Flow scheduling method and device, medium and program product

Abstract

The invention discloses a flow scheduling method and device, a medium and a program product, and relates to the technical field of heat dissipation, and the method comprises the steps: obtaining a video memory access frequency as a main parameter when a parallel computing function is executed, matching with other auxiliary parameters related to heat consumption, distributing a weight, determining a heat load scoring result, and then adjusting a cold plate micro-channel flow distribution coefficient, and determining an opening value of a flow control valve and sending a flow scheduling instruction to the liquid cooling system so as to dynamically adjust the cooling liquid flow of each cold plate micro-channel. In this way, accurate temperature control and heat dissipation of the video memory area of the computing device can be achieved, energy waste caused by unnecessary circulation of cooling liquid is avoided, the heat dissipation efficiency is remarkably improved, energy consumption is reduced, and the service life of the computing device is prolonged; in addition, the mode can autonomously complete full-flow operation such as video memory heat consumption monitoring, score calculation and flow adjustment, intelligent management of the heat dissipation process is achieved, manual operation errors and cost are reduced, and the stability and reliability of operation of computing equipment are guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of heat dissipation, in particular to a flow scheduling method, device, medium and program product. BACKGROUND

[0002] With the rapid development of artificial intelligence, big data analysis and other fields, the demand for high-performance computing is increasing. To meet this demand, computing devices are upgraded in performance, and the increase in video memory capacity and access speed directly leads to an increase in device heat dissipation problems. Liquid cooling heat dissipation technology is widely used due to its high efficiency, and the cooling liquid flows in the microchannels of the cold plate to carry away heat and maintain the normal temperature of the device.

[0003] In related technical solutions, the liquid cooling heat dissipation system usually adopts a fixed flow distribution method, that is, the flow of the cooling liquid in the microchannels of the cold plate is pre-set. When the heat dissipation of the computing device video memory is low, the fixed flow will cause waste of cooling liquid and increase the energy consumption of the circulating pump; when the heat dissipation of the video memory is high, the fixed flow cannot provide sufficient heat dissipation capacity, which will cause the temperature of the video memory to be too high, affecting the performance and service life of the computing device. SUMMARY

[0004] The present application provides a flow scheduling method, device, medium and program product to at least solve the problem of low heat dissipation efficiency, high energy consumption and affecting the performance and service life of the computing device caused by the fixed flow distribution method of the liquid cooling heat dissipation system in related technologies.

[0005] The present application provides a flow scheduling method, comprising: obtaining the video memory access frequency when the parallel computing function is executed; using the obtained video memory access frequency as a main parameter, and using other parameters related to heat dissipation as auxiliary parameters, assigning weights to the main parameter and the auxiliary parameters, and determining a heat load score result; based on the heat load score result, adjusting the flow distribution coefficient of the corresponding cold plate microchannel; According to the adjusted flow distribution coefficient, a flow control valve opening value is obtained, and a corresponding flow scheduling instruction is sent to the liquid cooling system to control the liquid cooling system to adjust the cooling liquid flow of each cold plate microchannel according to the flow control valve opening value, so that each region of the video memory is within the rated temperature range.

[0006] The present application also provides an electronic device comprising a memory for storing a computer program and a processor for executing the computer program to implement the steps of any of the above flow scheduling methods.

[0007] The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of any of the above-described traffic scheduling methods.

[0008] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above-described traffic scheduling methods.

[0009] The flow scheduling method provided by this invention obtains the memory access frequency during the execution of parallel computing functions as the main parameter, combines it with other auxiliary parameters related to heat consumption and assigns weights to determine the heat load score result, and then adjusts the flow distribution coefficient of the cold plate microchannel to obtain a determined flow control valve opening value and sends a scheduling command to the liquid cooling system. This dynamically adjusts the coolant flow rate of each cold plate microchannel, enabling precise temperature control and heat dissipation of the memory area of ​​the computing device. On the one hand, it can flexibly adapt the coolant flow rate according to the real-time heat consumption of the memory, avoiding energy waste caused by unnecessary coolant circulation, significantly improving heat dissipation efficiency and reducing energy consumption. On the other hand, the entire adjustment process is responsive, and the flow adjustment can be completed within milliseconds from monitoring heat consumption changes. It can promptly respond to situations where heat consumption rises rapidly, effectively prevent the temperature from rising sharply in a short period of time, ensure that the memory is always within the rated operating temperature range, reduce the damage of high temperature to hardware, and extend the service life of the computing device. In addition, this method does not require manual intervention and can autonomously complete the entire process of memory heat consumption monitoring, score calculation, and flow adjustment, realizing intelligent management of the heat dissipation process. This reduces manual operation errors and costs, and further ensures the stability and reliability of the computing device.

[0010] In addition, the present invention also provides corresponding electronic devices, computer-readable storage media and program products for the traffic scheduling method, which have the same or corresponding technical features as the traffic scheduling method mentioned above, and have the same effect. Attached Figure Description

[0011] To more clearly illustrate the embodiments of the present invention, the accompanying drawings used in the embodiments 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.

[0012] Figure 1 A flowchart of a traffic scheduling method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the cold plate microchannel structure provided in an embodiment of the present invention; Figure 3 This is a partially enlarged schematic diagram of the flow control valve provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a single channel cold plate assembly provided in an embodiment of the present invention. Detailed Implementation

[0013] With the rapid development of fields such as artificial intelligence and big data analytics, the demand for high-performance computing is increasing daily. Computing devices, with their powerful parallel computing capabilities, are widely used in these fields. To meet the ever-increasing computing demands, the performance of computing devices is continuously upgraded, with memory capacity and access speed constantly improving. This makes the heat dissipation problem during operation increasingly prominent. During the operation of computing devices, memory access is one of the major sources of heat generation. Liquid cooling technology, as a highly efficient heat dissipation method, is widely used in computing device cooling. It removes heat through the flow of coolant in the microchannels of a cold plate to maintain the normal operating temperature of the computing device. Traditional liquid cooling systems typically use a fixed flow distribution method, meaning the coolant flow rate in the microchannels of the cold plate is preset and does not adjust according to the real-time heat dissipation of the memory during computing device operation. This system mainly consists of a coolant circulation pump, cold plates, pipes, and flow control valves. The flow control valves operate at a fixed opening, allowing the coolant to flow at a fixed flow rate in each microchannel of the cold plate. This fixed flow distribution method has significant shortcomings. When the heat dissipation of the graphics memory in a computing device is low, a fixed flow rate leads to waste of coolant and increases the energy consumption of the circulating pump. Conversely, when the heat dissipation of the graphics memory is high, a fixed flow rate cannot provide sufficient heat dissipation, causing the graphics memory temperature to become too high, affecting the performance and lifespan of the computing device. To address this technical problem, this invention provides a flow rate scheduling method.

[0014] 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 of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0015] It should be noted that, in the description of this invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0016] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] The specific application environment architecture or specific hardware architecture on which the execution of the traffic scheduling method depends is described here.

[0018] The embodiments of the present invention provide a traffic scheduling method, and the method is described in detail in conjunction with the execution flow of the traffic scheduling method. Figure 1 A flowchart of the traffic scheduling method provided in the embodiments of the present invention is shown below. Figure 1 As shown, the method includes: S101. Obtain the frequency of video memory accesses during the execution of the parallel computing function.

[0019] It should be noted that the parallel computing function of this invention refers to a function that can be executed simultaneously on multiple processing units to accelerate computation. Here, the parallel computing function can be a CUDA kernel function or other functions. The computing device involved can be a Graphics Processing Unit (GPU) or other types of processors. For example, a CUDA kernel function can be executed in parallel on multiple threads of a GPU; such as a vector addition kernel function, which allows a large number of GPU threads to process vector element addition operations simultaneously.

[0020] In step S101, this invention obtains the frequency of video memory accesses during the execution of the parallel computing function. This is a crucial step in accurately capturing the actual workload of the video memory. In practical applications, when a computing device runs a deep learning training task, the parallel computing function frequently accesses the video memory to acquire and store data. The frequency of video memory accesses is directly related to the intensity of data interaction, which in turn is highly correlated with the amount of heat generated by the video memory. A higher access frequency means that the video memory processes a larger amount of data per unit time, and the heat generated is usually greater. Therefore, obtaining this data provides core input parameters that reflect the true operating state of the video memory for subsequent heat load scoring calculations, avoiding the potential lag that may exist when relying solely on indirect indicators such as temperature.

[0021] S102. The obtained memory access frequency is used as the main parameter, and other parameters related to heat consumption are used as auxiliary parameters. After assigning weights to the main parameter and auxiliary parameters, the heat load score result is determined.

[0022] In implementation, during step S102, the memory access frequency is used as the primary parameter, and heat dissipation-related parameters are used as auxiliary parameters with assigned weights. Since the memory access frequency directly reflects the intensity of data interaction in parallel computing, using it as the primary parameter can accurately pinpoint the core driving factor of heat load. Including other heat dissipation-related auxiliary parameters and assigning weights can compensate for the limitations of a single parameter and more comprehensively depict the actual heat load state. By performing heat load scoring calculation after merging the primary and auxiliary parameters with preset weights, the final determined heat load score result can better reflect the actual heat dissipation of the hardware, providing a more reliable basis for subsequent heat dissipation adjustments such as flow control.

[0023] S103. Based on the heat load score, adjust the flow distribution coefficient of the corresponding cold plate microchannel.

[0024] In implementation, when performing step S103, the heat load score directly reflects the actual heat generation intensity of the corresponding area of ​​each cold plate microchannel. Based on this, the flow distribution coefficient can be adjusted to ensure that heat dissipation resources are accurately matched with heat consumption requirements. This flow distribution method guided by heat load not only ensures the heat dissipation reliability of high heat generation areas, but also achieves energy consumption optimization of the entire liquid cooling system.

[0025] S104. Based on the adjusted flow distribution coefficient, obtain the flow control valve opening value and send the corresponding flow scheduling command to the liquid cooling system to control the liquid cooling system to adjust the coolant flow of each cold plate microchannel according to the flow control valve opening value, so that each area of ​​the memory is within the rated temperature range.

[0026] In implementation, during step S104, the opening value of the flow control valve is determined based on the adjusted flow distribution coefficient, and a flow scheduling command is sent to the liquid cooling system to adjust the coolant flow rate of each cold plate microchannel. The flow distribution coefficient is the target flow rate ratio, while the opening value is the specific action quantity executed by the hardware. The conversion between the two ensures that the control requirements can be accurately executed by the liquid cooling system. By controlling the coolant flow rate on demand through commands, the heat dissipation can be specifically improved when the heat dissipation of the video memory is high, and the flow waste can be reduced when the heat dissipation is low. Ultimately, the temperature of each area of ​​the video memory is stabilized within the rated range, which not only ensures the safety of the video memory hardware and the computing performance, but also achieves a balance between heat dissipation efficiency and energy consumption.

[0027] Figure 2 This is a schematic diagram of a cold plate microchannel structure provided in an embodiment of the present invention. Figure 2As shown, the main body 1 of the cold plate adopts an aluminum alloy microchannel array structure. A is the liquid cooling channel inlet, B is the liquid cooling channel outlet, and C is the heat sink part that actually contacts the computing device. The arrows indicate the direction of water flow. A flow control valve 2 (which can be called a micro control valve) is installed at the liquid cooling channel inlet A of each computing device. After receiving the flow scheduling command, the valve opening is controlled, thereby controlling the flow rate. The valve opening range is 0-100%.

[0028] Figure 3 This is a partially enlarged schematic diagram of the flow control valve provided in an embodiment of the present invention. Figure 3 As shown, the coolant enters from the liquid cooling channel inlet A and flows to the flow control valve 2. The flow control valve 2 can receive flow scheduling commands and adjust its own opening according to the commands, thereby controlling the flow of coolant into the microchannel below, thus achieving precise regulation of the coolant flow in the microchannel of the cold plate.

[0029] Figure 4 This is a schematic diagram of a single-channel cold plate assembly provided in an embodiment of the present invention. Figure 4 As shown, the liquid cooling plate assembly 3 serves as the core, with heat sink C on its surface for contacting computing devices (such as GPUs) to conduct heat. Coolant enters from the liquid cooling channel inlet A, and after its flow rate is regulated by the flow control valve 2, it flows inside the liquid cooling plate, absorbing the heat transferred by the heat sink, and finally flows out from the liquid cooling channel outlet B. The flow control valve 2 can receive flow scheduling commands and precisely adjust the coolant flow rate according to factors such as heat load, thereby achieving efficient and on-demand heat dissipation for the computing device and ensuring stable operation of the computing device at a suitable temperature.

[0030] In the flow scheduling method provided by this invention, the frequency of video memory access during the execution of parallel computing functions is first obtained as the main parameter. This is combined with other auxiliary parameters related to heat consumption and weighted accordingly to determine the heat load score. Then, the flow distribution coefficient of the cold plate microchannel is adjusted to obtain the opening value of the flow control valve and send a scheduling command to the liquid cooling system. This dynamically adjusts the coolant flow rate of each cold plate microchannel, enabling precise temperature control and heat dissipation of the video memory area of ​​the computing device. On the one hand, the coolant flow rate can be flexibly adapted based on the real-time heat consumption of the video memory, avoiding unnecessary energy waste caused by unnecessary coolant circulation, significantly improving heat dissipation efficiency and reducing energy consumption. On the other hand, the entire adjustment process is rapid, from monitoring heat consumption changes to completing flow adjustment within milliseconds. This allows for timely responses to rapid increases in heat consumption, effectively preventing a sharp rise in temperature in a short period, ensuring that the video memory is always within its rated operating temperature range, reducing damage to hardware from high temperatures, and extending the lifespan of the computing device. Furthermore, this method requires no manual intervention and can autonomously complete the entire process of video memory heat consumption monitoring, score calculation, and flow adjustment, achieving intelligent management of the heat dissipation process. This reduces human error and cost, and further ensures the stability and reliability of the computing device.

[0031] Furthermore, in a specific implementation, in the traffic scheduling method provided in the embodiments of the present invention, step S101, obtaining the video memory access frequency during the execution of the parallel computing function, may specifically include: when the parallel computing function starts execution, activating the monitoring code deployed in the computing device driver; using the monitoring code to capture video memory access operations and recording the time points when the operations occur; the access operations include read operations and write operations; using a sliding time window algorithm of a preset duration to count the total number of video memory accesses within the time window, and dividing the total number of video memory accesses by the preset duration to obtain the video memory access frequency.

[0032] In implementation, monitoring code can first be embedded in the computing device driver. This monitoring code is activated only when the parallel computing function starts, capturing and recording the real-time timing of memory read and write operations. This monitoring code can be lightweight, enabling non-intrusive monitoring of memory access operations during the execution of the parallel computing function. Then, a sliding window algorithm of preset duration is used to count the total number of memory accesses within the window. Finally, the memory access frequency is calculated based on the preset window duration. For example, if the preset duration is set to 1 second, and 1 million accesses are detected within a 1-second window, the frequency is 1 million accesses per second. This provides accurate and real-time data support for subsequent thermal load scoring and heat dissipation control. Monitoring is activated only during the execution of parallel computing functions, avoiding unnecessary resource occupation during non-computation periods. It also accurately covers both memory read and write operations and records time points, ensuring the timeliness and accuracy of data collection. Furthermore, the sliding time window algorithm, combined with preset duration calculations of access frequency, dynamically reflects the actual workload intensity of the memory at different times, providing key parameter inputs that closely match the actual heat dissipation state for thermal load scoring calculations. This helps the subsequent cooling system more accurately match coolant flow, avoiding untimely or excessive heat dissipation due to data lag. In practical applications, adaptive monitoring strategies can be designed for computing devices with different architectures, ensuring compatibility with multiple hardware platforms.

[0033] Furthermore, in a specific implementation, in the above-mentioned traffic scheduling method provided in the embodiments of the present invention, step S102 uses other parameters related to heat consumption as auxiliary parameters, assigns weights to the main parameters and auxiliary parameters, and determines the heat load score result. Specifically, it may include: obtaining the actual operating temperature of the video memory and the load data of the computing device related to heat consumption, and using the actual operating temperature of the video memory and the load data of the computing device as auxiliary parameters; assigning corresponding weights to the obtained video memory access frequency, actual operating temperature of the video memory, and load data of the computing device; after assigning weights, inputting the obtained video memory access frequency, actual operating temperature of the video memory, and load data of the computing device into the pre-constructed heat load score model; running the heat load score model, calculating the heat load score of the obtained video memory access frequency, actual operating temperature of the video memory, and load data of the computing device by weighted summation, and outputting the heat load score result.

[0034] In implementation, a multi-dimensional parameter fusion thermal load scoring model is constructed, focusing on memory access frequency while comprehensively considering other factors that significantly impact the heat dissipation of the computing device's memory, such as the actual operating temperature of the memory and the overall load of the computing device. To reflect the different degrees of influence of each factor on the thermal load, appropriate weights are assigned to each factor. This invention calculates the thermal load score using a weighted summation method, with the specific formula: Thermal Load Score = a × Memory Access Frequency + b × Memory Operating Temperature + c × Overall Load of Computing Device. Here, a, b, and c are the weights corresponding to memory access frequency, memory operating temperature, and overall load of the computing device, respectively, and a + b + c = 1. These weight values ​​are derived through extensive experiments and data analysis to ensure the accuracy and reliability of the model. For example, after multiple experimental verifications, in specific computing device application scenarios, setting a = 0.6, b = 0.3, and c = 0.1 can accurately reflect the thermal load situation. This weighted calculation method ensures that the contribution of each factor to the heat load is appropriate for the actual scenario, and the final output score can serve as a reliable basis for subsequent heat dissipation control.

[0035] Furthermore, this invention can categorize and statistically analyze the frequency of different access and calculation methods of computing devices, and assign differentiated calorific value weights based on the actual calculation volume and heat generation differences (e.g., assigning higher weights to multiplication operations). A weighted formula (e.g., operation frequency = 1 × addition operation volume + 10 × multiplication operation volume) is then used to calculate the score, allowing the heat load score to more closely reflect the actual heat consumption pattern. This method can accurately match the actual heat generation differences of the device under different operating states such as addition and multiplication, providing a more precise quantitative basis for subsequent heat load assessment and heat dissipation control, thereby improving the adaptability of the heat dissipation solution and the temperature control efficiency.

[0036] Furthermore, in a specific implementation, in the flow scheduling method provided in the embodiments of the present invention, before executing step S103 to adjust the flow distribution coefficient of the corresponding cold plate microchannel based on the heat load score, it may further include: determining the initial flow rate of the cold plate microchannel based on the heat dissipation required by the computing device in a non-working state; wherein, after the computing device is started, when the computing device is not involved in the calculation, the liquid cooling system operates at the initial flow rate; when the computing device starts to participate in the calculation, the opening degree is adjusted based on the opening degree corresponding to the initial flow rate.

[0037] In implementation, before adjusting the flow distribution coefficient of the cold plate microchannel based on the heat load score, the initial flow rate is first determined based on the heat dissipation of the computing device in its non-operating state. The initial flow rate is determined based on the heat dissipation required by the system in its non-operating state (when the computing device is not involved in the calculation). That is, the liquid cooling system operates at the initial flow rate when the device is started but not calculating, and the opening degree is adjusted based on the initial flow rate during calculation. This avoids energy waste caused by excessive flow rate when the device is not calculating, and also provides a stable starting point for the opening degree adjustment during calculation, preventing sudden changes in flow rate from affecting the stability of heat dissipation.

[0038] Furthermore, in a specific implementation, in the above-mentioned flow scheduling method provided in the embodiments of the present invention, step S103, based on the heat load scoring result, adjusts the flow allocation coefficient of the corresponding cold plate microchannel, which may specifically include: comparing the heat load scoring result with a preset scoring threshold; determining the memory heat dissipation status based on the comparison result between the heat load scoring result and the preset scoring threshold; and adjusting the flow allocation coefficient of the corresponding cold plate microchannel according to the memory heat dissipation status.

[0039] In practice, this invention determines the heat dissipation status of the video memory by comparing the heat load score with a preset score threshold, and then adjusts the flow distribution coefficient of the cold plate microchannel accordingly. This allows for a quick and accurate determination of whether the heat dissipation of the video memory is high, low, or normal, avoiding blind adjustments and ensuring that the flow distribution coefficient adjustment always matches the actual heat dissipation requirements. This ensures that heat dissipation can be enhanced in a timely manner when the heat dissipation is high, while preventing energy waste when the heat dissipation is low.

[0040] In specific implementation, the heat load scoring result is compared with the preset scoring threshold in the above steps. Specifically, this may include: writing the heat load scoring result into the corresponding storage area of ​​the extended register according to the preset data frame format; and comparing the heat load scoring result with the preset scoring threshold stored in the extended register through the extended register.

[0041] In implementation, this invention utilizes an extended register to input the calculated heat load score as an input parameter into the corresponding algorithm. This extended register can be a high-speed serial bus extension (Peripheral Component Interconnect Express, PCIe) register. This invention enables high-speed data exchange between the computing device and the liquid cooling system based on the communication protocol of the PCIe BAR (Base Address Register) space. The register contains a mapping table that stores the heat load score results, the values ​​of primary and secondary parameters, and the corresponding weights of the primary and secondary parameters, which can be mapped to the flow control valve opening value.

[0042] This invention writes the heat load score results into the corresponding storage area of ​​the extended register according to a preset data frame format, and compares them with the pre-stored threshold through the register. This not only avoids deviations in data transmission and storage through a standardized format, but also enables direct interaction between the score and the threshold by relying on the extended register, reducing the time consumed in intermediate steps and quickly providing an accurate basis for subsequent determination of the memory heat consumption status and adjustment of the flow distribution coefficient.

[0043] Furthermore, in specific implementation, in the above steps, the memory heat dissipation status is determined based on the comparison between the heat load score result and the preset score threshold. Specifically, this may include: when the heat load score result is higher than the preset score threshold, determining that the memory heat dissipation status exceeds the upper limit of the baseline range. Accordingly, the flow distribution coefficient of the corresponding cold plate microchannel is adjusted according to the memory heat dissipation status. Specifically, this may include: when the memory heat dissipation status exceeds the upper limit of the baseline range, increasing the flow distribution coefficient of the corresponding cold plate microchannel.

[0044] In implementation, when the heat load score exceeds a preset threshold, it is determined that the memory heat dissipation exceeds the baseline limit. Accordingly, the flow distribution coefficient of the corresponding cold plate microchannel is increased, allowing more coolant to flow to that area, rapidly enhancing heat dissipation efficiency, effectively suppressing further increases in memory temperature, avoiding the impact of high temperatures on memory performance and hardware lifespan, and ensuring stable operation of the computing device. For example, when the heat load score exceeds the preset threshold, the flow distribution coefficient of the corresponding cold plate microchannel is increased, adjusting the flow opening of a certain cold plate microchannel from 30% to 50%, thereby increasing the coolant flow to that microchannel by approximately 67%.

[0045] Furthermore, in specific implementation, in the above steps, based on the comparison results, the memory heat dissipation status is determined. Specifically, this may include: when the heat load score result is lower than a preset score threshold, determining that the memory heat dissipation status is below the lower limit of the benchmark range. Accordingly, based on the memory heat dissipation status, the flow distribution coefficient of the corresponding cold plate microchannel is adjusted. Specifically, this may include: when the memory heat dissipation status is below the lower limit of the benchmark range, decreasing the flow distribution coefficient of the corresponding cold plate microchannel.

[0046] In practice, when the heat load score is lower than the preset score threshold, it is determined that the heat consumption of the memory is lower than the baseline lower limit. Accordingly, the flow distribution coefficient of the corresponding cold plate microchannel is reduced. This can reduce the flow of coolant in that area in a timely manner, avoid unnecessary circulation of coolant and energy waste, and at the same time ensure that the memory temperature is still maintained within the rated operating range. This improves the energy utilization efficiency of the heat dissipation system while ensuring stable operation of the equipment.

[0047] This invention dynamically adjusts the flow distribution coefficient of the cold plate microchannels, enabling the liquid cooling system to closely follow the real-time heat dissipation changes of the computing device's memory and precisely control the flow distribution of coolant within the microchannels. For memory areas with high heat dissipation, the coolant flow rate is increased to quickly remove heat and lower the temperature; for areas with low heat dissipation, the coolant flow rate is reduced to optimize energy utilization. In this way, regardless of the computing device's workload, it can be ensured to operate stably within a suitable temperature range, effectively improving the computing performance and reliability of the device.

[0048] Furthermore, in a specific implementation, in the above-mentioned traffic scheduling method provided in the embodiments of the present invention, obtaining the actual operating temperature of the video memory related to heat dissipation may specifically include: reading the actual operating temperature of the video memory reported by the temperature sensors deployed in each area of ​​the video memory.

[0049] Correspondingly, while determining the heat load score result, it may also include: calculating the deviation between the actual operating temperature of the video memory and the pre-stored expected temperature; if the calculated deviation is not within the preset temperature deviation range, the weight adjustment mechanism of the heat load score model is triggered.

[0050] In implementation, this invention can install temperature sensors in the video memory area to monitor the actual temperature of the video memory in real time and feed the actual temperature data back to the thermal load scoring model. When the actual temperature deviates from the expected temperature pre-stored according to the thermal load scoring model by more than a certain range, the weights in the thermal load scoring model are dynamically adjusted to improve the accuracy of the thermal load scoring, thereby making the flow allocation more precise. The expected temperature can be referenced to the optimal operating temperature range of the computing device; if it exceeds this range, it is considered a deviation. By reading the actual operating temperature data from the temperature sensors in each area of ​​the video memory, and simultaneously calculating the deviation between the actual and expected temperatures while running the thermal load scoring model to calculate the score, if the deviation exceeds a preset range, the model weight adjustment is triggered. This ensures that the thermal load scoring model always closely matches the actual heat dissipation state of the video memory, ensuring more precise subsequent flow control based on the score and avoiding insufficient or excessive heat dissipation due to model errors.

[0051] Furthermore, in specific implementation, if the calculated deviation value is not within the preset temperature deviation range in the above steps, the weight adjustment mechanism of the heat load scoring model is triggered. Specifically, this may include: if the calculated deviation value is higher than the upper limit of the preset temperature deviation range, the weight of the actual operating temperature of the video memory in the heat load scoring model is increased; if the calculated deviation value is lower than the lower limit of the preset temperature deviation range, the weight of the actual operating temperature of the video memory in the heat load scoring model is decreased.

[0052] In implementation, when the deviation between the actual and expected temperatures exceeds the upper limit of the preset range, the weight of temperature in the scoring model is increased; conversely, when the deviation is below the lower limit, its weight is decreased. For example, when the actual temperature is more than 3°C higher than the expected temperature, the weight of the memory operating temperature in the heat load scoring model is appropriately increased. This allows the model to more sensitively reflect temperature changes, thereby adjusting the flow distribution coefficient more quickly and reducing the memory temperature. This makes the model more responsive to real temperature changes, avoids scoring distortion caused by fixed weights, ensures that the heat load assessment is more closely aligned with actual heat consumption, and further improves the accuracy of memory temperature control.

[0053] Furthermore, in a specific implementation, in the above-mentioned flow scheduling method provided in the embodiments of the present invention, after triggering the weight adjustment mechanism of the thermal load scoring model, it may further include: when the weight of the actual operating temperature of the video memory in the thermal load scoring model increases, if the thermal load scoring result is still higher than the preset scoring threshold, shortening the adjustment cycle of the flow allocation coefficient; if the thermal load scoring result is still lower than the preset scoring threshold, extending the adjustment cycle of the flow allocation coefficient.

[0054] In implementation, the weight of the actual operating temperature of the video memory in the scoring model is increased. If the heat load score is still higher than the threshold, the flow distribution coefficient adjustment cycle is shortened; if it is still lower than the threshold, the cycle is extended. This makes the heat dissipation response more adaptable to temperature-sensitive scenarios. When the heat consumption is still high after the temperature weight is increased, the heat dissipation is accelerated by faster adjustment to avoid the temperature from rising continuously. When the heat consumption is low, the ineffective energy consumption is reduced by slowing down the adjustment. This balances the timeliness of temperature control and energy utilization efficiency, further ensuring the stable operation of the video memory.

[0055] Furthermore, in a specific implementation, in the flow scheduling method provided in the embodiments of the present invention, step S104 obtains the flow control valve opening value according to the adjusted flow allocation coefficient and sends the corresponding flow scheduling instruction to the liquid cooling system. Specifically, it may include: writing the adjusted flow allocation coefficient into the corresponding storage area of ​​the extended register according to a preset data frame format; obtaining the flow control valve opening value corresponding to the adjusted flow allocation coefficient according to the mapping relationship between the flow allocation coefficient and the flow control valve opening stored in the extended register; accessing the extended register, reading the flow control valve opening value stored in the extended register, and generating the corresponding flow scheduling instruction based on the flow control valve opening value.

[0056] In implementation, this invention can write the adjusted flow distribution coefficient into the corresponding area of ​​the extended register according to a preset data frame format. Based on the mapping relationship between the distribution coefficient and valve opening stored in the extended register, the opening value is obtained. Then, the opening value is read to generate a flow scheduling command, ensuring data accuracy and guaranteeing that the opening value precisely corresponds to the control requirements. Directly reading register data to generate commands reduces conversion steps, providing reliable support for the precise execution of flow adjustments in the liquid cooling system. The extended register of this invention can also store the mapping relationship between heat load scores and flow control valve openings. The flow control valve opening is obtained based on the heat load score results, and then the opening value is read to generate a flow scheduling command, further improving the response rate.

[0057] It should be added that the extended registers (such as PCIe registers) used in this invention store the communication parameters of each module. A hybrid storage architecture of Static Random-Access Memory (SRAM) and Electrically Erasable Programmable Read-Only Memory (EEPROM) can be selected to balance data real-time performance and power-off persistence. Specifically, SRAM: 16KB, used to store real-time status registers (lost when power is off); EEPROM: 8KB, used to store fixed configuration registers (retained when power is off). The extended registers can be soldered to the computing device. The cold plate can be tightly attached to the computing device's memory chips using thermal grease, and then connected to a liquid cooling circulation system. Furthermore, this invention can add a monitoring circuit between the computing device's memory controller and the data bus. This circuit uses a high-speed data acquisition chip with a sampling rate of up to 4GSPS, capable of accurately capturing the timing and data characteristics of each memory access cycle. The monitoring circuit preprocesses data using a Field Programmable Gate Array (FPGA) to extract access type, address, and timestamp information, and then transmits it to the system controller or extended registers via a high-speed peripheral component interconnect interface. This allows for precise capture of memory access timing and characteristics, providing accurate data support for scenarios such as thermal load assessment, and ensuring the efficiency and reliability of storage and memory monitoring in computing devices.

[0058] Additionally, it should be noted that the monitoring code can be written in NVIDIA PTX (Parallel ThreadExecution) intermediate language and injected into the computing device's kernel space through a parallel computing function driver interface. The monitoring code can be compiled into a kernel module, the thermal load scoring model deployed as a system service, and the flow control algorithm embedded in the liquid cooling controller firmware. The monitoring code mainly comprises three functional modules: an access interception module, a counter management module, and a data transmission module. The access interception module uses inline assembly technology to capture video memory access instructions; the counter management module maintains multiple time window counters; and the data transmission module sends the statistical results to the host computer through a high-speed peripheral component interconnect interface. This ensures both the real-time performance and accuracy of the monitoring, while also improving system integration efficiency through modular design, providing efficient and reliable technical support for thermal load assessment and flow control.

[0059] The traffic scheduling method provided in this embodiment of the invention was tested on a computing device using a deep learning training task as the load. The deep learning training task can use a Bidirectional Encoder Representations from Transformers (BERT) model for training. Test results show that, compared with a traditional fixed-flow system, this invention reduces memory temperature by 12°C and improves heat dissipation efficiency by 35% under high load; and reduces energy consumption by 28% under low load.

[0060] It should be noted that, compared with traditional fixed-flow liquid cooling systems, this invention can dynamically adjust the coolant flow rate based on the real-time heat dissipation of the computing device's memory, achieving precise heat dissipation for the memory area. Experimental data shows that, under the same high-load operating conditions, the method of this invention can reduce the temperature of the computing device's memory by 10°C to 15°C, improving heat dissipation efficiency by 30% to 40%, effectively avoiding performance degradation caused by excessive temperature. When the heat dissipation of the computing device's memory is low, the method of this invention can automatically reduce the coolant flow rate, avoiding unnecessary coolant circulation and energy waste. Actual testing shows that under low-load conditions, energy consumption can be reduced by 20% to 30% compared to traditional systems, helping to reduce operating costs in large-scale computing applications such as data centers.

[0061] Furthermore, this invention precisely controls the temperature of the computing device's memory, keeping it within a suitable operating range, effectively reducing the damage to the computing device hardware caused by high temperatures, thereby extending the lifespan of the computing device. Based on relevant research and practical application experience, the method of this invention can extend the mean time between failures (MTBF) of computing devices by 20% to 30%, reducing the frequency of equipment maintenance and replacement.

[0062] The method of this invention is not limited to a specific model or type of GPU. By appropriately adjusting the monitoring code and thermal load scoring model, it can be adapted to computing devices with different architectures and performance levels. Furthermore, for different application scenarios, such as deep learning training, big data analysis, and graphics rendering, it can also achieve good heat dissipation effects by optimizing parameters such as weights, demonstrating broad applicability.

[0063] It should be noted that, in addition to its application in heat dissipation for video memory in computing devices, this invention can also be applied to heat dissipation for other chips or electronic components with similar heat dissipation characteristics, such as FPGAs and Central Processing Units (CPUs). When applied to CPUs, the monitoring target is simply changed from the frequency of video memory access during parallel function execution to the frequency of cache access during CPU instruction execution. A corresponding thermal load scoring model is then constructed, and the flow distribution coefficient of the cold plate microchannel is dynamically adjusted using an extended interface in a similar manner to achieve precise heat dissipation for the CPU.

[0064] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0065] Embodiments of the present invention also provide a traffic scheduling device. From the perspective of functional modules, this embodiment of the device includes: The monitoring module is used to obtain the frequency of video memory accesses during the execution of parallel computing functions; The model running module is used to take the acquired memory access frequency as the main parameter and other parameters related to heat consumption as auxiliary parameters. After assigning weights to the main parameter and auxiliary parameters, the heat load score result is determined. The coefficient adjustment module is used to adjust the flow distribution coefficient of the corresponding cold plate microchannel based on the heat load score results; The instruction sending module is used to obtain the opening value of the flow control valve according to the adjusted flow distribution coefficient, and send the corresponding flow scheduling instruction to the liquid cooling system to control the liquid cooling system to adjust the coolant flow of each cold plate microchannel according to the opening value of the flow control valve, so that each area of ​​the memory is within the rated temperature range.

[0066] In the flow scheduling device provided in this embodiment of the invention, the interaction of the four modules allows for the acquisition of the memory access frequency during parallel computing function execution as the primary parameter. This is combined with other auxiliary parameters related to heat consumption and weighted accordingly to determine the heat load score. This, in turn, adjusts the flow distribution coefficient of the cold plate microchannel, obtains the flow control valve opening value, and sends a flow scheduling command to the liquid cooling system. This dynamically adjusts the coolant flow rate of each cold plate microchannel, enabling precise temperature control and heat dissipation of the computing device's memory area. On the one hand, it allows for flexible adaptation of the coolant flow rate based on the real-time heat consumption of the memory, avoiding unnecessary energy consumption caused by unnecessary coolant circulation. This significantly improves heat dissipation efficiency and reduces energy consumption. On the other hand, the entire adjustment process is responsive, with the entire process from monitoring heat dissipation changes to completing flow adjustment within milliseconds. It can promptly address situations where heat dissipation rises rapidly, effectively preventing a sharp increase in temperature in a short period of time, ensuring that the video memory is always within the rated operating temperature range, reducing damage to hardware from high temperatures, and extending the lifespan of computing devices. In addition, the device can autonomously complete the entire process of video memory heat dissipation monitoring, scoring calculation, and flow adjustment, realizing intelligent management of the heat dissipation process. This not only reduces human error and costs but also further ensures the stability and reliability of computing device operation.

[0067] Since the embodiments of the traffic scheduling device and the traffic scheduling method correspond to each other, the descriptions of the features in the embodiments corresponding to the traffic scheduling device can be found in the relevant descriptions of the embodiments corresponding to the traffic scheduling method, and will not be repeated here. Furthermore, it has the same beneficial effects as the traffic scheduling method mentioned above.

[0068] Furthermore, in a specific implementation, in the traffic scheduling device provided in the embodiments of the present invention, the monitoring module can be specifically used to activate the monitoring code deployed in the computing device driver when the parallel computing function starts execution; use the monitoring code to capture the access operation of the video memory and record the time point when the operation occurs; the access operation includes read operation and write operation; use a sliding time window algorithm of preset duration to count the total number of video memory accesses within the time window, and divide the total number of video memory accesses by the preset duration to obtain the video memory access frequency.

[0069] Furthermore, in a specific implementation, in the traffic scheduling device provided in the embodiments of the present invention, the model running module can be specifically used to acquire the actual operating temperature of the video memory and the load data of the computing device related to heat consumption, and use the actual operating temperature of the video memory and the load data of the computing device as auxiliary parameters; assign corresponding weights to the acquired video memory access frequency, actual operating temperature of the video memory and the load data of the computing device; after assigning weights, input the acquired video memory access frequency, actual operating temperature of the video memory and the load data of the computing device into the pre-built heat load scoring model; run the heat load scoring model, calculate the heat load score of the acquired video memory access frequency, actual operating temperature of the video memory and the load data of the computing device by weighted summation, and output the heat load score result.

[0070] Furthermore, in a specific implementation, in the flow scheduling device provided in the embodiments of the present invention, the coefficient adjustment module can be used to compare the heat load score result with the preset score threshold; determine the memory heat consumption status based on the comparison result between the heat load score result and the preset score threshold; and adjust the flow distribution coefficient of the corresponding cold plate microchannel according to the memory heat consumption status.

[0071] Furthermore, in a specific implementation, in the flow scheduling device provided in the embodiments of the present invention, the instruction sending module can be specifically used to write the adjusted flow allocation coefficient into the corresponding storage area of ​​the extended register according to a preset data frame format; obtain the flow control valve opening value corresponding to the adjusted flow allocation coefficient according to the mapping relationship between the flow allocation coefficient and the flow control valve opening value stored in the extended register; access the extended register, read the flow control valve opening value stored in the extended register, and generate the corresponding flow scheduling instruction based on the flow control valve opening value.

[0072] Embodiments of the present invention also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above-described traffic scheduling method embodiments.

[0073] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above-described traffic scheduling method embodiments when it is run.

[0074] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0075] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above-described traffic scheduling method embodiments.

[0076] Embodiments of the present invention also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in any of the above-described traffic scheduling method embodiments.

[0077] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0078] The foregoing has provided a detailed description of the traffic scheduling method, device, medium, and program product provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only intended to help understand the method and core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of this invention.

Claims

1. A method of flow scheduling, the method comprising: The method comprises the following steps: obtaining the frequency of accessing the GPU memory when the parallel computing function is executed; using the obtained frequency of accessing the GPU memory as a main parameter, and using other parameters related to heat consumption as auxiliary parameters, assigning weights to the main parameter and the auxiliary parameters, and determining a thermal load score result; based on the thermal load score result, adjusting the flow distribution coefficient of the corresponding cold plate microchannel; according to the adjusted flow distribution coefficient, obtaining the opening value of the flow control valve, and sending the corresponding flow scheduling instruction to the liquid cooling system to control the liquid cooling system to adjust the cooling liquid flow of each cold plate microchannel according to the opening value of the flow control valve, so that each region of the GPU is within the rated temperature range.

2. The traffic scheduling method of claim 1, wherein, obtaining the frequency of accessing the GPU memory when the parallel computing function is executed, comprising: when the parallel computing function is started, activate the monitoring code deployed in the driver of the computing device; using the monitoring code to capture the access operation of the GPU memory, and recording the time point when the operation occurs; the access operation includes read operation and write operation; using a preset sliding time window algorithm to count the total number of GPU memory accesses within the time window, and dividing the total number of GPU memory accesses by the preset time length to obtain the frequency of accessing the GPU memory.

3. The traffic scheduling method of claim 1, wherein, using other parameters related to heat consumption as auxiliary parameters, assigning weights to the main parameter and the auxiliary parameters, and determining a thermal load score result, comprising: obtaining the actual working temperature of the GPU memory and the load data of the computing device related to heat consumption, and using the actual working temperature of the GPU memory and the load data of the computing device as auxiliary parameters; assigning corresponding weights to the obtained GPU memory access frequency, GPU memory actual working temperature and computing device load data; after assigning weights, input the obtained GPU memory access frequency, GPU memory actual working temperature and computing device load data into a pre-constructed thermal load scoring model; running the thermal load scoring model, and calculating the thermal load score of the obtained GPU memory access frequency, GPU memory actual working temperature and computing device load data by weighted summation, and outputting the thermal load score result.

4. The traffic scheduling method of claim 3, wherein, Before adjusting the flow distribution coefficient of the corresponding cold plate microchannel based on the thermal load score result, it further comprises: determining the initial flow of the cold plate microchannel according to the heat dissipation required by the computing device in the non-working state; wherein, after the computing device is started, when the computing device does not participate in calculation, the liquid cooling system operates at the initial flow; when the computing device starts to participate in calculation, the opening is adjusted based on the opening corresponding to the initial flow.

5. The method of flow scheduling according to claim 4, wherein, based on the thermal load score result, adjusting the flow distribution coefficient of the corresponding cold plate microchannel, comprising: comparing the thermal load score result with a preset score threshold; based on the comparison result between the thermal load score result and the preset score threshold, determining the GPU heat consumption state; adjusting the flow distribution coefficient of the corresponding cold plate microchannel according to the GPU heat consumption state.

6. The traffic scheduling method of claim 5, wherein, based on the comparison result between the thermal load score result and the preset score threshold, determining the GPU heat consumption state, comprising: when the thermal load score result is higher than the preset score threshold, it is determined that the GPU heat consumption state exceeds the upper limit of the reference range; Adjusting a flow distribution coefficient of a corresponding cold plate microchannel according to the GPU heat consumption state, including: When the GPU heat consumption state exceeds the upper limit of the reference range, increasing the flow distribution coefficient of the corresponding cold plate microchannel.

7. The method of traffic scheduling according to claim 6, wherein, Determining the GPU heat consumption state based on the comparison result, including: When the thermal load score result is lower than the preset score threshold, determining that the GPU heat consumption state is lower than the lower limit of the reference range; Adjusting a flow distribution coefficient of a corresponding cold plate microchannel according to the GPU heat consumption state, including: When the GPU heat consumption state is lower than the lower limit of the reference range, decreasing the flow distribution coefficient of the corresponding cold plate microchannel.

8. The method of traffic scheduling according to claim 7, wherein, Obtaining a GPU actual working temperature related to heat consumption, including: Reading a GPU actual working temperature reported by a temperature sensor deployed in each region of the GPU; When determining the thermal load score result, further including: Calculating a deviation value between the GPU actual working temperature and a pre-stored expected temperature; If the calculated deviation value is not within a preset temperature deviation range, triggering a weight adjustment mechanism of the thermal load scoring model.

9. The method of traffic scheduling according to claim 8, wherein, If the calculated deviation value is not within a preset temperature deviation range, triggering a weight adjustment mechanism of the thermal load scoring model, including: If the calculated deviation value is higher than an upper limit of the preset temperature deviation range, increasing the weight of the GPU actual working temperature in the thermal load scoring model; If the calculated deviation value is lower than a lower limit of the preset temperature deviation range, decreasing the weight of the GPU actual working temperature in the thermal load scoring model.

10. The traffic scheduling method of claim 9, wherein, After triggering the weight adjustment mechanism of the thermal load scoring model, further including: When the weight of the GPU actual working temperature in the thermal load scoring model is increased, if the thermal load score result is still higher than the preset score threshold, shortening an adjustment period of the flow distribution coefficient; if the thermal load score result is still lower than the preset score threshold, lengthening the adjustment period of the flow distribution coefficient.

11. The method of traffic scheduling according to claim 5, wherein, Comparing the thermal load score result with a preset score threshold, including: Writing the thermal load score result into a corresponding storage area of an extension register according to a preset data frame format; Comparing the thermal load score result with a preset score threshold pre-stored in the extension register through the extension register.

12. The method of traffic scheduling according to claim 1, wherein, According to the adjusted flow distribution coefficient, obtaining a flow control valve opening value, and sending a corresponding flow scheduling instruction to a liquid cooling system, including: Writing the adjusted flow distribution coefficient into a corresponding storage area of an extension register according to a preset data frame format; According to a mapping relationship between the flow distribution coefficient and the flow control valve opening stored in the extension register, obtaining a flow control valve opening value corresponding to the adjusted flow distribution coefficient; Accessing the extension register, reading the flow control valve opening value stored in the extension register, and generating a corresponding flow scheduling instruction based on the flow control valve opening value.

13. An electronic device, comprising: Including: A memory for storing a computer program; A processor for executing the computer program to implement the steps of the flow scheduling method according to any one of claims 1 to 12.

14. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program, when executed by a processor, implements the steps of the traffic scheduling method according to any one of claims 1 to 12.

15. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the traffic scheduling method according to any one of claims 1 to 12.

Images