessment method for the server cluster described in this embodiment as the target server.

[0022] Optionally, peak power design can refer to the maximum power demand of the cluster as a whole when all GCUs simultaneously reach their instantaneous peak power consumption under full load operation.

[0023] Optionally, the peak power design assessment of a cluster can refer to the peak power design assessment of a single server cluster, or it can refer to the peak power design assessment of a large-scale server cluster composed of multiple server clusters. The peak power design assessment method for server clusters described in this embodiment is applicable to both of the above assessment scenarios.

[0024] Optionally, server cluster time synchronization can refer to the server cluster synchronizing with a public network time protocol (NTP) server, as well as synchronizing the time of all servers within the server cluster, thereby solving the clock deviation problem across servers. Whether it is a large-scale server cluster or a single server cluster, the system clocks of each server maintain high-precision alignment.

[0025] Specifically, when the target server determines that the cluster power design peak assessment phase has begun, server cluster time synchronization may be performed, which may include: When the target server determines that it has entered the peak power design assessment phase, it obtains the master server election result based on the local Internet Protocol address and the address lexicographical order of the server cluster, and performs any of the following: If the target server is determined to be the master server, then time synchronization is performed with the public network time protocol server, and time synchronization service is provided to the slave servers in the server cluster. If it is determined that the target server is not selected as the master server, it will act as a slave server and use a clock step calibration mechanism to synchronize time with the master server in the server cluster.

[0026] Optionally, the Internet Protocol address is a unique identifier assigned to each server in the server cluster, and the local Internet Protocol address is the Internet Protocol address of the target server; the address lexicographical order of the server cluster can refer to the arrangement of the Internet Protocol addresses of all servers in the server cluster according to the rules of descending or ascending; each server in the server cluster can independently store the address lexicographical order, and if a new server is added to the server cluster, it can be rebroadcast to each server for storage after the address lexicographical order is updated.

[0027] Optionally, the master server election results may include the master server's Internet Protocol address and the determination result of whether the local machine has been elected as the master server.

[0028] Optionally, the master server election result can be obtained based on the local Internet Protocol address and the lexicographical order of the server cluster addresses, which may include: Using the target server, based on the address lexicographical order of the server cluster, obtain the minimum Internet Protocol address in the address lexicographical order, and determine whether the local Internet Protocol address is the same as the minimum Internet Protocol address; If they are the same, the target server is determined to be the master server; otherwise, the target server is determined not to be the master server, and the master server is determined according to the minimum Internet Protocol address.

[0029] The advantage of this setup is that the election process for the master server is fully automated, requiring no manual intervention, and ensuring that the server cluster can maintain time synchronization even when dynamically adjusted.

[0030] Optionally, Network Time Protocol (NTP) is a network protocol that can be used to synchronize the clocks of computer systems; a public network time protocol server can refer to a time source server deployed on the public Internet, used to provide a high-precision time reference for the main server. Public network time protocol servers can typically provide millisecond-level time accuracy.

[0031] Optionally, in this embodiment of the invention, a high-precision network time protocol service with an accuracy of about 1 millisecond can be selected, such as the Chrony service.

[0032] Optionally, if the target server is selected as the master server, the target server needs to synchronize its time with the public network time protocol server, maintain a high-precision clock, and provide time synchronization services for the slave servers; if the target server is a slave server, it synchronizes its time with the master server to keep its own clock consistent with the cluster reference time.

[0033] Optionally, the time synchronization service provided by the master server can specifically refer to: when a slave server requests time synchronization, the master server's clock is transmitted to the slave server through the local area network time protocol service.

[0034] Optionally, the clock step calibration mechanism can refer to gradually aligning the clock of the slave server with that of the master server by adjusting the slave server's clock multiple times. The adjustment step size can be the time difference between the slave server and the master server. The clock step calibration mechanism can be preset with the number of adjustments. When the preset number of adjustments is reached, the time synchronization is considered complete.

[0035] Understandably, based on the clock step calibration mechanism and the low latency characteristics within the local area network, rapid convergence of the server clock can be achieved. Assume the clock deviation between the master and slave servers is Δt. ij After time synchronization, the clock deviation Δt between the master server and the slave server will be... ij =[(T2-T1)-(T4-T3)] / 2≤2δ LAN +ε NTP Where T1 is the timestamp of the time alignment request sent from the slave server to the master server, T2 is the timestamp of the master server receiving the time alignment request, T3 is the timestamp of the master server returning the master server clock to the slave server, and T4 is the timestamp of the slave server receiving the master server clock. The one-way latency within the local area network is δ. LAN , ε NTP The inherent precision of the Network Time Protocol (NTP) and the low latency characteristics of local area networks (LANs) result in a one-way latency δ within the LAN.LAN Generally, the precision is less than or equal to 0.1 milliseconds. When selecting a Network Time Protocol (NTP) service, its own precision ε... NTP Typically, it can be 1 millisecond, and the clock deviation between the slave server and the master server is less than 2 milliseconds.

[0036] Furthermore, in the scenario of power design peak assessment of a single server cluster, after each slave server in the server cluster synchronizes with the master server, since the time difference between each server in the server cluster is in the millisecond range, and the typical duration of the power peak of the GCU is generally 10-50 milliseconds, the server cluster time synchronization method of the present invention can make the triggering time error of each GCU thread much smaller than the duration of the power peak, thus ensuring that the power consumption peak of each GCU remains consistent on the time axis.

[0037] Furthermore, in the scenario of power design peak assessment for large-scale server clusters, since the master server in each server cluster synchronizes its clock with the public network time protocol server based on the high-precision network time protocol service, the clock reference deviation of each server cluster is less than 2 milliseconds. Even if the clock reference deviation is superimposed with the clock deviation between each server in the server cluster, the clock deviation between each server in a large-scale cluster will not exceed 5 milliseconds, which is still far lower than the typical duration of GCU power peak.

[0038] Therefore, it can be seen that the server cluster time synchronization method of the present invention can achieve high-precision clock alignment between multiple server clusters and between multiple servers within a single server cluster. Moreover, the method does not involve hardware structure and does not require the use of expensive high-precision dedicated hardware. It has wide applicability and supports high-precision time synchronization in server cluster scenarios of different sizes.

[0039] Optionally, the operating system kernel of each server in the server cluster maintains a UTC (Coordinated Universal Time) clock. After the server cluster time is synchronized, the UTC clocks of each server are aligned. The UTC clock represents time with a timestamp. In this embodiment, the UTC clock can be in milliseconds.

[0040] Figure 2 As an alternative time synchronization relational topology graph, such as Figure 2 As shown, the master server synchronizes its time with the public network time protocol server via public network time synchronization, and then the master server synchronizes its time with each server in the server cluster within the local area network.

[0041] Among these steps, after determining that the target server has entered the peak power design assessment phase, and after obtaining the master server election result based on the local Internet Protocol address and the address lexicographical order of the server cluster, the process may further include: By using the target server, the server system type is identified, and time synchronization service is configured in the corresponding configuration address according to the server system type; Based on the target server system type, configure the firewall and open the time synchronization port to enable time synchronization for the server cluster.

[0042] Optionally, existing server systems can be pre-classified into specified server system types. Different server system types correspond to different time synchronization services, different configuration addresses, and different firewall types, but all use high-precision time synchronization services.

[0043] Optionally, through the target server, a pre-set script can be used to automatically read system files, obtain the server system of the target server, and determine the server system type to which the server system of the target server belongs; Based on the pre-established correspondence between server system type, configuration address, and time synchronization service, determine the corresponding configuration address and time synchronization service, and configure the time synchronization service in the corresponding configuration address.

[0044] Optionally, time synchronization service configuration can refer to setting time synchronization-related parameters in the corresponding configuration address according to the server system type, thereby ensuring that the server can synchronize time with the specified network time protocol server.

[0045] Optionally, the target server can be used to determine the firewall type corresponding to the server system type, and firewall rules can be automatically configured directly based on the firewall type, while opening the time synchronization port required for the Network Time Protocol service.

[0046] S120. Using the target server, at the start time of the current triggering round, calculate the next triggering time of the GCU thread based on the preset delay time, triggering cycle, and start time.

[0047] Optionally, a GCU thread can refer to a worker thread running on the GCU performing full-load computation.

[0048] Optionally, during the peak power design assessment of the server cluster, multiple rounds of GCU thread synchronization triggering are required. Ultimately, the peak power design assessment needs to be performed based on the complete power data generated by the multiple triggering processes.

[0049] Optionally, after the server cluster time synchronization is completed, each server synchronously starts the load assessment process and enters the first triggering round. After each GCU thread triggering and execution is completed, the next triggering round is entered.

[0050] Optionally, when entering the current triggering round, the UTC clock of the target server can be directly read as the starting time of the current triggering round. Since the service area cluster time synchronization has been completed in step S110, the clocks of each server are highly aligned, and the error of the starting time read by each server is extremely small.

[0051] Optionally, each server in the server cluster is pre-configured with the same delay time and trigger period. The delay time is used to provide a buffer time for the first synchronous trigger of the GCU thread, and the trigger period is used to control the interval between each trigger of the GCU thread. The trigger period must be greater than the maximum execution time of the GCU thread.

[0052] Optionally, the next trigger time refers to the absolute point in time when all servers in the server cluster start the GCU thread simultaneously.

[0053] Figure 3 This is a schematic diagram of the timeline for one possible GCU synchronous triggering method. For example... Figure 3 As shown, when the delay time is set to 1 second and the trigger period T is set to 2 seconds, according to UTC time, starting from the beginning of the first trigger round, after time t=1 second, the GCU threads of GCU0 and GCU1 in servers A, B, and C are synchronously triggered. After time t=3 seconds, the GCU threads of GCU0 and GCU1 in servers A, B, and C are synchronously triggered again. After time t=5 seconds, the GCU threads of GCU0 and GCU1 in servers A, B, and C are synchronously triggered again, and so on.

[0054] Specifically, calculating the next trigger time for the GCU thread at the start time of the current trigger round via the target server, based on a preset delay time, trigger period, and the start time, may include: The target server calculates a reference time based on the start time and the delay time, and performs a modulo operation on the reference time and the trigger period to obtain the modulo result. The target server calculates the time to the next cycle boundary based on the remainder result and the triggering period. The target server performs a summation operation on the reference time and the time from the next cycle boundary, and uses the summation result as the next trigger time for the GCU thread.

[0055] Understandably, although Figure 3 In the synchronous trigger timeline shown, the GCU thread triggers according to the specified trigger period each time. However, in each trigger round, the next trigger time needs to be calculated in the above way, instead of directly superimposing the trigger period with the previous trigger time. This avoids the error of the previous round affecting the subsequent round and ensures the trigger accuracy of long-term operation.

[0056] Optionally, the start time and the delay time can be added together to obtain the reference time. For example, if the UTC timestamp of the start time is 1747190400000 and the delay time is 1000 milliseconds, then the reference time is 1747190400000 + 1000 = 1747190401000.

[0057] Optionally, after performing a modulo operation on the reference time and the trigger period, the remainder result represents the margin between the reference time and the boundary of the previous period. For example, if the trigger period is 2000 milliseconds and the reference time is 1747190401000, then performing a modulo operation on the reference time and the trigger period (1747190401000%2000=1000) yields a remainder of 1000, indicating that the distance between the reference time and the boundary of the previous period is 1000 milliseconds.

[0058] Furthermore, by calculating the difference between the trigger period and the remainder result, we can obtain the time to the next cycle boundary. Continuing the previous example, the trigger period is 2000 milliseconds, and the remainder result is 1000 milliseconds. Therefore, the time to the next cycle boundary is 2000-1000=1000, which means there are still 1000 milliseconds to go before the next cycle boundary.

[0059] Continuing with the previous example, if the reference time is 1747190401000 and the time from the next cycle boundary is 1000, then the timestamp of the next trigger moment is 1747190401000 + 1000 = 1747190402000.

[0060] S130. Through the target server, periodically read the server clock, determine whether the next trigger time has been reached based on the currently read server clock, and when it is determined that the next trigger time has been reached, synchronously trigger all GCU threads in the target server to align with the trigger time of the GCU threads of the other servers in the server cluster.

[0061] Optionally, based on the currently read server clock, obtain the current UTC timestamp of the server, and determine whether the UTC timestamp is equal to the timestamp corresponding to the next trigger time; if they are equal, determine that the next trigger time has arrived; if they are not equal, determine the time for the next read server clock based on the time difference between the currently read server clock and the next trigger time.

[0062] Optionally, upon reaching the next trigger time, all GCU threads in the target server are simultaneously triggered, causing them to enter a full-load operation state at the same time.

[0063] It is understandable that, since all servers in the server cluster are time-aligned, during the peak power design assessment phase of the cluster, all servers in the server cluster execute the peak power design assessment method of the server cluster described in this embodiment as the target server. Therefore, the GCU threads in all servers in the server cluster can achieve synchronous triggering, enabling the server cluster to enter a full-load operation state within a millisecond-level time window, ensuring the superposition of power consumption peaks.

[0064] S140. Through the power analysis system, collect multi-dimensional power data of each server in the server cluster during the execution of GCU threads, and analyze the results to obtain the peak power design evaluation of the cluster.

[0065] Optionally, the power analysis system may include power monitoring equipment and power analysis equipment. The power monitoring equipment can be used to collect multi-dimensional power data of each server in the server cluster during the execution of GCU threads. The power analysis equipment can be used to summarize the collected multi-dimensional power data and, in combination with the load scale and runtime of the server cluster, analyze and obtain the peak power design evaluation result of the cluster.

[0066] Optionally, multi-dimensional power data may include power consumption data of each server and each GCU in the server cluster, peak power of each triggering round, average power of each triggering round, and power change curve.

[0067] Optionally, the peak power design assessment results may include information such as the actual peak power of the server cluster, the time of peak occurrence, the duration of peak, average power consumption, load size matching relationship, and power supply and distribution recommendations.

[0068] Specifically, the power analysis system collects multi-dimensional power data from each server in the server cluster during the execution of GCU threads, and analyzes the data to obtain the peak power design assessment results for the cluster, which may include: The power analysis system collects power consumption data, peak power for each triggering round, average power for each triggering round, and power change curves for each server and each GCU in the server cluster. The power analysis system aggregates multi-dimensional power data and combines it with the load scale and runtime of the server cluster to obtain the peak power design assessment results for the cluster.

[0069] Optionally, the server's power consumption data can refer to the real-time power consumption data of a single server during the testing process; the power consumption data of each GCU can refer to the independent real-time power consumption data of each GCU chip inside the server.

[0070] Optionally, the peak power of each trigger round can refer to the maximum instantaneous power consumption value collected in that round of testing after a single synchronous trigger; the average power of each trigger round can refer to the arithmetic mean of the power consumption data within a single trigger round, used to reflect the power consumption level of a stable load; the power change curve can refer to a continuous time-series waveform formed with time as the horizontal axis and power consumption as the vertical axis, which can intuitively display information such as the rising edge of power consumption, peak duration, and fluctuation characteristics.

[0071] Optionally, load scale can refer to parameters such as the number of servers participating in the peak power design assessment of this server cluster, the total number of GCUs, and the number of load threads; runtime can refer to the duration of the full-load stress test of the GCUs.

[0072] The technical solution of this invention adopts the following approach: When the target server determines that it has entered the peak power design assessment stage, it performs server cluster time synchronization; at the start of the current triggering round, it calculates the next triggering time of the GCU thread based on a preset delay time, triggering cycle, and start time; it periodically reads the server clock, determines whether the next triggering time has been reached based on the currently read server clock, and when the next triggering time is determined, it synchronously triggers all GCU threads in the target server to align with the triggering times of the GCU threads of the other servers in the server cluster; and it uses a power analysis system to collect multi-dimensional power data from each server in the server cluster during the execution of the GCU threads and analyzes the data to obtain the peak power design assessment results of the cluster. This approach enables high-precision time synchronization of the server cluster without relying on dedicated hardware, ensuring high alignment of the triggering times of the GCU threads of all servers in the cluster. This allows the peak power consumption of each GCU to be completely superimposed on the time axis, truly reflecting the peak power design of the cluster, and thus enabling accurate assessment of the peak power design of the server cluster.

[0073] Example 2 Figure 4 This is a flowchart of a peak power design assessment method for a server cluster provided in Embodiment 2 of the present invention. Based on the above embodiments, this embodiment specifically illustrates the calculation and monitoring process for the next triggering moment. Figure 4 As shown, the method includes: S210. When the target server determines that the cluster power design peak assessment stage has been entered, server cluster time synchronization is performed.

[0074] S220. Calculate the reference time based on the start time and delay time using the target server, and perform a modulo operation on the reference time and trigger period to obtain the modulo result.

[0075] S230. Calculate the time to the next cycle boundary based on the remainder result and the triggering cycle through the target server.

[0076] S240. Through the target server, sum the reference time and the time from the next cycle boundary, and use the summation result as the next trigger time of the GCU thread.

[0077] Specifically, the process of periodically reading the server clock through the target server and determining whether the next trigger time has arrived based on the currently read server clock may include: Using the target server, based on the current maintained reading frequency, the server clock is read periodically, and it is determined whether the currently read server clock is equal to the next trigger time. If so, then the next trigger time is determined; If not, the remaining time to the next trigger time is calculated based on the next trigger time and the currently read server clock, and the reading frequency is adjusted according to the remaining time.

[0078] Optionally, the reading frequency is a clock reading time interval, used to control the data reading accuracy.

[0079] Optionally, the remaining time until the next trigger time can refer to the difference between the next trigger time and the currently read server clock, used as the time difference between the currently read server clock and the next trigger time.

[0080] Optionally, the reading frequency can be dynamically adjusted according to the remaining time. The remaining time range can be preset, and a mapping relationship between the remaining time range and the reading frequency can be established. The longer the remaining time, the higher the corresponding reading frequency, and the shorter the remaining time, the lower the corresponding reading frequency. In this embodiment, in order to achieve millisecond-level synchronization triggering accuracy, the minimum reading frequency should be set to less than 1 millisecond, such as 100 microseconds.

[0081] In one optional example, if the remaining duration is greater than 200 milliseconds, the corresponding read frequency is 100 milliseconds; if the remaining duration is less than or equal to 200 milliseconds but greater than 100 milliseconds, the corresponding read frequency is 50 milliseconds; if the remaining duration is less than or equal to 100 milliseconds but greater than 10 milliseconds, the corresponding read frequency is 5 milliseconds; if the remaining duration is less than or equal to 10 milliseconds but greater than 5 milliseconds, the corresponding read frequency is 1 millisecond; and if the remaining duration is less than or equal to 5 milliseconds, the corresponding read frequency is 100 microseconds.

[0082] The advantage of this setting is that it can gradually shorten the read frequency as the next trigger moment approaches, ensuring high-precision triggering while avoiding long-term CPU (Central Processing Unit) spinning, thus significantly reducing CPU resource usage.

[0083] S250. Through the target server, periodically read the server clock, determine whether the next trigger time has been reached based on the currently read server clock, and when it is determined that the next trigger time has been reached, synchronously trigger all GCU threads in the target server to align with the trigger time of the GCU threads of the other servers in the server cluster.

[0084] S260. Through the power analysis system, collect power consumption data of each server and each GCU in the server cluster, peak power of each triggering round, average power of each triggering round, and power change curve.

[0085] S270. Through the power analysis system, the collected multi-dimensional power data is aggregated and combined with the load scale and runtime of the server cluster to obtain the peak power design assessment results of the cluster.

[0086] Optionally, through the power analysis system, the collected multi-dimensional power data can first be summarized and organized, and the real-time power consumption, peak power, average power and power change curves of each server and each GCU after synchronization triggering can be aligned in time sequence to form a complete cluster power consumption dataset. Then, the total peak power consumption of the cluster in each triggering round is extracted, and the maximum value is selected from the peak values ​​of multiple rounds as the original peak power of the cluster.

[0087] Furthermore, the matching between the original power peak and the load scale can be verified by combining the server cluster load scale to ensure that the peak is generated by all GCUs synchronously at full load. In addition, the continuous stability of the peak can be analyzed by combining the stress test runtime to eliminate the influence of transient interference and non-full load data.

[0088] Optionally, through the power analysis system, the actual power design peak value of the cluster can be obtained through extreme value statistics, time series verification, and load matching analysis, and a complete evaluation result including peak value, peak duration, average power consumption, and power supply and distribution recommendations can be output.

[0089] The technical solution of this invention adopts the following approach: When the target server determines that it has entered the peak power design assessment stage, it performs server cluster time synchronization; at the start of the current triggering round, it calculates the next triggering time of the GCU thread based on a preset delay time, triggering cycle, and start time; it periodically reads the server clock, determines whether the next triggering time has been reached based on the currently read server clock, and when the next triggering time is determined, it synchronously triggers all GCU threads in the target server to align with the triggering times of the GCU threads of the other servers in the server cluster; and it uses a power analysis system to collect multi-dimensional power data from each server in the server cluster during the execution of the GCU threads and analyzes the data to obtain the peak power design assessment results of the cluster. This approach enables high-precision time synchronization of the server cluster without relying on dedicated hardware, ensuring high alignment of the triggering times of the GCU threads of all servers in the cluster. This allows the peak power consumption of each GCU to be completely superimposed on the time axis, truly reflecting the peak power design of the cluster, and thus enabling accurate assessment of the peak power design of the server cluster.

[0090] Example 3 Figure 5 This is a flowchart of a method for evaluating the peak power design of a server cluster according to Embodiment 3 of the present invention. Based on the above embodiments, this embodiment specifically illustrates the process of individually evaluating the peak power design of a target server within the server cluster. Figure 5 As shown, the method includes: S310. When it is determined that the power design peak assessment stage of the target server is to be entered, if the target server contains multiple GCUs, then register the software barrier synchronization point and determine the total number of GCU threads based on the number of GCUs in the target server and the number of subtask threads of each GCU.

[0091] Optionally, during the peak power design assessment phase of the target server, the target server can be a single server currently being assessed in the server cluster. By synchronously triggering all GCU threads in the target server, the peak power design of the target server can be accurately captured.

[0092] Optionally, if the target server contains only one GCU, then only the thread of that single GCU needs to be triggered, and there is no issue of not being able to trigger synchronously.

[0093] Optionally, if the target server contains multiple GCUs, it is necessary to ensure that the GCU threads of each GCU are triggered synchronously.

[0094] Optionally, the software barrier synchronization point can refer to a pre-created, purely software-implemented synchronization waiting point. All GCU threads must reach the software barrier synchronization point and wait until all GCU threads have reached the software barrier synchronization point before being allowed to pass.

[0095] Optionally, the number of GCUs refers to the total number of GCU chips installed in the target server, and the number of GCU subtask threads can refer to the number of GCU threads registered on each GCU chip.

[0096] Optionally, the total number of GCU threads is the sum of the registered GCU threads in all GCU chips in the target server. The total number of GCU threads can be obtained by adding up the number of subtask threads of each GCU.

[0097] S320. Through the target server, whenever the GCU thread reaches the software barrier synchronization point, the GCU thread is blocked so that it enters a waiting state and the counter is updated.

[0098] Optionally, blocking the GCU thread can mean pausing the GCU thread's execution and causing it to enter a waiting state.

[0099] Optionally, a counter can be used to record the number of GCU threads that have reached the software barrier synchronization point. The counter is incremented by 1 each time a GCU thread is blocked at the software barrier synchronization point.

[0100] S330. When the counter value is determined to be equal to the total number of GCU threads by the target server, all GCU threads are broadcast to wake up, so that all GCU threads can cross the software barrier synchronization point at the same time.

[0101] Optionally, when the counter value is equal to the total number of GCU threads, it indicates that all GCU threads in the target server have arrived and entered the waiting state. At this time, a wake-up signal can be sent to all GCU threads at the software barrier synchronization point. Each GCU thread is woken up at the same time and ends the waiting state, and begins to execute the stress test task.

[0102] It is understandable that after each GCU thread finishes execution, the GCU thread can run again and reach the software barrier synchronization point, waiting for the next round of GCU thread synchronization to be awakened, so as to achieve multi-round GCU thread synchronization triggering within the target server.

[0103] Figure 6 This is a schematic diagram illustrating an optional synchronous wake-up method using a single server GCU thread. For example... Figure 6As shown, taking a single server containing four GCU threads (GCU0, GCU1, GCU2, and GCU3) as an example, whenever a GCU thread reaches the software barrier synchronization point, the counter arrived is incremented by 1, and the GCU thread is blocked. When the counter value reaches 4, that is, when the counter value is equal to the total number of GCU threads, all GCU threads are broadcast to wake up. All GCU threads then synchronously enter full-load operation. Based on the broadcast-wake-up GCU threads, the synchronization error can be less than 1 millisecond, which is much smaller than the duration of the power peak.

[0104] The advantage of this setup is that, in existing technologies, in single-machine multi-GCU scenarios, each GCU thread typically achieves near-synchronous triggering by setting a sleep function and a preset fixed delay time. However, when system scheduling jitter exists, the error generated by system scheduling jitter is typically between 1 and 10 milliseconds, and the error accumulates and increases with the number of threads, resulting in the peak power consumption of the GCUs being misaligned. The method of this application can achieve high-precision synchronous triggering of multiple GCUs on a single machine, effectively solving the problems of reliance on system scheduling synchronization and error accumulation in existing technologies.

[0105] S340. Through the power analysis system, collect multi-dimensional power data of the target server during the execution of the GCU thread, and analyze the power design peak evaluation results of the target server.

[0106] Optionally, during the peak power design assessment phase of the target server, multi-dimensional power data may include real-time power consumption, peak power, average power, power change curves, and other power consumption time-series data of the target server and each GCU within the target server in each triggering round.

[0107] Optionally, the peak power design assessment results for the target server may include data such as the target server's peak power design, peak duration, average power consumption, stability analysis, and power supply and distribution recommendations.

[0108] The technical solution of this invention adopts the following approach: When the target server is determined to enter the power design peak assessment stage, if the target server contains multiple GCUs, a software barrier synchronization point is registered, and the total number of GCU threads is determined based on the number of GCUs and the number of subtask threads of each GCU. Whenever a GCU thread reaches the software barrier synchronization point, the GCU thread is blocked to enter a waiting state, and a counter is updated. When the counter value is determined to be equal to the total number of GCU threads, all GCU threads are broadcast to wake up, so that all GCU threads cross the software barrier synchronization point synchronously at the same time. By collecting multi-dimensional power data of the target server during the execution of GCU threads through a power analysis system and analyzing it to obtain the power design peak assessment result of the target server, thread scheduling deviation can be effectively eliminated, ensuring that the power consumption peaks of each GCU in a single machine are completely superimposed, improving the accuracy of power design peak assessment. At the same time, each GCU thread adopts a blocking and waiting mechanism, which does not occupy unnecessary CPU resources, resulting in low system overhead and stable operation.

[0109] Example 4 Figure 7 This is a schematic diagram of a power design peak assessment device for a server cluster provided in Embodiment 4 of the present invention. Figure 7 As shown, the device includes: a time synchronization module 410, a trigger time calculation module 420, a synchronization trigger module 430, and a cluster power data acquisition and analysis module 440.

[0110] The time synchronization module 410 is used to perform server cluster time synchronization when the target server determines that the cluster power design peak assessment stage has been entered.

[0111] The trigger time calculation module 420 is used to calculate the next trigger time of the GCU thread at the start time of the current trigger round through the target server, based on the preset delay time, trigger period and the start time.

[0112] The synchronization trigger module 430 is used to periodically read the server clock through the target server, determine whether the next trigger time has been reached based on the currently read server clock, and when it is determined that the next trigger time has been reached, synchronously trigger all GCU threads in the target server to align with the trigger time of the GCU threads of the other servers in the server cluster.

[0113] The cluster power data acquisition and analysis module 440 is used to collect multi-dimensional power data of each server in the server cluster during the execution of the GCU thread through the power analysis system, and analyze the data to obtain the peak power design evaluation result of the cluster.

[0114] The technical solution of this invention adopts the following approach: When the target server determines that it has entered the peak power design assessment stage, it performs server cluster time synchronization; at the start of the current triggering round, it calculates the next triggering time of the GCU thread based on a preset delay time, triggering cycle, and start time; it periodically reads the server clock, determines whether the next triggering time has been reached based on the currently read server clock, and when the next triggering time is determined, it synchronously triggers all GCU threads in the target server to align with the triggering times of the GCU threads of the other servers in the server cluster; and it uses a power analysis system to collect multi-dimensional power data from each server in the server cluster during the execution of the GCU threads and analyzes the data to obtain the peak power design assessment results of the cluster. This approach enables high-precision time synchronization of the server cluster without relying on dedicated hardware, ensuring high alignment of the triggering times of the GCU threads of all servers in the cluster. This allows the peak power consumption of each GCU to be completely superimposed on the time axis, truly reflecting the peak power design of the cluster, and thus enabling accurate assessment of the peak power design of the server cluster.

[0115] Based on the above embodiments, the time synchronization module 410 can be specifically used for: When the target server determines that it has entered the peak power design assessment phase, it obtains the master server election result based on the local Internet Protocol address and the address lexicographical order of the server cluster, and performs any of the following: If the target server is determined to be the master server, then time synchronization is performed with the public network time protocol server, and time synchronization service is provided to the slave servers in the server cluster. If it is determined that the target server is not selected as the master server, it will act as a slave server and use a clock step calibration mechanism to synchronize time with the master server in the server cluster.

[0116] Based on the above embodiments, a time synchronization configuration module may also be included, for: By using the target server, the server system type is identified, and time synchronization service is configured in the corresponding configuration address according to the server system type; Based on the target server system type, configure the firewall and open the time synchronization port to enable time synchronization for the server cluster.

[0117] Based on the above embodiments, the trigger time calculation module 420 can be specifically used for: The target server calculates a reference time based on the start time and the delay time, and performs a modulo operation on the reference time and the trigger period to obtain the modulo result. The target server calculates the time to the next cycle boundary based on the remainder result and the triggering period. The target server performs a summation operation on the reference time and the time from the next cycle boundary, and uses the summation result as the next trigger time for the GCU thread.

[0118] Based on the above embodiments, the synchronization triggering module 430 can be specifically used for: Using the target server, based on the current maintained reading frequency, the server clock is read periodically, and it is determined whether the currently read server clock is equal to the next trigger time. If so, then the next trigger time is determined; If not, the remaining time to the next trigger time is calculated based on the next trigger time and the currently read server clock, and the reading frequency is adjusted according to the remaining time.

[0119] Based on the above embodiments, the cluster power data acquisition and analysis module 440 can be specifically used for: The power analysis system collects power consumption data, peak power for each triggering round, average power for each triggering round, and power change curves for each server and each GCU in the server cluster. The power analysis system aggregates multi-dimensional power data and combines it with the load scale and runtime of the server cluster to obtain the peak power design assessment results for the cluster.

[0120] Based on the above embodiments, a single-server power design peak assessment module may also be included, for: When the target server is determined to enter the peak power design assessment stage, if the target server contains multiple GCUs, a software barrier synchronization point is registered, and the total number of GCU threads is determined according to the number of GCUs in the target server and the number of subtask threads of each GCU. Through the target server, whenever the GCU thread reaches the software barrier synchronization point, the GCU thread is blocked to put the GCU thread into a waiting state and update the counter. When the target server determines that the counter value is equal to the total number of GCU threads, it broadcasts a wake-up call to all GCU threads so that each GCU thread can synchronously cross the software barrier synchronization point at the same time. The power analysis system collects multi-dimensional power data of the target server during the execution of the GCU thread, and analyzes the data to obtain the peak power design evaluation result of the target server.

[0121] The power design peak assessment device for server clusters provided in this embodiment of the invention can execute the power design peak assessment method for server clusters provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0122] Example 5 Figure 8 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0123] like Figure 8 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0124] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0125] Processor 11 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the peak power design evaluation method for server clusters described in embodiments of the present invention. That is: When the target server determines that the cluster power design peak assessment phase has been entered, server cluster time synchronization is performed. Using the target server, at the start time of the current triggering round, the next triggering time of the GCU thread is calculated based on the preset delay time, triggering period, and the start time. The target server periodically reads the server clock, determines whether the next trigger time has been reached based on the currently read server clock, and when it is determined that the next trigger time has been reached, synchronously triggers all GCU threads in the target server to align with the trigger times of the GCU threads of the other servers in the server cluster. The power analysis system collects multi-dimensional power data of each server in the server cluster during the execution of GCU threads, and analyzes the data to obtain the peak power design evaluation results of the cluster.

[0126] In some embodiments, the peak power design assessment method for a server cluster can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the peak power design assessment method for a server cluster described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the peak power design assessment method for a server cluster by any other suitable means (e.g., by means of firmware).

[0127] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0128] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0129] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0130] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0131] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0132] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0133] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0134] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for evaluating peak power design in a server cluster, executed in cooperation between a target server in an AI server cluster and a power analysis system, characterized in that, include: When the target server determines that the cluster power design peak assessment phase has been entered, server cluster time synchronization is performed. Using the target server, at the start time of the current triggering round, the next triggering time of the Graphics Computing Unit (GCU) thread is calculated based on the preset delay time, triggering period, and the start time. The target server periodically reads the server clock, determines whether the next trigger time has been reached based on the currently read server clock, and when it is determined that the next trigger time has been reached, synchronously triggers all GCU threads in the target server to align with the trigger times of the GCU threads of the other servers in the server cluster. The power analysis system collects multi-dimensional power data of each server in the server cluster during the execution of GCU threads, and analyzes the data to obtain the peak power design evaluation results of the cluster.

2. The method according to claim 1, characterized in that, When the target server determines that the cluster power design peak assessment phase has begun, server cluster time synchronization is performed, including: When the target server determines that it has entered the peak power design assessment phase, it obtains the master server election result based on the local Internet Protocol address and the address lexicographical order of the server cluster, and performs any of the following: If the target server is determined to be the master server, then time synchronization is performed with the public network time protocol server, and time synchronization service is provided to the slave servers in the server cluster. If it is determined that the target server is not selected as the master server, it will act as a slave server and use a clock step calibration mechanism to synchronize time with the master server in the server cluster.

3. The method according to claim 2, characterized in that, After determining that the target server has entered the peak power design assessment phase, and obtaining the master server election result based on the local Internet Protocol address and the address lexicographical order of the server cluster, the process further includes: By using the target server, the server system type is identified, and time synchronization service is configured in the corresponding configuration address according to the server system type; Based on the target server system type, configure the firewall and open the time synchronization port to enable time synchronization for the server cluster.

4. The method according to claim 1, characterized in that, Using the target server, at the start time of the current triggering round, based on the preset delay time, triggering period, and the start time, the next triggering time of the GCU thread is calculated, including: The target server calculates a reference time based on the start time and the delay time, and performs a modulo operation on the reference time and the trigger period to obtain the modulo result. The target server calculates the time to the next cycle boundary based on the remainder result and the triggering period. The target server performs a summation operation on the reference time and the time from the next cycle boundary, and uses the summation result as the next trigger time for the GCU thread.

5. The method according to claim 1, characterized in that, The target server periodically reads the server clock and determines whether the next trigger time has arrived based on the currently read server clock, including: Using the target server, based on the current maintained reading frequency, the server clock is read periodically, and it is determined whether the currently read server clock is equal to the next trigger time. If so, then the next trigger time is determined; If not, the remaining time to the next trigger time is calculated based on the next trigger time and the currently read server clock, and the reading frequency is adjusted according to the remaining time.

6. The method according to claim 1, characterized in that, The power analysis system collects multi-dimensional power data from each server in the server cluster during GCU thread execution and analyzes it to obtain the peak power design assessment results for the cluster, including: The power analysis system collects power consumption data, peak power for each triggering round, average power for each triggering round, and power change curves for each server and each GCU in the server cluster. The power analysis system aggregates multi-dimensional power data and combines it with the load scale and runtime of the server cluster to obtain the peak power design assessment results for the cluster.

7. The method according to claim 1, characterized in that, Also includes: When the target server is determined to enter the peak power design assessment stage, if the target server contains multiple GCUs, a software barrier synchronization point is registered, and the total number of GCU threads is determined according to the number of GCUs in the target server and the number of subtask threads of each GCU. Through the target server, whenever the GCU thread reaches the software barrier synchronization point, the GCU thread is blocked to put the GCU thread into a waiting state and update the counter. When the target server determines that the counter value is equal to the total number of GCU threads, it broadcasts a wake-up call to all GCU threads so that each GCU thread can synchronously cross the software barrier synchronization point at the same time. The power analysis system collects multi-dimensional power data of the target server during the execution of the GCU thread, and analyzes the data to obtain the peak power design evaluation result of the target server.

8. A peak power design assessment device for a server cluster, executed in cooperation with a power analysis system within an AI server cluster, characterized in that, include: The time synchronization module is used to perform server cluster time synchronization through the target server when it is determined that the cluster power design peak assessment stage has been entered. The trigger time calculation module is used to calculate the next trigger time of the graphics computing unit (GCU) thread at the start time of the current trigger round, based on the preset delay time, trigger period, and the start time, through the target server. The synchronization triggering module is used to periodically read the server clock through the target server, determine whether the next triggering time has been reached based on the currently read server clock, and when it is determined that the next triggering time has been reached, synchronously trigger all GCU threads in the target server to align with the triggering time of the GCU threads of the other servers in the server cluster. The cluster power data acquisition and analysis module is used to collect multi-dimensional power data of each server in the server cluster during the execution of the GCU thread through the power analysis system, and analyze the data to obtain the peak power design evaluation results of the cluster.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the power design peak assessment method for the server cluster according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the power design peak assessment method for the server cluster according to any one of claims 1-7.

Images