Uninterrupted operation system for microgrid-interconnected data center

Abstract

An uninterrupted operation system for a microgrid-linked data center is provided, comprising an administrator terminal that links the microgrid and the data center, a linkage unit that links the microgrid and the data center, a thermal buffer calculation unit that calculates the thermal buffer capacity by measuring the melting fraction of a Phase Change Material (PCM) placed in a liquid cooling loop within the data center when the power generation of the microgrid decreases, and an integrated control server that includes a power limiting unit that gradually lowers the frequency and voltage of the GPUs to a minimum frequency threshold where network synchronization between GPUs of at least one AI server in the data center is maintained based on the depletion rate of the thermal buffer capacity.

Description

Technology Field

[0001] The present invention relates to an uninterrupted operation system for a microgrid-linked data center, and provides a system that gradually lowers the frequency and voltage of the data center's GPUs based on the depletion rate of the thermal buffer capacity of the PCM when the power generation of the microgrid is insufficient. Background Technology

[0002] Due to the recent implementation of the Distributed Energy Act, data centers are reducing their dependence on external power grids and adopting self-generation sources such as gas generators and solar power. However, gas generators present a problem where power gaps occur during the ramp-up phase, where output is not secured immediately but gradually increases. While Energy Storage Systems (ESS) are being introduced to compensate for this, lithium-based ESSs have limitations, entailing a risk of large-scale fires and high costs. Meanwhile, although applying power capping during power shortages could be considered, data centers are composed of distributed learning structures where multiple GPUs perform computations simultaneously. This necessitates network synchronization based on the NVIDIA Collective Communication Library (NCCL), and even a frequency drop in just a few GPUs can cause the entire synchronization to collapse. Consequently, problems such as degraded learning performance or system interruptions arise, presenting a structural limitation that makes it difficult to address with simple power control methods.

[0003] At this time, methods for controlling the internal temperature of a data center or performing air conditioning of a data center using ice thermal storage linked to an ESS have been researched and developed. In this regard, prior art Korean Registered Patent No. 10-2775121 (published March 4, 2025) and Korean Registered Patent No. 10-1773946 (published September 4, 2017) disclose a configuration that prevents server overheating and fire by step-by-step adjusting the amount of liquid sprayed based on a temperature sensor and a control module to detect the internal temperature of a data center and perform cooling or fire extinguishing functions by comparing it with a set temperature, and a configuration that reduces energy consumption by utilizing nighttime electricity to store cold energy and reducing or stopping the operation rate of a chiller according to the outside temperature conditions, in order to improve the cooling efficiency of a data center by combining an ESS, an ice thermal storage system, and outside air cooling.

[0004] However, in the former case, only a simple cooling and extinguishing structure that sprays an active liquid according to temperature standards is disclosed, and a configuration for storing heat generated in the data center is not disclosed. In the latter case as well, only a configuration of an ESS-based air conditioner is disclosed, and a configuration for controlling the computational performance of the GPU in conjunction with the heat generated by the GPU when a power gap occurs is not disclosed. Thus, in a renewable energy-based microgrid environment, power gaps occur due to fluctuations in power generation, but data centers face limitations in cost and scalability due to their reliance on ESS, which is an electric battery. Furthermore, since power, heat, and computational control are separated, there is a problem in that it is difficult to simultaneously maintain uninterrupted service and verify the degradation of IT assets. Accordingly, research and development of technology capable of simultaneously achieving uninterrupted operation of microgrid-linked data centers and preventing equipment degradation is required. The problem to be solved

[0005] One embodiment of the present invention can provide an uninterrupted operation system for a microgrid-linked data center, wherein, when the power generation of the microgrid decreases, the melting fraction of a Phase Change Material (PCM) placed in a liquid cooling loop is measured to calculate the thermal buffer capacity, and based on the depletion rate of the thermal buffer capacity, the frequency and voltage of the GPUs are gradually lowered down to a minimum frequency threshold at which network synchronization between GPUs of at least one AI server in the data center is maintained, thereby preventing degradation without interrupting the operation of the GPUs until the power generation returns to normal. However, the technical problem that this embodiment aims to solve is not limited to the technical problem described above, and other technical problems may exist. means of solving the problem

[0006] As a technical means for achieving the aforementioned technical task, one embodiment of the present invention includes an integrated control server comprising: a linkage unit that links a microgrid and a data center; a thermal buffer calculation unit that calculates a thermal buffer capacity by measuring the melting fraction of a phase change material (PCM) placed in a liquid cooling loop within the data center when the power generation of the microgrid decreases; and a power limiting unit that gradually lowers the frequency and voltage of the GPUs based on the depletion rate of the thermal buffer capacity to a minimum frequency threshold at which network synchronization between GPUs of at least one AI server in the data center is maintained. Effects of the invention

[0007] According to any one of the means for solving the problem of the present invention described above, the thermal buffer capacity and depletion rate can be quantitatively calculated using the PCM-based melting progress rate, and the limit that the PCM can withstand can be determined in real time by comparing this with the power gap time. At the same time, by considering the minimum frequency threshold at which network synchronization between multiple GPUs is maintained, the GPU frequency and voltage can be gradually lowered, thereby maintaining the stability of the AI ​​server even in a power shortage situation and preventing unnecessary performance degradation or system collapse. Furthermore, uninterrupted operation of the data center is made possible without relying on the ESS, advanced energy-computing control integrating PCM and GPU control is implemented, and the effects of preventing degradation of IT assets and improving reliability can be achieved simultaneously by minimizing the accumulation of thermal stress. Brief explanation of the drawing

[0008] FIG. 1 is a drawing for explaining an uninterrupted operation system of a microgrid-linked data center according to one embodiment of the present invention. Figure 2 is a block diagram illustrating an integrated control server included in the system of Figure 1. FIG. 3 is a drawing for explaining an embodiment in which an uninterrupted operation solution according to an embodiment of the present invention is implemented. FIG. 4 is an example drawing for explaining the concept of an uninterrupted operation solution according to one embodiment of the present invention. FIG. 5 is an operation flowchart illustrating a method for providing an uninterrupted operation solution according to an embodiment of the present invention. Specific details for implementing the invention

[0009] Embodiments of the present invention are described below with reference to the attached drawings so that those skilled in the art can easily implement the invention. However, the present invention may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present invention in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification are denoted by similar reference numerals.

[0010] Throughout the specification, when a part is described as being "connected" to another part, this includes not only cases where they are "directly connected" but also cases where they are "electrically connected" with other elements interposed between them. Furthermore, when a part is described as "including" a component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components, and it should be understood that this does not preclude the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0011] Terms such as “about,” “substantially,” etc., used throughout the specification, are used to mean at or near the stated value when inherent manufacturing and material tolerances are presented in the stated meaning, and are used to prevent unscrupulous infringers from unfairly exploiting the disclosure in which precise or absolute values ​​are mentioned to aid in understanding the invention. Terms such as “step” or “step of” used throughout the specification of the invention do not mean “step for”.

[0012] In this specification, the term "part" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. Additionally, one unit may be realized using two or more pieces of hardware, and two or more units may be realized by one piece of hardware. Meanwhile, "part" is not limited to software or hardware, and "part" may be configured to reside in an addressable storage medium or configured to run on one or more processors. Accordingly, as an example, "part" includes components such as software components, object-oriented software components, class components, and task components, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functions provided within the components and "parts" may be combined into a smaller number of components and "parts" or further separated into additional components and "parts." In addition, the components and '~parts' may be implemented to play one or more CPUs within the device or secure multimedia card.

[0013] Some of the operations or functions described herein as being performed by a terminal, device, or device may instead be performed by a server connected to said terminal, device, or device. Likewise, some of the operations or functions described as being performed by a server may also be performed by a terminal, device, or device connected to said server.

[0014] In this specification, some of the operations or functions described as mapping or matching with a terminal may be interpreted as meaning mapping or matching the terminal's unique number or personal identification information, which is the terminal's identifying data.

[0015] The present invention will be described in detail below with reference to the attached drawings.

[0016] FIG. 1 is a diagram illustrating an uninterrupted operation system for a microgrid-linked data center according to an embodiment of the present invention. Referring to FIG. 1, the uninterrupted operation system (1) for a microgrid-linked data center may include at least one administrator terminal (100), an integrated control server (300), at least one microgrid (400), and at least one data center (500). However, since the uninterrupted operation system (1) for a microgrid-linked data center of FIG. 1 is merely an embodiment of the present invention, the present invention is not to be interpreted as being limited by FIG. 1.

[0017] At this time, each component of FIG. 1 is generally connected through a network (Network, 200). For example, as shown in FIG. 1, at least one manager terminal (100) can be connected to an integrated control server (300) through the network (200). And, the integrated control server (300) can be connected to at least one manager terminal (100), at least one microgrid (400), and at least one data center (500) through the network (200). Also, at least one microgrid (400) can be connected to the integrated control server (300) through the network (200). And, at least one data center (500) can be connected to at least one manager terminal (100), the integrated control server (300), and at least one microgrid (400) through the network (200).

[0018] Here, a network refers to a connection structure capable of exchanging information among individual nodes, such as multiple terminals and servers. Examples of such networks include Local Area Networks (LANs), Wide Area Networks (WANs), the World Wide Web (WWW), wired and wireless data networks, telephone networks, and wired and wireless television networks. Examples of wireless data communication networks include, but are not limited to, 3G, 4G, 5G, 3GPP (3rd Generation Partnership Project), 5GPP (5th Generation Partnership Project), 5G NR (New Radio), 6G (6th Generation of Cellular Networks), LTE (Long Term Evolution), WIMAX (World Interoperability for Microwave Access), Wi-Fi, Internet, LAN (Local Area Network), Wireless LAN (Wireless Local Area Network), WAN (Wide Area Network), PAN (Personal Area Network), RF (Radio Frequency), Bluetooth network, NFC (Near-Field Communication) network, satellite broadcasting network, analog broadcasting network, DMB (Digital Multimedia Broadcasting) network, etc.

[0019] In the following, the term "at least one" is defined as a term including both singular and plural forms, and it will be obvious that even if the term "at least one" does not exist, each component may exist in a singular or plural form and may mean singular or plural. Furthermore, whether each component is provided in a singular or plural form may be changed according to the embodiment.

[0020] At least one administrator terminal (100) may be a terminal of an administrator that links the microgrid (400) and the data center (500) using a web page, app page, program, or application related to the non-stop operation solution.

[0021] Here, at least one administrator terminal (100) may be implemented as a computer capable of connecting to a remote server or terminal via a network. Here, the computer may include, for example, a navigation system, a laptop equipped with a web browser, a desktop, a laptop, etc. At this time, at least one administrator terminal (100) may be implemented as a terminal capable of connecting to a remote server or terminal via a network. At least one manager terminal (100) may include all kinds of handheld-based wireless communication devices, such as navigation, PCS (Personal Communication System), GSM (Global System for Mobile communications), PDC (Personal Digital Cellular), PHS (Personal Handyphone System), PDA (Personal Digital Assistant), IMT (International Mobile Telecommunication)-2000, CDMA (Code Division Multiple Access)-2000, W-CDMA (W-Code Division Multiple Access), Wibro (Wireless Broadband Internet) terminal, smartphone, smartpad, tablet PC, etc.

[0022] The integrated control server (300) may be a server that provides a web page, app page, program, or application for an uninterrupted operation solution. Additionally, the integrated control server (300) may be a server that monitors the power generation of the microgrid (400) when the microgrid (400) and the data center (500) are linked at the administrator terminal (100), and when the power generation decreases or a ramp-up period occurs, measures the melting fraction of the phase change material (PCM) placed in the liquid cooling loop within the data center (500) to calculate the thermal buffer capacity, and based on the depletion rate of the thermal buffer capacity, gradually lowers the frequency and voltage of the GPUs to a minimum frequency threshold where network synchronization between GPUs of at least one AI server in the data center (500) is maintained. Additionally, the integrated control server (300) may be a server that, after predicting the power gap time, if the power gap time is greater than the thermal buffer time calculated by dividing the thermal buffer capacity by the depletion rate, operates the inference node of the AI ​​server within the data center (500) normally, but moves the computation context of the learning node to external storage. Furthermore, the integrated control server (300) may be a server that performs measurement and verification by storing the GPU temperature profile in a distributed ledger and then inserting it as the hash value of the metadata within the AI ​​server's certificate when disposing of the AI ​​server as an IT asset, thereby proving that the AI ​​server is an asset that has not undergone degradation. Here, the integrated control server (300) may be implemented as a computer capable of connecting to a remote server or terminal via a network. Here, the computer may include, for example, a laptop, desktop, or laptop equipped with a navigation system and a web browser.

[0023] At least one microgrid (400) is a power grid that transmits generated energy to a data center (500) with or without using a web page, app page, program, or application related to an uninterrupted operation solution. In particular, it refers to a small-scale power grid that integrates the production, storage, supply, and load of power within a specific area by connecting distributed self-generation facilities, such as solar, wind, hydroelectric, or fuel-based generators, with an external power grid, and can autonomously control power supply and demand in a connected state with grid power or in an independent operation state.

[0024] At least one data center (500) may be an energy-computing integrated operating environment that receives power supplied from the microgrid (400) by being linked with the microgrid (400) with or without using a web page, app page, program, or application related to the non-disruptive operation solution, controls the frequency and voltage of the GPU according to the control of the integrated control center (300), and controls the computation context of the learning node to be moved to external storage when the power gap time is expected to be greater than the thermal buffer time.

[0025] FIG. 2 is a block diagram for explaining an integrated control server included in the system of FIG. 1, FIG. 3 is a diagram for explaining an embodiment in which an uninterrupted operation solution according to an embodiment of the present invention is implemented, and FIG. 4 is an example diagram for explaining a concept for understanding an uninterrupted operation solution according to an embodiment of the present invention.

[0026] Referring to FIG. 2, the integrated control server (300) may include an interlocking unit (310), a heat buffer calculation unit (320), a power limiting unit (330), a maintenance unit (340), and a metering certification unit (350).

[0027] When an integrated control server (300) or another server (not shown) operating in conjunction with an embodiment of the present invention transmits an uninterrupted operation solution application, program, app page, web page, etc. to at least one administrator terminal (100), at least one microgrid (400), and at least one data center (500), the at least one administrator terminal (100), at least one microgrid (400), and at least one data center (500) may install or open the uninterrupted operation solution application, program, app page, web page, etc. Additionally, a service program may be operated on at least one administrator terminal (100), at least one microgrid (400), and at least one data center (500) using a script executed in a web browser. Here, a web browser refers to a program that enables the use of web (WWW: World Wide Web) services and receives and displays hypertext described in HTML (Hyper Text Mark-up Language), and includes, for example, Chrome, Microsoft Edge, Safari, Firefox, Whale, UC Browser, etc. Additionally, an application refers to an application on a terminal, and includes, for example, an app running on a mobile terminal (smartphone).

[0028] Before explaining Fig. 2, the background of how the present invention was derived based on the basic concepts of microgrids and data centers and the recently implemented Distributed Energy Act will be explained below. The content described below will not be repeated in Fig. 2.

[0029] Microgrids and Data Centers

[0030] Referring to Fig. 4a, the microgrid is a power supply layer that supplies power to the data center, including solar, wind, and gas generators and electric batteries, and generates power gaps or fluctuation signals according to changes in power generation. The data center is an integrated operation layer that distributes power resources to each piece of equipment and manages the operational status of the entire system, including a power distribution system, a cooling system, a scheduler, and a monitoring system. The infrastructure within this data center has a layered structure as follows: [Power (hundreds of kW) → Rack (Power / Cooling) → AI Server (Computing Node) → GPU (Computation)]. To explain this sequence, first, the rack shown in Fig. 4b is an infrastructure layer that supplies power to the AI ​​server through power supply units and cooling devices and physically processes the generated heat. The AI ​​server, which is inserted horizontally into the rack as shown in Figs. 4c to 4e, is a computing node layer composed of a CPU, GPU, memory, and network interface that performs artificial intelligence tasks such as inference and learning. GPUs (the yellow object of the NVIDIA logo in the drawing is the GPU) embedded in one AI server such as Fig. 4e are devices that perform parallel computations in the AI ​​server, and are computation layers that process deep learning computations and generate heat during the computation process.

[0031] Distributed Energy Method

[0032] Recently, the government has been promoting the Distributed Energy Act, a policy aimed at decentralizing power production and consumption at the regional level rather than concentrating them in one place. The core of this policy is to encourage power-consuming facilities to generate their own electricity, rather than relying on large-scale power plants. In line with this policy shift, there is a growing number of cases where data centers are establishing self-generation systems by installing internal gas-powered generators or solar power facilities, instead of relying solely on the external power grid.

[0033] <Background>

[0034] There is a structural issue here: facilities such as gas generators do not deliver maximum output immediately upon demand, but rather have the characteristic of gradually increasing output over a certain period. This process is called "ramp-up," and during this phase, a power gap occurs where the system fails to supply sufficient power to meet the data center's requirements. To compensate for this power gap, an Energy Storage System (ESS) is commonly adopted. This device stores electricity in batteries and supplies it when needed, primarily using lithium batteries. However, lithium batteries pose a risk of thermal runaway, which can lead to fire or explosion if overheated. Furthermore, meeting the capacity requirements of large-scale data centers incurs costs amounting to tens of billions of won, placing a significant burden in terms of both safety and economic feasibility.

[0035] On the other hand, when power is insufficient, a method called power capping can be applied to simply limit the data center's power consumption. However, this method can cause critical problems in data centers. Training AI models requires multiple GPUs to perform computations simultaneously and exchange data, and the NVIDIA Collective Communications Library (NCCL) is responsible for this communication and synchronization between GPUs. In an NCCL-based architecture, if the performance of even a single GPU drops, the overall computation speed is tuned to that GPU. Therefore, if the performance of some GPUs is reduced due to power capping, the overall training speed may slow down drastically, or in severe cases, synchronization between GPUs may break, causing the training itself to proceed abnormally.

[0036] In summary, due to distributed energy policies, data centers have adopted self-generation, and while they are mandatorily introducing ESS to resolve power gaps that occur during this process, using lithium-type ESS entails high costs and fire risks, and responding with power capping results in performance collapse issues due to the AI ​​learning structure. Accordingly, in one embodiment of the present invention, power, heat, and computation are not viewed separately but are considered simultaneously. If power is insufficient, the system first sustains itself using a phase change material (PCM). If it is still insufficient, the frequency and voltage of the GPU are gradually lowered to a level that does not break the AI ​​learning structure. If this is also insufficient and it is predicted that the power gap time will exceed the thermal buffer time, the inference node is maintained while the computation context of the learning node is moved to external storage, thereby ensuring the continuity of the inference service and enabling the actual uninterrupted operation of the data center.

[0037] Based on the basic concept described above, an explanation will be provided with reference to FIGS. 2 and FIGS. 3.

[0038] Referring to (a) of FIG. 3a, the linkage unit (310) can link the microgrid (400) and the data center (500). The administrator terminal (100) can link the microgrid (400) and the data center (500). At this time, as will be described later, the first step is to check the power generation of the microgrid (400) and synchronize the operation of the data center (500) according to the power situation in order to calculate the thermal buffer capacity of the data center (500) based on the power generation of the microgrid (400), control the frequency and voltage of the GPU accordingly, and further control the transfer of the computation context of the learning node to external storage.

[0039] Referring to Fig. 3a (b), the heat buffer calculation unit (320) can calculate the heat buffer capacity by measuring the melting fraction of the Phase Change Material (PCM) placed in the liquid cooling loop within the data center (500) when the power generation of the microgrid (400) decreases. At this time, PCM refers to a material that absorbs or releases heat when its state changes, such as between solid and liquid. While ordinary materials only continue to rise in temperature as the temperature increases, PCM refers to a material that absorbs heat while maintaining its temperature almost the same once it reaches a certain temperature. The heat absorbed at this time is called latent heat. If we consider the melting of ice, it is similar to the principle that while the temperature remains the same, heat continues to be absorbed as the ice melts from 0 degrees.

[0040] Here, while various types of PCMs can be used, paraffin-based PCMs may be the most realistic option when considering the safety margin given that the normal operating temperature of a GPU is approximately 60 to 85 degrees. This is because paraffin-based PCMs allow for phase transition temperatures to be controlled between 50 and 70 degrees, and they are the most chemically stable and non-corrosive. However, it goes without saying that various other materials can also be used, such as fatty acid-based PCMs, which have high heat storage capacity and are relatively stable, and inorganic PCMs, which have very large heat storage capacity and are inexpensive.

[0041] At this time, the PCM is not simply added as a liquid, but is packaged for use, for example, in the form shown in Table 1 below. Of course, the following forms are merely embodiments of the present invention and are not limited thereto.

[0042] Integrated heatsink Cold Plate + PCM Rack-unit PCM module Liquid cooling loop interlocking type Inserting the PCM inside the GPU / CPU heatsink is the most direct method. PCM placed behind the liquid cooling plate Insert PCM cartridge inside server rack Connecting the PCM tank to the liquid cooling line Effect: Heat absorption right next to the chip Effect: Serves as a backup when cooling is delayed Effect: Full heat buffering Effect: Absorbs heat from the center

[0043] At this time, the melting progress rate indicates how much of the total PCM has already melted. To express the phase change state of the PCM in the simplest and clearest way, the melting progress rate can be defined as shown in Equation 1 below.

[0044]

[0045] M(t) is the melting progress rate of the PCM at time t, mmelt(t) is the mass of the melted PCM at time t, and mpcm is the total mass of the PCM. This equation indicates the proportion of the total PCM mass that has already melted; M(t)=0 means the entirety is in a solid state, and M(t)=1 means the entirety is in a liquid state. In other words, as M(t) increases, it means that more heat has already been absorbed.

[0046] However, mathematical formula 1 is difficult to measure in actual control. For example, with paraffin, it is not easy to take out and measure the mass of melted paraffin and the mass of unmelted paraffin from the paraffin contained in the liquid cooling loop. Accordingly, the melting progress rate can be calculated based on a temperature-based approximation formula in which the melting progress rate increases linearly as the temperature of the PCM rises within the phase change region after the temperature of the PCM is measured.

[0047]

[0048] Equation 2 calculates the melting progress rate based on the fact that the melting progress rate increases linearly as the temperature rises within the phase change range. For example, if the phase change range is between 0 and 1 and the temperature rises by 0.5 within the phase change range, the melting progress rate can be approximated as 50%.

[0049] The thermal buffer capacity can be calculated by multiplying the mass of the PCM by the latent heat already stored in the PCM, and then multiplying the result by the remainder obtained by subtracting the melting progress rate from 1. Here, latent heat refers to the intrinsic latent heat per unit mass of the PCM, which is the value listed in the PCM datasheet. In this case, the thermal buffer capacity is the capacity of heat that the PCM can absorb further. That is, it is the total amount of additional heat that can be absorbed before reaching the temperature limit within the phase change range. At this time, the phase change range is limited to the liquid stage and does not consider the gas stage. This is because transitioning to a boiling phase change is an area that must absolutely be avoided in data centers, as it not only increases pressure and degrades expensive IT assets but also poses a risk of explosion. Accordingly, it is possible to determine how much heat the PCM can absorb while reaching melting—that is, the thermal buffer capacity—which can be calculated as shown in Equation 3 below.

[0050]

[0051] Qrembuf(t) is the thermal buffer capacity at time t, mcpm is the mass of the PCM, L is the latent heat of the PCM, and M(t) is the melting progress rate at time t. Here, the latent heat is data already listed in the datasheet, and since M(t) is calculated using the temperature approximation formula described above, the thermal buffer capacity can also be derived. Accordingly, as M(t) increases, the thermal buffer capacity decreases, and as M(t) decreases, the thermal buffer capacity increases.

[0052] Referring to (c) of FIG. 3a, the power limiting unit (330) can gradually lower the frequency and voltage of the GPUs based on the depletion rate of the thermal buffer capacity to a minimum frequency threshold at which network synchronization between GPUs of at least one AI server of the data center (500) is maintained. At this time, the depletion rate of the thermal buffer capacity refers to how quickly the thermal buffer capacity decreases, and can be expressed as a mathematical formula as shown in Equation 4 below.

[0053]

[0054] Dbuf(t) is the rate of exhaustion of the column buffer capacity at time t, and Qrembuf(t) is the column buffer capacity at time t. Since the column buffer capacity decreases over time, a negative sign is attached to the rate of decrease to define a positive rate of exhaustion. Accordingly, the larger Dbuf(t) is, the faster the column buffer capacity is being consumed.

[0055] At this time, if the depletion rate is expressed using the melting progress rate, it can be expressed as shown in Equation 5 below.

[0056]

[0057] In this case, dM(t) / dt is the rate of change of the melting progress rate over time. In other words, the faster the melting progress rate increases—that is, the faster the PCM melts—the faster the thermal buffer capacity is depleted.

[0058] Meanwhile, referring to (d) of Fig. 3a, the minimum frequency threshold refers to the lowest operating frequency at which training can be maintained normally without synchronization being broken when the GPU performs training of an AI model. Training of an AI model proceeds not by a single GPU, but by multiple GPUs simultaneously performing calculations and exchanging calculation results with each other. At this time, the NVIDIA Collective Communications Library (NCCL) is used for exchanging calculation results. If the performance of the GPUs is reduced due to insufficient power, some GPUs may become slow while others become fast, causing synchronization delays and bottlenecks, which can lead to a deadlock in severe cases. It is similar to a two-legged race where no matter how well one person runs, if the other runs poorly, their feet get tangled and they fall. This is precisely the breakdown of synchronization in GPUs; however, while a two-legged race involves only two people, the situation is even more dangerous in the case of GPUs because tens to thousands of GPUs are actually connected simultaneously. Accordingly, the minimum frequency threshold can be defined as the minimum running speed that can be maintained without falling over in a two-person three-legged race.

[0059] Accordingly, the GPU frequency and voltage are gradually lowered only up to the minimum frequency threshold. The reason for lowering both frequency and voltage is that reducing only the frequency limits power reduction. In other words, GPU power is P=CV 2 f represents power consumption, capacitance, voltage, and frequency. As indicated in this formula, voltage is squared; therefore, if we assume that the power reduction with frequency is linear, the power reduction with voltage has a greater influence due to the square. On the other hand, the reason for setting the threshold only for frequency and not for voltage is that the minimum voltage threshold is automatically determined based on the minimum frequency threshold.

[0060] division Control purpose variable Power control decrease Frequency + Voltage Stability constraints maintain Frequency

[0061] <Minimum frequency threshold>

[0062] In order to obtain the minimum frequency threshold at which the synchronization of the aforementioned NCCL is not broken, the computation time of one step of GPU i is first assumed as in Equation 6 below.

[0063]

[0064] Tcomp,i(t) is the computation time of GPU i, Wi(t) is the normalized amount of computation that GPU i must process at the corresponding step, and fi(t) is the computation frequency of GPU i. Equation 6 means that for the same amount of workload, the lower the frequency, the longer the computation time.

[0065] At this time, the synchronization maintenance condition of NCCL can be set as shown in Equation 7 below.

[0066]

[0067] Tcomm,i(t) is the NCCL communication time for GPU i, and Tmaxsync is the maximum allowable step time for maintaining synchronization. In other words, this means that the total time required for each GPU—the sum of computation and communication time—must not exceed the allowable limit. If the limit is exceeded, the GPUs will wait for each other, or collective communication will become a bottleneck, causing synchronization efficiency to collapse.

[0068] When the minimum frequency threshold in the above-described mathematical formula is rearranged as in the following mathematical formula 8, and when the strictest value is selected for all of the multiple GPUs, it is derived as in mathematical formula 9.

[0069]

[0070]

[0071] fcriti is the minimum frequency threshold required for GPU i to maintain NCCL synchronization, and fcritNCCL is the minimum frequency threshold required to maintain NCCL synchronization across the entire rack. The meaning of Equation 9 is that since the minimum frequency required for each GPU may differ, the highest requirement among them must be satisfied to prevent the entire system (based on the rack unit in Equation 9) from collapsing. Accordingly, this implies that the frequency must not fall below this value.

[0072] Frequency Gradual Down-adjustment

[0073]

[0074] f(k) is the GPU frequency at control cycle k, and f(k+1) is the GPU frequency at the next control cycle. fcritNCCL is the minimum frequency threshold for maintaining NCCL at control cycle k, and Δf is the frequency reduction per control cycle. In other words, it is an equation that limits the GPU frequency to be lowered stepwise by Δf without dropping it significantly all at once, while ensuring it does not fall below the minimum frequency threshold at which NCCL is maintained.

[0075] Voltage Step Down Adjustment

[0076] As mentioned above, voltage is linked to frequency. This can be written as a general formula as shown in Equation 11 below.

[0077]

[0078] Alternatively, it can be written as a linear approximation as V(k+1)=Vmin+α(f(k+1))-fmin). Here, V(k+1) is the GPU voltage in the next control cycle, g(·) is the voltage mapping function corresponding to the frequency, Vmin is the minimum operating voltage, fmin is the minimum operating frequency, and α is the voltage change coefficient according to the change in frequency. The meaning of Equation 11 is that when the frequency is lowered, the voltage is lowered along with it to a level where safe operation is possible at that frequency.

[0079] When the frequency and voltage of the GPU are lowered in this way, the power consumption per rack is actually reduced. This can be derived as shown in Equation 12.

[0080]

[0081] Prack(k) represents the total power consumption of a rack containing multiple AI servers at control cycle k, N represents the number of GPUs, Ci represents the effective switching capacitance of GPU i, Vi(k) represents the voltage of GPU i, fi(k) represents the frequency of GPU i, and Pstatic(k) represents static leakage power and incidental power consumption. Equation 12 implies that power consumption per rack can be reduced by lowering both the frequency and the voltage. Of course, the example standard was set at the rack level, and it does not necessarily mean that power consumption is reduced only at the rack level.

[0082] Additionally, the power limiting unit (330) can increase the reduction value of the GPU's frequency and voltage as the depletion rate increases, and decrease the reduction value of the GPU's frequency and voltage as the depletion rate decreases. That is, since increasing the reduction value always can degrade the performance of the GPU, in order to prevent excessive performance degradation, it can control the heat strongly when it accumulates rapidly and reduce it gently when it accumulates slowly.

[0083]

[0084] Speed ​​of depletion ↑ Reduce more Depletion rate ↓ Less reduction

[0085] Considering all the parameters discussed so far, an objective function for finding the control point can be established as follows. In other words, the goal is to find a control point that minimizes power consumption to the available supply, reduces the depletion rate, leaves a significant amount of thermal buffer capacity remaining, and ensures the GPU frequency is neither too high nor too low compared to the NCCL threshold.

[0086]

[0087] J(k) is the optimization objective function at control period k, and w1~w4 are weights reflecting the importance of each term, with a sum of 1. Prack(k) is the power consumed by the rack at control period k, and Psup(k) is the power supplied to the rack at control period k. The other parameters are as described above.

[0088] NVIDIA NVML (Management Library) API or DCGM (Data Center GPU Manager), etc., can be used as a means to obtain the GPU parameters required in the above-described mathematical formula, and real-time telemetry data can be collected through this.

[0089] Unlike Power Capping-based Dynamic Voltage and Frequency Scaling (DVFS), which reduces frequency and voltage to limit power consumption to a certain level, one embodiment of the present invention performs synchronization constraint-based control by calculating a minimum frequency threshold that maintains network synchronization by considering the computation time and communication time for each of multiple GPUs, and setting a frequency down-limit based on the minimum frequency threshold. In particular, by setting the maximum value among the minimum frequency thresholds calculated for each GPU as a common minimum frequency at the rack level, it has a structure that prevents synchronization collapse caused by computational imbalance between GPUs, clearly distinguishing the technical challenges, control criteria, and operation methods from DVFS control that merely satisfies a power upper limit.

[0090] Differences and characteristics from power capping-based DVFS 1. Synchronization standard, not power standard 2. Calculate minimum frequency threshold per GPU 3. Set the maximum value among them as the lag standard

[0091] division Power capping-based DVFS The present invention Control criteria Adjusts frequency and voltage based on the target power upper limit Control is achieved by comprehensively considering thermal buffer status, power gap duration, and inter-GPU synchronization conditions. Control objectives Limit power consumption to below a certain level Achieved uninterrupted system operation and stability even in power shortage situations. Frequency determination method Downscale frequency across all GPUs or CPUs The minimum threshold frequency is calculated based on the computation and communication times for each GPU, and the maximum value is set on a rack-by-rack basis. Whether to consider synchronization Does not consider synchronization or communication latency between GPUs Reflects NCCL synchronization maintenance conditions by considering computation time and communication time between GPUs. Voltage control method Reduces voltage by simply linking it with frequency Voltage is controlled in conjunction by considering the frequency lower limit (synchronization threshold). Thermal management method Heat reduction is a side effect, and there is no separate heat buffer concept. Directly controls the thermal state based on the melting progress and depletion rate of a phase change material-based thermal buffer. Power shortage response methods There is a possibility of performance degradation or system shutdown in case of insufficient power. Combines column buffers and computational control to respond in stages and maintain uninterrupted operation Workload control Applying blanket performance degradation without distinguishing workload types Differentiate between inference nodes and learning nodes, and selectively suspend or evacuate learning nodes. Whether to preserve state State may be lost if operation is interrupted Preserve the state by moving the learning node's computation context to external memory. System stability Bottlenecks and synchronization issues may occur if the load is imbalanced. Ensures system stability by controlling based on conditions for maintaining synchronization between GPUs. applied area Single device or single node-centric control Integrated control of microgrids at the data center level Differences It corresponds to a general control technology for power saving. Differentiation secured through a complex control structure integrating heat, power, and computation

[0092] Referring to (a) of FIG. 3b, the maintenance unit (340) inputs the power generation amount of the microgrid (400) into a pre-established prediction model to predict the power gap time, calculates the thermal buffer time by dividing the thermal buffer capacity of the PCM by the depletion rate, and if the power gap time exceeds the thermal buffer time, the inference node that performs inference of the AI ​​model maintains normal operation, and the training node that performs training of the AI ​​model can move the computation context of the training node to external storage.

[0093] <Power Outage Time Prediction>

[0094] First, predict how long the power shortage will last.

[0095]

[0096] τgap(t) is the predicted power gap time at time t, (hat)trecovery is the expected time when power supply returns to normal levels, and t is the current time. Here, the term "power gap" does not mean that the power supply is zero, but rather refers to a state where power supply is insufficient. For example, this refers to a case where demand is 10 MW and supply is 7 MW. In this case, power is being supplied, but not sufficiently, resulting in a shortage of 3 MW; this is referred to as a power gap. Accordingly, a power gap does not mean a state of having no power at all, but rather a state where the required amount cannot be supplied.

[0097] These power gap times are predicted within the power system using, for example, generator ramp-up models, renewable energy output prediction models, and grid power recovery time models. For example, if the first time t+Δt such that Psup(t+Δt)≥Preq is found under the following conditions, this time becomes the expected time when power supply is restored to normal levels.

[0098]

[0099] At this time, the column buffer time is converted into the time that can withstand the column buffer capacity as shown in (b) of Fig. 3b, and is as follows in Equation 17.

[0100]

[0101] τsurv(t) is the estimated survival time at time t, i.e., the thermal buffer time, Qrembuf(t) is the thermal buffer capacity, and Dbuf(t) is the depletion rate of the thermal buffer capacity. ε is a very small constant to avoid division by zero. That is, by dividing the currently remaining thermal buffer capacity by the current depletion rate, one can calculate how many seconds or minutes can be sustained under the current conditions, and this value is called the thermal buffer time as described above. In one embodiment of the present invention, a decision is made by changing the physical quantity called thermal into the perspective of time.

[0102] Accordingly, as shown in (c) of FIG. 3b, if the power idle time exceeds the thermal buffer time, the training node is stopped or evacuated to CXL, while the inference node is maintained. That is, since the inference node, which performs inference on the AI ​​model, is directly related to service provision, it is maintained in normal operation to prevent response delays to user requests or service interruptions. On the other hand, since the training node, which performs training on the AI ​​model, causes relatively high power consumption and heat generation, the computation context of the training node is transferred to external storage in the event of power shortage or thermal limit, thereby temporarily removing the training load and rapidly reducing the power and thermal burden, while allowing training to resume in the same state thereafter. In addition, the GPU frequency may be further lowered, and some AI servers may be switched to sleep mode. Conversely, if the power idle time is less than the thermal buffer time, the thermal buffer capacity is sufficient to sustain it, so the workload can be maintained and only simple DVFS can be performed.

[0103] External storage can be an external shared memory pool based on the CXL (Compute Express Link) 3.0 fabric. In this case, the CXL 3.0-based external shared memory pool can be described as high-speed external RAM shared by multiple servers. In this case, high-speed evasion can be performed to the external shared memory pool. Here, high-speed evasion refers to evacuating the computation context by minimizing CPU intervention through CXL 3.0 memory semantics and DMA (Direct Memory Access)-based Page Migration technology.

[0104] With reference to (d) of FIG. 3b, the Measurement and Verification unit (350) can prove that the AI ​​server is an asset that has not undergone degradation by recording the temperature profile of the AI ​​server in a Distributed Ledger when the frequency and voltage of the GPU are gradually lowered in the Power Limiting unit (330), and then storing the hash value of the temperature profile in the metadata of the AI ​​server's certificate when the AI ​​server is disposed of via IT Asset Disposition Service.

[0105] Verification refers to a procedure or method of measuring and verifying the results regarding a system's performance, energy savings, and status. In other words, it means that the thermal stress (temperature profile) applied to the AI ​​server is directly recorded in a distributed ledger, and when the AI ​​server is recycled or remanufactured for sale, this serves as certification that the AI ​​server has never deteriorated. According to one embodiment of the present invention, heat is controlled through a PCM, and since it can be proven that there is no impact on the GPU or, more broadly, on the AI ​​server, verification becomes possible. Consequently, the AI ​​server of the present invention can be valued at a high disposal price as a defect-free asset during IT asset disposal.

[0106] FIG. 4a is a drawing attached as an example to explain a microgrid and a data center, FIG. 4b is a drawing attached as an example to explain a rack, FIG. 4c to 4e are drawings attached as examples to explain an AI server inserted into a rack and a GPU within the AI ​​server, and FIG. 4f is a drawing attached as an example to explain a cooling system of an AI server. The sources of FIG. 4a to 4f are as follows, and it is made clear that the drawings attached to FIG. 4 are attached only as examples for explanation and have absolutely no relation to the present invention.

[0107] floor plan source Fig. 4a https: / / www.gevernova.com / content / dam / Energy_Consulting / global / en_US / images / solutions / data-centers / updated-infographics-march-2025 / Data-centers-infographic-btm-hybrid-microgrid.png Fig. 4b EPRADO Standing Rack Cabinet (Model Name: EPRADO-R19-42U / 800X800) Fig. 4c https: / / docs.oracle.com / cd / E55040_01 / html / E54129 / figures / 127935.jpg Fig. 4d https: / / pim-media.intel.com / pub-api / v1 / imageservice / customize?url=https: / / marketplace.intel.com / file-asset / a5Y3b0000008ICvEAM_a5b3b0000004fCxAAI?.jpg&height=400&width=500 Fig. 4e https: / / www.asrockrack.com / general / products.asp Fig. 4f https: / / www.chromausa.com / applications / ai-server-test /

[0108] As for the uninterrupted operation solution provision method of FIGS. 2 to 4, the details not described are identical to or can be easily inferred from the details described above regarding the uninterrupted operation solution provision method through FIG. 1, so the following description will be omitted.

[0109] FIG. 5 is a diagram illustrating the process of transmitting and receiving data between each component included in the uninterrupted operation system of a microgrid-linked data center of FIG. 1 according to an embodiment of the present invention. Hereinafter, an example of the process of transmitting and receiving data between each component will be described through FIG. 5, but the present invention is not to be interpreted as being limited to such an embodiment, and it is obvious to those skilled in the art that the process of transmitting and receiving data illustrated in FIG. 5 may be modified according to various embodiments described above.

[0110] Referring to FIG. 5, the integrated control server links the microgrid and the data center (S5100), and when the power generation of the microgrid decreases, it measures the melting fraction of the phase change material (PCM) placed in the liquid cooling loop within the data center and calculates the thermal buffer capacity (S5200).

[0111] And, the integrated control server gradually lowers the frequency and voltage of the GPUs based on the depletion rate of the thermal buffer capacity to a minimum frequency threshold at which network synchronization between GPUs of at least one AI server in the data center is maintained (S5300).

[0112] The order of the steps described above (S5100~S5300) is merely an example and is not limited thereto. That is, the order of the steps described above (S5100~S5300) may vary, and some of these steps may be executed simultaneously or deleted.

[0113] As for the details regarding the method of providing an uninterrupted operation solution of Fig. 5 that are not described, they are identical to or can be easily inferred from the details described above regarding the method of providing an uninterrupted operation solution through Figs. 1 to 4, so further explanation will be omitted.

[0114] A method for providing an uninterrupted operation solution according to one embodiment described through FIG. 5 may also be implemented in the form of a recording medium containing instructions executable by a computer, such as an application or program module executed by a computer. A computer-readable medium may be any available medium accessible by a computer and includes both volatile and non-volatile media, and both removable and non-removable media. Additionally, a computer-readable medium may include all computer storage media. Computer storage media include both volatile and non-volatile, removable and non-removable media implemented by any method or technique for storing information such as computer-readable instructions, data structures, program modules, or other data.

[0115] The method for providing an uninterrupted operation solution according to one embodiment of the present invention described above may be executed by an application basically installed on a terminal (which may include a program included in a platform or operating system, etc., basically installed on the terminal), or by an application (i.e., a program) directly installed by a user on a master terminal through an application providing server, such as an application store server, an application, or a web server related to the service. In this sense, the method for providing an uninterrupted operation solution according to one embodiment of the present invention described above may be implemented as an application (i.e., a program) that is basically installed on a terminal or directly installed by a user, and may be recorded on a computer-readable recording medium such as a terminal.

[0116] The foregoing description of the present invention is for illustrative purposes only, and those skilled in the art will understand that other specific forms can be easily modified without altering the technical spirit or essential features of the present invention. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. For example, each component described as a single unit may be implemented in a distributed manner, and components described as distributed may likewise be implemented in a combined form.

[0117] The scope of the present invention is defined by the claims set forth below rather than by the detailed description above, and all modifications or variations derived from the meaning and scope of the claims and equivalent concepts thereof should be interpreted as being included within the scope of the present invention.

Claims

Claim 1 An uninterrupted operation system for a microgrid-linked data center comprising: an integration unit linking a microgrid and a data center; a thermal buffer calculation unit that calculates a thermal buffer capacity by measuring the melting fraction of a Phase Change Material (PCM) placed in a liquid cooling loop within the data center when the power generation of the microgrid decreases; and an integrated control server that stepwise lowers the frequency and voltage of the GPUs up to a minimum frequency threshold where network synchronization between GPUs of at least one AI server of the data center is maintained based on the depletion rate of the thermal buffer capacity; wherein the power limiting unit increases the reduction value of the frequency and voltage of the GPUs as the depletion rate increases, and decreases the reduction value of the frequency and voltage of the GPUs as the depletion rate decreases. Claim 2 An uninterrupted operation system for a microgrid-linked data center according to claim 1, wherein the melting progress rate is calculated based on a temperature-based approximation formula in which the melting progress rate increases linearly as the temperature of the PCM rises within the phase change section after the temperature of the PCM is measured, the heat buffer capacity is calculated by multiplying the mass of the PCM by the latent heat stored in the PCM and then multiplying the remainder obtained by subtracting the melting progress rate from 1, and the latent heat is the inherent latent heat per unit mass of the PCM. Claim 3 delete Claim 4 The uninterrupted operation system of a microgrid-linked data center according to claim 1, further comprising: a maintenance unit in which the integrated control server inputs the power generation amount of the microgrid into a pre-established prediction model to predict the power gap time, calculates the heat buffer time by dividing the heat buffer capacity of the PCM by the depletion rate, and if the power gap time exceeds the heat buffer time, maintains the normal operation of the inference node performing inference of the AI ​​model and moves the computation context of the training node performing training of the AI ​​model to external storage. Claim 5 A non-disruptive operation system for a microgrid-linked data center according to claim 4, characterized in that the external storage is an external shared memory pool based on CXL (Compute Express Link) 3.0 fabric. Claim 6 The uninterrupted operation system of a microgrid-linked data center according to claim 1, further comprising: a Measurement and Verification unit in which the integrated control server records the temperature profile of the AI ​​server in a Distributed Ledger when the frequency and voltage of the GPU are gradually lowered in the power limiting unit, and then, when the AI ​​server is disposed of via IT Asset Disposition Service, binds and stores the hash value of the temperature profile to the metadata of the AI ​​server's certificate to prove that the AI ​​server is an asset that has not undergone degradation. Claim 7 A method for operating a data center running on an integrated control server, comprising: a step of linking a microgrid and a data center; a step of calculating a thermal buffer capacity by measuring the melting fraction of a Phase Change Material (PCM) placed in a liquid cooling loop within the data center when the power generation of the microgrid decreases; and a step of gradually lowering the frequency and voltage of the GPUs based on the depletion rate of the thermal buffer capacity to a minimum frequency threshold at which network synchronization between GPUs of at least one AI server of the data center is maintained; wherein the lowering step is characterized by increasing the reduction value of the frequency and voltage of the GPUs as the depletion rate increases, and lowering the reduction value of the frequency and voltage of the GPUs as the depletion rate decreases. Claim 8 A computer-readable recording medium storing a program for executing the method of claim 7 on a computer.

Images