Optimization systems, devices, methods, and programs for optimizing the operating costs of facilities.

Abstract

This invention provides an optimization system, optimization apparatus, optimization method, and program that minimizes data center operating costs while preventing thermal damage to electronic components by treating the temperature management of individual servers or computers within a data center as an optimization problem using quantum annealing. [Solution] A method using an optimization system having an annealing device, comprising: S1 generating an optimization problem relating to the operating costs of a facility; S2 defining binary variables for executing the optimization problem on the annealing device; S3 obtaining one or more physical parameter values ​​relating to the physical characteristics of the facility; S4 converting the optimization problem into a quadratic unconstrained binary optimization problem using one or more physical parameter values; S5 transmitting and receiving the quadratic unconstrained binary optimization problem to and from the annealing device; S6 obtaining a solution from the annealing device; and S7 controlling the operation of the facility based on the obtained solution.

Description

[Technical Field]

[0001] The present disclosure relates to an optimization system, an optimization apparatus, an optimization method, and a program for minimizing the operation cost of a facility, particularly a data center. [Background Art]

[0002] In recent years, with the prosperity of generative services represented by ChatGPT (registered trademark), demand for data centers that collect servers equipped with high-performance graphics cards has been increasing. On the other hand, these servers generate a large amount of heat during data processing. Furthermore, if a graphics card is heated above a predetermined temperature, its failure rate increases. Therefore, data centers require more advanced temperature management, and the power consumption of air conditioners has also reached a level that cannot be ignored. There are multiple individually controllable air conditioners in a data center, and the cooling effect of turning them on or off on individual servers depends on factors such as the layout of servers in the data center, air temperature, and the material of racks. Therefore, it is generally extremely difficult to uniquely calculate which air conditioner should be operated at what intensity based on the current temperature of each individual server.

[0003] As a conventional technology related to temperature management in data centers, for example, Japanese Patent Application Laid-Open No. 2023-104385 is available (hereinafter referred to as Patent Document 1). Patent Document 1 discloses a technique for balancing power supply and reducing costs by lowering the set temperature of air conditioners in a data center when power supply from green power or the like becomes excessive. [Prior Art Documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Application Laid-Open No. 2023-104385

[0005] However, the technology disclosed in Patent Document 1 lowers the set temperature of the entire data center and cannot manage the temperature of individual servers within the data center.

[0006] On the other hand, quantum annealing is considered a promising method for solving the complex optimization problems described above. Quantum annealing is a technique that solves optimization problems by utilizing the fact that the physical Ising model, given conditions between spins, results in a combination of spin states that minimizes the overall energy of the system. However, no conventional techniques are known to have treated the temperature management of individual servers within a data center as an optimization problem using quantum annealing. [Overview of the project] [Problems that the invention aims to solve]

[0007] This disclosure is made in view of the above circumstances, and its purpose is to provide an optimization system, optimization device, optimization method, and program for minimizing data center operating costs while preventing thermal damage to very expensive electronic components such as graphics cards, by treating the temperature management of individual servers or computers within a data center as an optimization problem using quantum annealing. [Means for solving the problem]

[0008] In one embodiment, the optimization system of the present disclosure is an optimization system having an annealing device, comprising: a problem generation unit that generates an optimization problem relating to the operating costs of a facility; a physical parameter acquisition unit that acquires one or more physical parameter values ​​relating to the physical characteristics of a facility; a secondary unconstrained binary optimization conversion unit that uses one or more physical parameter values ​​to convert the optimization problem into a secondary unconstrained binary optimization problem; and a facility operation control unit that controls the operation of the facility based on the solution to the secondary unconstrained binary optimization problem obtained from the annealing device, wherein the secondary unconstrained binary optimization conversion unit converts the optimization problem into a secondary unconstrained binary optimization problem that includes at least the total amount of the facility's electricity costs and equipment consumption costs.

[0009] In another embodiment, the optimization device of the present disclosure is an optimization device that communicates with an external annealing device and comprises a problem generation unit that generates an optimization problem relating to the operating costs of a facility; a physical parameter acquisition unit that acquires one or more physical parameter values ​​relating to the physical characteristics of a facility; a secondary unconstrained binary optimization conversion unit that uses one or more physical parameter values ​​to convert the optimization problem into a secondary unconstrained binary optimization problem; a processor that functions as a facility operation control unit that controls the operation of the facility based on the solution of the secondary unconstrained binary optimization problem obtained from the annealing device; and a computer having memory, wherein the secondary unconstrained binary optimization conversion unit converts the optimization problem into a secondary unconstrained binary optimization problem that includes at least the total amount of the facility's electricity costs and equipment consumption costs.

[0010] In another embodiment, the optimization method of the present disclosure is a computer-operated optimization method comprising: a problem generation step of generating an optimization problem relating to the operating costs of a facility; a physical parameter acquisition step of acquiring one or more physical parameter values ​​relating to the physical characteristics of a facility; a secondary unconstrained binary optimization conversion step of using one or more physical parameter values ​​to convert the optimization problem into a secondary unconstrained binary optimization problem; and a facility operation control step of controlling the operation of the facility based on the solution to the secondary unconstrained binary optimization problem obtained from an annealing device, wherein the secondary unconstrained binary optimization conversion step includes a step of converting the optimization problem into a secondary unconstrained binary optimization problem that includes at least the total amount of the facility's electricity costs and equipment consumption costs.

[0011] In yet another embodiment, the program of the present disclosure is a program that causes a computer to execute a problem generation procedure for generating an optimization problem relating to the operating costs of a facility; a physical parameter acquisition procedure for obtaining one or more physical parameter values ​​relating to the physical characteristics of a facility; a quadratic unconstrained binary optimization conversion procedure for converting the optimization problem into a quadratic unconstrained binary optimization problem using one or more physical parameter values; and a facility operation control procedure for controlling the operation of a facility based on the solution to the quadratic unconstrained binary optimization problem obtained from an annealing device, wherein the quadratic unconstrained binary optimization conversion procedure includes a procedure for converting the optimization problem into a quadratic unconstrained binary optimization problem that includes at least the total amount of the facility's electricity costs and equipment consumption costs. [Effects of the Invention]

[0012] According to the optimization system, optimization apparatus, optimization method, and program described herein, the temperature management of individual servers or computers within a data center can be treated as a quantum annealing optimization problem, thereby minimizing the operating costs of the data center while preventing thermal damage to very expensive electronic components such as graphics cards. [Brief explanation of the drawing]

[0013] [Figure 1]Figure 1 is an overall conceptual diagram showing an example of applying an optimization system and optimization apparatus according to the embodiments of this disclosure to a data center. [Figure 2] Figure 2 is a schematic diagram showing the inside of a server room in a data center. [Figure 3] Figure 3 is a block diagram showing an optimization system, optimization apparatus, and peripheral equipment according to an embodiment of the present disclosure. [Figure 4] Figure 4 is a flowchart showing an optimization method according to an embodiment of the present disclosure. [Figure 5] Figure 5 is a schematic diagram showing the correlation between task execution, air conditioner operation, and computer temperature over time. [Figure 6] Figure 6 is a graph summarizing the results from Figure 5 in relation to the computer's temperature and time. [Figure 7] Figure 7 is a table showing the correspondence between m-qubit variables and binary variables that represent the strength of a load balancer according to other embodiments of this disclosure. [Figure 8] Figure 8 is a graph illustrating a method for determining approximate trouble costs according to other embodiments of this disclosure. [Modes for carrying out the invention]

[0014] Embodiments of the optimization system, optimization apparatus, optimization method, and program described herein will be explained below with reference to the drawings. Throughout the drawings, the same components are denoted by the same reference numerals. The embodiments shown in the drawings are illustrative and do not limit the disclosure in any way. Furthermore, the embodiments and their modifications described below can be combined in any way, and such combinations are included within the scope of the disclosure.

[0015] In the following explanation, data centers will be used as an example of facilities to be optimized, but the facilities are not limited to data centers. For example, other examples of facilities include agricultural greenhouses, cultivation or aquaculture facilities, marine product storage facilities, bioproduct storage facilities (such as human or animal sperm, eggs, embryos, and blood), pharmaceutical storage facilities (such as vaccines), zoos and botanical gardens, aquariums, office buildings, and logistics centers.

[0016] Figure 1 is an overall conceptual diagram showing an example of applying an optimization system 10 according to an embodiment of the present disclosure to a data center 1. The data center 1 consists of a plurality of server rooms 2. Each server room 2 includes a plurality of racks 3, a plurality of air conditioners 4, a power supply unit 5, and a task distribution unit 6. The power supply unit 5 may also be called a temperature control unit 5.

[0017] Each rack 3 has a chassis and multiple shelves for placing computers, arranged at equal intervals inside the chassis. Each air conditioner 4 may have a cooling capacity of, for example, 40 to 60 kW, and approximately 3 to 5 units are installed in one server room 2. In this example, for the sake of explanation, one air conditioner 4 is provided for each rack 3, but this is not limited to this, and one air conditioner 4 may be provided for two racks 3.

[0018] The power supply unit 5 receives power from an external power supply system and distributes it to the air conditioner 4. The power supply unit 5 also functions as a temperature control unit, adjusting the cooling intensity and airflow volume of the air conditioner 4, and controlling the ON / OFF status of the air conditioner 4 to maintain a set temperature. The task distribution unit 6 receives task commands from an external network 7 and assigns tasks to be executed to designated computers (servers) mounted in the rack 3.

[0019] The optimization system 10 is connected to the power supply unit (temperature control unit) 5 and the task distribution unit 6 via the operation instruction unit 8 and the network 7. As will be described later, the operation instruction unit 8 transmits instruction signals to the power supply unit 5 and the task distribution unit 6 via the network 7 according to the solution obtained by the annealing device. The power supply unit 5 operates the air conditioner 4 according to the instruction signals, and the task distribution unit 6 assigns tasks to any computer (server) in the rack 3 according to the instruction signals. In other words, the operation instruction unit 8, the power supply unit (temperature control unit) 5, and the task distribution unit 6 are configured to optimally operate the server room 2, and by extension the data center 1, based on the solution obtained from the annealing device.

[0020] Figure 2 is a schematic diagram of the inside of server room 2 in data center 1. In the following description, computers may also be servers. Computers 1 (31), 2 (32), 3 (33), and 4 (34) are mounted in rack 1 (3) on the left, and computers 5 (35) ... Computer M (36) are mounted in rack 2 (3) on the right. An air conditioner 1 (4) is installed adjacent to rack 1 (3), and an air conditioner 2 (4) is installed adjacent to rack 2 (3). Each rack 3 is installed on a floor panel 30 with a double-floor structure in which many small holes are formed.

[0021] Each air conditioner 4 generates a flow of cool air 40 under the double floor in a direction parallel to the front of each rack 3 and parallel to the floor surface. The cool air flow 40 rises as cool air 41 onto the double floor through small holes in the floor panel 30. The cool air 41 is drawn to the front of each rack 3, cooling the computers 31-36 inside the rack 3, and is discharged from the back of the rack 3 as hot exhaust air 42. The exhaust air 42 returns to the air conditioner 4 as a flow of warm air 43 in the opposite direction to the cool air flow 40.

[0022] Computers 31-36 are each equipped with a temperature sensor (not shown). The temperature of computers 31-36 is measured by the temperature sensor, for example, every minute, every 5 minutes, or every 10 minutes. The time interval for temperature measurement can be arbitrary. The measured temperatures of computers 31-36 will be different, for example, computer 1 (31) will show 48.6°C, computer 2 (32) will show 53.4°C, computer 3 (33) will show 55.0°C, computer 4 (34) will show 50.6°C, computer 5 (35) will show 49.3°C, and computer M (36) will show 65.6°C. This temperature information is sent to the power supply unit (temperature control unit) 5.

[0023] Figure 3 is a block diagram illustrating an optimization system, optimization device, and peripheral devices according to an embodiment of the present disclosure. The optimization device system comprises an optimization device 11, which is a classical computer, and an annealing device 19. The optimization device 11 comprises a control device 12, a memory 14, a storage device 16, and an interface 18. The optimization device 11 may be connected via a network 7 to peripheral devices such as a server room 2, a machine learning device 20, an input device 22, and a display device 24.

[0024] The control device 12 may be a computer including one or more processors (e.g., a CPU). The memory 14 may be volatile memory such as DRAM (Dynamic Random Access Memory) or SRAM (Static Random Access Memory). The storage device 16 may be flash memory such as ROM (Read Only Memory) or SSD (Solid State Drive), or non-volatile memory such as HDD (Hard Disk Drive).

[0025] Interface 18 includes an input unit, an output unit, and a communication unit. Interface 18 connects to an input device 22, a display device 24, and a machine learning device 20 via a network 7. Network 7 may be the Internet, Wi-Fi, or other network. Input device 22 may be a keyboard, mouse, touch panel, stylus, etc. Display device 24 may be a liquid crystal display, an organic EL display, etc. Machine learning device 20 may be a machine learning machine using AI (Artificial Intelligence).

[0026] The annealing device 19 may be a quantum annealing computer, a digital annealer (registered trademark), a CMOS annealing machine, a simulated branching machine, etc. Furthermore, the annealing device 19 may be implemented as a cloud service such as Leap (D-Wave Systems), AWS (Amazon Web Services), Braket, Microsoft Azure Quantum, or IBM Quantum Experience. That is, the annealing device 19 may be installed in a location spatially separated from the control device 12; for example, the annealing device 19 may be installed overseas. The annealing device 19 is a hardware device capable of executing a different quantum annealing algorithm than that of the optimization device 11. In addition to the annealing device 19, the optimization system 10 may include an optimization solver (not shown). The optimization solver may be implemented as software that implements a prim-dual interior-point method or Newton's method capable of solving optimization problems.

[0027] The control device 12 functions as a problem generation unit for generating optimization problems, a variable definition unit, a physical parameter acquisition unit, a quadratic unconstrained binary optimization conversion unit, a quadratic unconstrained binary optimization problem transmission / reception unit, and a facility operation control unit.

[0028] First, let's explain the function of the problem generation unit. The problem generation unit generates an optimization problem concerning the operating costs of a facility such as data center 1. Here, the operating cost is the sum of costs based on the facility's power consumption, such as electricity costs, costs based on the facility's equipment consumption, such as equipment depreciation costs, and costs based on the productivity of the facility's equipment (in the case of data center 1, computers 31-36), such as profits. total Generally, it is defined as follows (Equation 1). C total = Σ i C i (E,W,-R) (Formula 1) Here, C i is the cost of the i-th server room 2, E is the electricity cost, W is the equipment depreciation cost, and R is the profit. The problem generation unit is C total An optimization problem is defined as a problem to be solved by minimizing a certain value. The problem generation unit stores the defined optimization problem in the optimization problem storage unit in memory 14.

[0029] Next, the function of the variable definition section will be explained. The variable definition section defines binary variables for executing the optimization problem in the annealing device 19. In order to solve the optimization problem using the annealing device 19, it is necessary to represent the optimization problem using an Ising model, which is a model of a magnetic material. The Ising model is a simplified model for dealing with phase transitions in ferromagnets, and is represented by spin states with up and down orientations that represent the properties of the magnetic material, interaction coefficients that represent the force of the interaction between two spins, and external field coefficients that represent the force of an externally applied magnetic field. In the Ising model, for example, the energy is determined by the interaction between spins that take binary values ​​of +1 and -1, and the spin states are updated so that the energy of the Ising model is minimized. The combination of parameters that minimizes the evaluation index of the optimization problem is obtained as the combination of spin states that minimizes the energy by mapping the optimization problem to correspond to the energy of the Ising model and converging the Ising model.

[0030] The QUBO (Quadratic Unconstrained Binary Optimization) format is a representation of the overall spin evaluation function using a binary variable (0 or 1) instead of +1 and -1 spins. The variable definition section defines a binary variable (0 or 1) in order to formalize the optimization problem (Equation 1) defined in the problem generation section into QUBO format.

[0031] The specific functions of the variable definition unit in this embodiment will now be explained. For example, suppose that N air conditioners 4 are installed in a data center 1. Each air conditioner 4 can be individually controlled by a power supply unit (temperature control unit) 5, and its power consumption is determined according to the strength of the air conditioning. Also, suppose that M computers 31 to 36 are housed in multiple racks 3 within the data center 1. Here, a computer refers to any processing system in which temperature and power consumption can be measured. In other words, multiple servers may be considered as one computer, or each rack 3 of servers may be considered as one computer.

[0032] Tasks are constantly pouring into Data Center 1 for processing. To distinguish between expected tasks from one point in time to another, a number is assigned. Each task has a predetermined reward and power consumption. For example, the reward for the i-th task is (number 1), and the power consumption is (number 2).

number

number

[0033] In this example, the variable definition section defines two variables that can be manipulated during optimization. The first variable is related to the task allocation by the task distribution section 6 and is defined as a variable (equation 3) that indicates whether or not the i-th task is executed on the j-th computer.

number

number

number

number

[0034] The second variable relates to the power supplied to the air conditioner 4 by the power supply unit (temperature control unit) 5, i.e., the power consumption of the air conditioner 4, and is defined as a variable (equation 7) that indicates whether the i-th air conditioner 4 is operated at intensity j (for example, j takes values ​​of 0 (intensity 0: off), 1 (intensity 1: weak), 2 (intensity 2: medium), and 3 (intensity 3: strong)).

number

[0035] Here, similar to (Equation 3), (Equation 7) is also a binary variable. When (Equation 8) is true, the i-th air conditioner 4 is operated at intensity j.

number

[0036] The variable definition unit stores the two variables (number 3) and (number 7) defined as described above in the variable definition storage unit within memory 14.

[0037] Next, the functions of the physical parameter acquisition unit will be explained. The physical parameter acquisition unit acquires one or more physical parameter values ​​related to the physical characteristics of the facility. Here, the physical characteristics of the facility refer to the physical characteristics of the equipment within the data center 1 that change according to the execution of tasks by the computer and / or the operating intensity of the air conditioner 4. Examples of physical parameter values ​​related to physical characteristics include the current temperature of the computer, the temperature rise per unit power consumption of the computer, the temperature drop of the computer when the intensity of the air conditioner 4 is set to a certain value, and the thermal fatigue degree of the computer. Of these, the temperature rise per unit power consumption of the computer and the temperature drop of the computer when the intensity of the air conditioner 4 is set to a certain value will differ depending on factors such as the actual computer placement, the material of the rack 3, and the outside air temperature, and will therefore be measured in practice.

[0038] Of the physical parameters described above, the current temperature of the computer is detected by a temperature sensor installed in each computer and sent to the physical parameter acquisition unit via the network 7. The temperature rise per unit power consumption of the computer and the temperature drop of the computer when the intensity of the air conditioner 4 is set to a certain value are actually measured in the server room 2 and sent to the physical parameter acquisition unit via the network 7. This will be explained later. In addition, the thermal fatigue level of the computer is calculated from the number of thermal cycles of the computer, which is stored and updated in the storage device 16, and sent to the physical parameter acquisition unit. The physical parameter acquisition unit stores the acquired physical parameter values ​​in the physical parameter storage section of memory 14.

[0039] Next, we will explain the function of the quadratic unconstrained binary optimization transformation unit. The quadratic unconstrained binary optimization transformation unit uses one or more of the above-mentioned physical parameter values ​​to transform the optimization problem (Equation 1) into a quadratic unconstrained binary optimization problem (QUBO format). First, we formulate the electricity cost E in (Equation 1). The power consumption when the intensity of the air conditioner 4 is i is expressed by (Equation 9).

number

Mathematical Expression

Mathematical Expression

Mathematical Expression

Mathematical Expression

[0040] Next, the profit R in (Equation 1) is formulated. The profit R is formulated by (Equation 14).

Mathematical Expression

[0041] Next, the facility consumption amount W in (Equation 1) is formulated. This defines a quantity related to temperature for calculating the failure rate of computers. Among the physical parameter values described above, let T be the temperature rise amount per unit power consumption of a computer unit . Let L be the heat resistant temperature of the server. L is a constant. The current temperature of the i-th server (computer) is represented by (Equation 15). Let w be the facility consumption degree per unit temperature i expressed by w iThis value represents the amount of wear and tear that occurs for every 1 degree Celsius increase in the computer's temperature. The amount by which the computer i's temperature decreases when the air conditioner j is set to intensity k is represented by (Equation 16). The equipment wear and tear cost W is formulated by (Equation 17).

number

number

number

[0042] The quadratic unconstrained binary optimization transformation unit formulates (Equation 1) as an evaluation function in QUBO form as shown in (Equation 18).

number

[0043] Next, the function of the quadratic unconstrained binary optimization problem transmission / reception unit will be explained. The quadratic unconstrained binary optimization problem transmission / reception unit outputs the quadratic unconstrained binary optimization problem to the annealing device 19. The quadratic unconstrained binary optimization problem is formulated by converting the optimization problem into QUBO format by the quadratic unconstrained binary optimization conversion unit, and in this example it is represented as (Equation 18). The quadratic unconstrained binary optimization problem transmission / reception unit outputs the evaluation function (Equation 18) stored in the evaluation function storage unit of memory 14 to the annealing device 19 via a network such as the Internet. The annealing device 19 calculates the set of optimal solutions that minimize the evaluation function (Equation 18) in QUBO format, that is, the optimal parameter set for task distribution to the computer (Equation 3) and operation control of the air conditioner 4 (Equation 7). In the example where the annealing device 19 is a digital annealer, the digital annealer randomly changes all variables. As a result the evaluation formula goes up and down, but if only changes are made in the direction of going down, the optimal solution cannot be obtained, so changes that initially worsen the evaluation are probabilistically allowed. The optimal value is found by gradually reducing the probability of deterioration. By continuously improving the evaluation in this way, the optimal solution can be obtained in a short time. The second-order unconstrained binary optimization problem transmission and reception unit stores the solution obtained from the annealing device 19 via the network in the solution storage unit of the memory 14.

[0044] Next, the functions of the facility operation control unit will be described. The facility operation control unit controls the operation of the facility based on the solution obtained from the annealing device 19 via the network. Facility operation refers to the distribution of tasks to the computers in the data center 1 and the operation control of the air conditioners 4. The facility operation control unit sends control signals to the operation instruction unit 8, and the operation instruction unit 8 transmits instruction signals via the network 7 to the server room 2, i.e., the power supply unit (temperature control unit) 5 and the task distribution unit 6 shown in Figure 1. Based on the instruction signals, the power supply unit (temperature control unit) 5 controls the ON / OFF and intensity of the air conditioners 4, and the task distribution unit 6 controls the allocation of tasks to the computers.

[0045] Figure 4 is a flowchart showing an optimization method according to an embodiment. The optimization method according to this embodiment includes a problem generation step (step S1) for generating an optimization problem relating to the operating costs of a facility, a variable definition step (step S2) for defining binary variables in order to execute the optimization problem in the annealing device 19, a physical parameter acquisition step (step S3) for acquiring one or more physical parameter values ​​relating to the physical characteristics of the facility, a quadratic unconstrained binary optimization conversion step (step S4) for converting the optimization problem into a quadratic unconstrained binary optimization problem using one or more physical parameter values, a quadratic unconstrained binary optimization problem transmission / reception step (steps S5, S6) for transmitting and receiving the quadratic unconstrained binary optimization problem to and from the annealing device 19, and a facility operation control step (step S7) for controlling the operation of the facility based on the solution obtained from the annealing device 19.

[0046] In step S1, the problem generation unit generates the optimization problem (Equation 1) and stores it in the optimization problem storage unit of memory 14.

[0047] In step S2, the variable definition unit defines a first variable (number 3) and a second variable (number 7), and stores the defined first variable (numbers 3, 4, and 5) and second variable (numbers 7 and 8) in the variable definition storage unit of memory 14. The conditions (number 6) for each variable are also stored together.

[0048] In step S3, the physical parameter acquisition unit obtains the current temperature of the server (computer) (number 15) and the temperature rise T per unit power consumption of the computer. unit The amount of temperature drop of computer i when the intensity of air conditioner j is set to k (equation 16), and the degree of wear and tear of the computer per unit temperature w. i , and the thermal fatigue degree F of the i-th computer i The system obtains the following parameters: the reward for the i-th task (number 1), the power required for the i-th task (number 2), the power consumption of the air conditioner 4 when its intensity is i (number 9), and the electricity cost P per unit power consumption of the computer or air conditioner 4. These parameters are input from the input device 22 and stored in the storage device 16.

[0049] Among the physical parameters, the current temperature of the server (computer) (number 15) is detected by a temperature sensor built into the computer, transmitted to the physical parameter acquisition unit at predetermined time intervals, and overwritten in the physical parameter storage unit of memory 14.

[0050] The rate of wear and tear on a computer per unit temperature lol i This varies depending on the processing speed and purchase date of the computers in Server Room 2. The horizontal axis represents processing speed, and the vertical axis represents purchase date. i It is preferable to create a table of values ​​and store it in the storage device 16. The table may be created by inputting from the input device 22 or downloaded via the network 7. Note that the computer's wear rate per unit temperature w i may be a constant.

[0051] The thermal fatigue level F of the i-th computer i This can be determined from the number of thermal cycles of the i-th computer. The number of thermal cycles refers to the number of times the temperature detected by the temperature sensor built into the computer changes to a value above a certain threshold. Thermal fatigue degree F i is, F i It can be calculated as = α·n, where α is a coefficient and n is the number of thermal cycles. The thermal fatigue degree F of the computer obtained in this way is i Store it in the storage device 16.

[0052] Temperature rise T per unit power consumption of a computer unit Furthermore, the temperature drop of computer i (Equation 16) when the intensity of air conditioner j is set to k will vary depending on factors such as the actual computer placement, the material of rack 3, and the ambient temperature, as described above, and will therefore be measured. The measurement method will be explained in detail below.

[0053] Figure 5 is a schematic diagram showing the correlation between task execution, operation of air conditioner 4, and computer temperature in a time series. In this example, it is assumed that the type of task is the same. When a task is assigned to a computer, the computer executes the task. As the task is executed, the computer's temperature rises, and air conditioner 4 turns on. In this example, for the sake of explanation, air conditioner 4 is turned on at the same time as the task starts, but the timing of turning on air conditioner 4 is not limited to this.

[0054] Referring to the upper part of Figure 5, the air conditioner 4 is operating at intensity 1, or "low". At this time, the computer's temperature rises at a steep angle, reaches a maximum temperature, and then gradually decreases. Next, referring to the middle part of Figure 4, the air conditioner 4 is operating at intensity 2, or "medium". At this time, the computer's temperature rises at a slightly smaller angle than at intensity 1, reaches a maximum temperature slightly lower than at intensity 1, and then decreases at a slightly steeper angle than at intensity 1. Next, referring to the lower part of Figure 4, the air conditioner 4 is operating at intensity 3, or "high". At this time, the computer's temperature rises at a slightly smaller angle than at intensity 2, reaches a maximum temperature slightly lower than at intensity 2, and then decreases at a slightly steeper angle than at intensity 2. It can also be seen that the maximum temperature difference of the computer is intensity 1 < intensity 2 < intensity 3.

[0055] Figure 6 is a graph summarizing the results from Figure 5 in relation to the computer's temperature and time. The temperature rise T per unit power consumption of the computer. unit This can be determined from the slope of the graph shown in Figure 6. The computer's temperature rises at a predetermined rate between t0 and t1, regardless of the intensity of the air conditioner 4. This indicates that it takes time for the computer to cool down after the air conditioner 4 is turned on. In other words, the effect of the air conditioner 4 can be ignored between t0 and t1, so the temperature rise T per unit power consumption of the computer can be determined from the slope of the graph during this period and the computer's power consumption. unit It is possible to find this.

[0056] On the other hand, when the intensity of the air conditioner j is set to k, the temperature drop of the computer i (Equation 16) can be determined from the graph shown in Figure 6 as the temperature ΔT that drops within a predetermined time Δt from the time when the computer's temperature is at its maximum. The physical parameter acquisition unit then calculates the temperature rise T per unit power consumption of the computer obtained in this way. unit The amount of temperature reduction of the computer (equation 16) is stored in the physical parameter storage section of memory 14.

[0057] As shown in Figure 6, the temperature rise T per unit power consumption of the computer. unit The temperature drop of the computers (Equation 16) is obtained by actual measurements, but can also be predicted using the machine learning device 20. The configuration of computers mounted in rack 3 within server room 2, the material of rack 3, the ambient temperature, the amount of tasks, and the temperature rise T per unit power consumption of the computers are all factors considered. unit Furthermore, actual measurement sample data of the computer's temperature drop (Equation 16) is stored in a database, and the graph in Figure 6 can be predicted for any computer through AI-based machine learning. This allows the temperature rise T per unit power consumption to be predicted from the graph in Figure 6 using the machine learning device 20, without actually performing measurements. unit The computer's temperature drop (Equation 16) can also be obtained. The physical parameter acquisition unit obtains the temperature rise T per unit power consumption of the computer, which is determined using the machine learning device 20. unit The amount of temperature reduction of the computer (equation 16) may also be stored in the physical parameter storage section of the memory 14 via the network 7.

[0058] In step S4, the second-order unconstrained binary optimization transformation unit accesses the optimization problem storage unit, variable definition storage unit, and physical parameter storage unit of memory 14, as well as the storage device 16, to obtain the binary variable (Equation 3), physical parameter value (Equation 15, T unit Number 16, w i , F iThe optimization problem generated by the problem generation unit is converted into QUBO format by reading out the other parameter values ​​(Equation 1, Equation 2, Equation 9, P), formulating the electricity cost E (Equation 10), profit R (Equation 14), and equipment consumption cost W (Equation 17), and formulating the evaluation function as (Equation 18). The second-order unconstrained binary optimization conversion unit stores the evaluation function (Equation 18) in the evaluation function storage unit of memory 14.

[0059] In step S5, the second-order unconstrained binary optimization problem transmission / reception unit accesses the evaluation function storage unit in memory 14 to read the evaluation function (equation 18) and transmits it to the annealing device 19 via the network. The annealing device 19 calculates the solution that minimizes the evaluation function (equation 18).

[0060] In step S6, the second-order unconstrained binary optimization problem transmission / reception unit obtains the solution from the annealing device 19 via the network and stores it in the solution storage unit of the memory 14.

[0061] In step S7, the facility operation control unit accesses the memory 14's decryption unit and sends a control signal to the operation instruction unit 8. The operation instruction unit 8 transmits instruction signals via the network 7 to the server room 2, i.e., the power supply unit (temperature control unit) 5 and the task distribution unit 6 shown in Figure 1. Based on the instruction signals, the power supply unit (temperature control unit) 5 controls the ON / OFF status and intensity of the air conditioner 4, and the task distribution unit 6 controls the allocation of tasks to the computers.

[0062] Next, other embodiments of this disclosure will be described with reference to the drawings. The embodiments described above are applicable to facilities other than data centers (for example, agricultural greenhouses, cultivation or aquaculture facilities, marine product storage facilities, bioproduct storage facilities (human or animal sperm, eggs, embryos, blood, etc.), pharmaceutical (vaccine, etc.) storage facilities, zoos, aquariums, office buildings, logistics centers, etc.), but the other embodiments described below are more preferably applied when the facility is a data center.

[0063] The description of configurations similar to those in the embodiments described above will be omitted, and only the different configurations will be described. In the other embodiments, the definition of the variables that cause the computer to execute tasks, the definition of power consumption, the definition of electricity costs, the formula for trouble costs, and the formula for total costs differ from those in the embodiments described above. These will be described in detail below.

[0064] In this embodiment, a load balancer is used when assigning tasks to computers. Essentially, all tasks to be processed in a data center are handled by computers within servers. However, if tasks are distributed equally among all computers in the server, computers with higher temperatures will experience a greater processing load compared to those with lower temperatures. Therefore, a load balancer is used to distribute fewer tasks to computers with higher temperatures and more tasks to computers with lower temperatures.

[0065] In this embodiment, the strength of the load balancer is represented by the first variable q. Variable q is expressed using m qubits. m The level can be set in stages. The variable q can be expressed using a binary variable as (equation 19).

number

[0066] Figure 7 shows the strength of a load balancer according to another embodiment, with an m-qubit variable q and a binary variable (q1, q2, q3, ... q m This table shows the correspondence with q. When q=0, the load balancer strength is 0, and processing is distributed equally to all server computers. As the load balancer strength (i.e., the value of q) increases, fewer tasks are assigned to computers with higher temperatures, and more tasks are assigned to computers with lower temperatures.

[0067] The load balancer intensity q takes a discrete value between 0 and 1. An example of how to determine the specific distribution is shown below. Let T be the temperature distribution of the servers. That is, the temperature of the i-th server is T. i If so, the temperature distribution of N servers can be expressed as (Equation 20).

number

number

number

number

number

[0068] Furthermore, this embodiment introduces a new approach to power consumption. It focuses on the fact that, of the power consumed in a data center, the power consumption of air conditioners is the only power that can be optimized. In other words, as mentioned above, all tasks that should be processed in a data center are basically processed by computers within servers. This means that the power consumption of servers can be excluded from optimization, and by narrowing the target of optimization to the power consumption of air conditioners, it becomes possible to manage costs in a way that is more in line with the actual conditions of a data center.

[0069] Based on the above reasoning, we redefine the power consumption P using the formula shown in (Equation 25).

number

[0070] Based on the power consumption P defined in (Equation 25), the total electricity cost E for the data center's air conditioning is defined by the formula shown in (Equation 26). However, it is assumed that the indoor temperature near the air conditioning unit is the same for all air conditioning units.

number

[0071] Furthermore, in the present embodiment, a new concept is introduced for the equipment consumption cost. In the following description, the term trouble cost is used instead of equipment consumption cost. If a computer continues to operate at a temperature exceeding a certain level, the probability of failure increases. Therefore, when considering the long-term operation cost of a data center, the cost caused by such failure (for example, the purchase price of the hardware itself, equipment replacement work cost, etc.) can never be ignored. In addition to failure, problems such as a decrease in calculation speed and a decrease in calculation accuracy also occur during operation at high temperatures. The cost consisting of a plurality of factors (including factors not attributable to the computer) resulting from such high-temperature operation is collectively referred to as trouble cost. The trouble cost can be calculated as a function W(T) of the temperature T by considering the occurrence probability and cost of these factors.

[0072] Incidentally, although the occurrence probability of trouble differs for each factor, there are also cases where a plurality of factors are intricately intertwined to form another trouble factor. Furthermore, the specifications and processing capacity of individual computers are not uniform. Therefore, it is generally difficult to strictly calculate the trouble cost function W(T) as a function of the temperature T. On the other hand, the trouble cost itself is a parameter for which a rough calculation is possible, and has rough characteristics. For example, the trouble cost W(T) has a certain temperature T high up to (T high ≧T) takes a substantially zero value, and when T high is exceeded (T high <T) it comes to take a high value. Furthermore, as the temperature increases, the rate of increase of the trouble cost becomes larger. The trouble cost function W(T) can be estimated based on these characteristics and actually measured trouble costs (if any). Note that the trouble cost function W(T) may be obtained by extrapolating actually measured values of trouble cost, or may be predicted by the machine learning device 20 through machine learning using AI.

[0073] Next, we will explain how to obtain the approximate trouble cost W' by applying a linear approximation to the trouble cost function. Figure 8 is a graph illustrating the method for obtaining the approximate trouble cost W'. The horizontal axis of the graph shown in Figure 8 is T high The graph shows the temperature rise T (°C) from the source, and the vertical axis shows the trouble cost (for example, in units of 10,000 yen). In the example shown in Figure 8, the trouble cost function W(T) is W(T) = T 2 It is assumed that the trouble cost function W(T) can be expressed as a function of . However, the trouble cost function W(T) is not limited to this.

[0074] First, the temperature range T is divided into several sections. In the example shown in Figure 8, the range is divided into three sections (section 1, section 2, and section 3) based on the values ​​T=0, 3, 6, and 9. The method of division is not limited to this, and it can be divided at any value of T. Also, the interval between each section can be larger for lower temperatures and smaller for higher temperatures. Next, adjacent W(T) values ​​corresponding to T=0, 3, 6, and 9 are connected by straight lines, and a linear function W(T) = aT + b is obtained for each section. Here, the slope a of the linear function is given by w (Equation 17). i This corresponds to a value that represents how much wear and tear occurs for every 1 degree Celsius increase in the computer's temperature. The temperature of computer i at the start of the calculation is T. i Therefore, T i Since it can be measured, we can determine which division of the domain it belongs to. In the example in Figure 8, T i It belongs to category 2. i The slope a of the linear function W(T) = aT + b corresponding to section 2 to which it belongs is given by computer i's w i As shown in Figure 8, the computer's temperature, when the task is executed, is measured from the starting temperature T. i From being cooled by the air conditioner, the temperature T i It rises to '. That temperature T i Approximate trouble cost W'(T)(T)(T) from the linear function W(T)=aT+b corresponding to section 2 in ' i ') is obtained.

[0075] Thus, the approximate formula W' for the overall trouble cost can be formulated as shown in (Equation 27).

Formula

[0076] Therefore, as a QUBO-format evaluation function in the present embodiment, the total cost C can be expressed as shown in (Formula 28).

Formula

[0077] Note that in the present embodiment, the physical parameters to be actually observed are the trouble cost W(T), (Formula 16), T i, and X. W(T) can be calculated by multiplying the probability of occurrence for each trouble factor by the cost, or by plotting past trouble data and extrapolating, or by obtaining it from AI machine learning data. (Equation 16) can be obtained from Q, or from the specifications of the air conditioner, etc. T i This can be obtained from the computer's temperature sensor. For X, it is possible to predict the workload for each time period from past data.

[0078] The optimization method according to the above-described embodiment or other embodiments can be executed by program. The program causes a computer to execute a problem generation procedure for generating an optimization problem relating to the operating costs of a facility; a variable definition procedure for defining binary variables in order to execute the optimization problem in an annealing device; a physical parameter acquisition procedure for obtaining one or more physical parameter values ​​relating to the physical characteristics of the facility; a quadratic unconstrained binary optimization conversion procedure for converting the optimization problem into a quadratic unconstrained binary optimization problem using one or more physical parameter values; a quadratic unconstrained binary optimization problem transmission and reception procedure for transmitting and receiving the quadratic unconstrained binary optimization problem to and from the annealing device; and a facility operation control procedure for controlling the operation of the facility based on the solution obtained from the annealing device.

[0079] The program's information processing is concretely realized using hardware resources such as the control unit 12 (CPU), memory 14, storage device 16, and interface 18. The program may be provided as an installable or executable file, recorded on various computer-readable storage media. Alternatively, the program may be stored on a computer connected to a network such as the Internet and downloaded via the network, or provided directly via the network.

[0080] The embodiments, other embodiments, and modifications described above in the specification and drawings are merely examples of the present disclosure and do not limit the scope of the present disclosure. Furthermore, it will be known to those skilled in the art that various modifications and changes are possible without deviating from the spirit and nature of the present disclosure, and these are also included within the scope of the present disclosure. [Explanation of Symbols]

[0081] 1 Data center, 2 Server room, 3 Rack, 4 Air conditioner, 5 Power supply unit, temperature control unit, 6 Task distribution unit, 7 Network, 8 Operation instruction unit, 10 Optimization system, 11 Optimization device, 12 Control device, 14 Memory, 16 Storage device, 18 Interface, 19 Annealing device, 20 Machine learning device, 22 Input device, 24 Display device, 30 Floor panel, 31-36 Computer, 40 Cool air flow, 41 Cool air, 42 Exhaust, 43 Warm air flow

Claims

1. An optimization system having an annealing device, A problem generation unit that generates an optimization problem related to the operating costs of a facility, A physical parameter acquisition unit that acquires one or more physical parameter values ​​relating to the physical characteristics of the facility, A quadratic unconstrained binary optimization transformation unit that uses one or more of the aforementioned physical parameter values ​​to transform the optimization problem into a quadratic unconstrained binary optimization problem, Based on the solution to the quadratic unconstrained binary optimization problem obtained from the annealing device, a facility operation control unit controls the operation of the facility. Equipped with, The aforementioned quadratic unconstrained binary optimization transformation unit transforms the optimization problem into a quadratic unconstrained binary optimization problem that includes at least the total amount of electricity costs and equipment consumption costs for the facility. An optimization system characterized by the following features.

2. An optimization device that communicates with an external annealing device, A problem generation unit that generates an optimization problem related to the operating costs of a facility, A physical parameter acquisition unit that acquires one or more physical parameter values ​​relating to the physical characteristics of the facility, A quadratic unconstrained binary optimization transformation unit that uses one or more of the aforementioned physical parameter values ​​to transform the optimization problem into a quadratic unconstrained binary optimization problem, Based on the solution to the quadratic unconstrained binary optimization problem obtained from the annealing device, a facility operation control unit controls the operation of the facility. A processor that functions as, memory and Equipped with a computer having, The aforementioned quadratic unconstrained binary optimization transformation unit transforms the optimization problem into a quadratic unconstrained binary optimization problem that includes at least the total amount of electricity costs and equipment consumption costs for the facility. An optimization device characterized by the following features.

3. The optimization system according to claim 1, characterized in that the facility includes a load and a cooling device for cooling the load.

4. The optimization apparatus according to claim 2, characterized in that the facility includes a load and a cooling device for cooling the load.

5. The optimization system according to claim 3, characterized in that the one or more physical parameter values ​​include at least one of the current temperature of the load, the amount of temperature reduction of the load by the cooling device, the amount of temperature increase of the load per unit power consumption, the degree of wear of the load per unit temperature, the actual trouble cost, and the predicted amount of tasks per unit time.

6. The optimization apparatus according to claim 4, wherein the one or more physical parameter values ​​include at least one of the current temperature of the load, the amount of temperature reduction of the load by the cooling device, the amount of temperature increase of the load per unit power consumption, the degree of wear of the load per unit temperature, the actual trouble cost, and the predicted amount of tasks per unit time.

7. A problem generation process that generates an optimization problem related to the operating costs of a facility, A physical parameter acquisition step of acquiring one or more physical parameter values ​​relating to the physical characteristics of the said facility, A quadratic unconstrained binary optimization transformation step that transforms the optimization problem into a quadratic unconstrained binary optimization problem using one or more of the aforementioned physical parameter values, Based on the solution to the quadratic unconstrained binary optimization problem obtained from the annealing device, a facility operation control process is performed to control the operation of the facility. Equipped with, The aforementioned quadratic unconstrained binary optimization transformation step includes a step of transforming the optimization problem into a quadratic unconstrained binary optimization problem that includes at least the total amount of electricity costs and equipment consumption costs for the facility. The optimization method that a computer uses.

8. On the computer, A problem generation procedure for generating an optimization problem related to the operating costs of a facility, A physical parameter acquisition procedure for acquiring one or more physical parameter values ​​relating to the physical characteristics of the said facility, A quadratic unconstrained binary optimization transformation procedure that transforms the optimization problem into a quadratic unconstrained binary optimization problem using one or more of the aforementioned physical parameter values, A facility operation control procedure for controlling the operation of the facility, based on the solution to the quadratic unconstrained binary optimization problem obtained from the annealing device, Make it run, The aforementioned quadratic unconstrained binary optimization transformation procedure includes a step of transforming the optimization problem into a quadratic unconstrained binary optimization problem that includes at least the total amount of electricity costs and equipment consumption costs for the facility. program.

Images