Power amount reduction control device and power amount reduction control method

Abstract

A power amount reduction control device (100) is provided with: an air-conditioning control learning data generation unit (230) that generates, as air-conditioning control learning data (201), air-conditioning power consumption for air-conditioning control turns and an inlet temperature transition for each server temperature monitoring turn, which are associated with each air-conditioning situation; an APL division and arrangement unit (320) that generates an APL arrangement configuration pattern indicating the number of divisions and arrangement destinations of an APL; a server control learning data generation unit (340) that generates, as server control learning data (301), the minimum server FAN rotation control value and server power consumption in association with each server situation; and an optimal control calculation unit (240) that determines an APL arrangement configuration, and the minimum server FAN rotation control value of each accelerator-equipped server, as an optimal solution using the air-conditioning control learning data and the server control learning data.

Description

Power reduction control device and power reduction control method

[0001] The present invention relates to a power reduction control device and a power reduction control method for reducing the amount of power consumed in a data center (hereinafter sometimes referred to as "DC").

[0002] Air conditioning accounts for a large proportion of the power consumption in data centers (DCs), and as the number and size of DCs increases, there is a need to reduce the amount of power consumed by air conditioning. Furthermore, the amount of data processed in DCs is increasing year by year, making it necessary to improve the power consumption efficiency of the DCs as a whole (the amount of power consumed by the entire DC for a certain amount of data processing).

[0003] Furthermore, it is assumed that load processing will be performed within a DC in an environment where a central processing unit (CPU) server, accelerators, etc. are mixed. Here, accelerators are, for example, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), tensor processing units (TPUs), etc.

[0004] Meanwhile, Non-Patent Document 1 discloses a technology for optimizing the overall power consumption of a data center by taking into account the power consumption of air conditioners and servers (IT equipment). The datacenter-oriented air-conditioning-linked IT load placement optimization method described in Non-Patent Document 1 predicts future load trends on IT equipment by collecting operation and monitoring information from the datacenter's IT equipment, and calculates the power increase of the air-conditioning equipment in accordance with the increase in power consumption of the IT equipment. Then, an optimization problem is solved to minimize the objective function, which is the power consumption of the datacenter, so as to increase the load concentration rate on the IT equipment over time, i.e., to reduce the number of operating IT devices. This calculates the placement of IT loads (virtual machines) on the IT equipment that minimizes the power consumption of the datacenter.

[0005] Furthermore, a technique has been proposed for saving power and reducing the calculation processing speed by controlling the rotation speed of the fan provided in the server based on the predicted temperature of the server (Non-Patent Documents 2 and 3).

[0006] Jun Okitsu and four others, "IT Load Placement Optimization Method Linked with Air Conditioning for Environmentally Conscious Data Centers," FIT (Forum on Information Technology) 2010, 9th Information Science and Technology Forum, RC-009. Z. Wang, C. Bash, N. Tolia, M. Marwah, X. Zhu, and P. Ranganathan, "Optimal fan speed control for thermal management of servers," ASME InterPack Conference 2009, IPACK2009, vol. 2, pp. 709-719, 2010. X. Han and Y. Joshi, "Energy reduction in server cooling via real-time thermal control," in Annual IEEE Semiconductor Thermal Measurement and Management Symposium, pp. 74-81, 2012.

[0007] In servers equipped with accelerators set as DCs, one issue is board temperature rise due to heat generated by the accelerator while an application (APL) is running. This is because, generally, CPU servers are not equipped with a function to directly detect FPGA board temperature information, so even if the board temperature rises, the server's fan rotation cannot keep up, and it is expected that the application will stop running in some cases.

[0008] As a specific example, the experimental results shown in Figure 23 were obtained for an HPE ProLiant DL20 Gen10 server equipped with a single Arria10 FPGA manufactured by Intel ("Intel" is a registered trademark). Here, when the server's intake temperature was 19°C, Mandelborot (a fractal algorithm) was run at 534 MHz as an application to be processed by the FPGA. The manufacturer set the core temperature threshold to 100°C, the board temperature threshold to 80°C, and the supply temperature threshold to 120°C.

[0009] The default setting of the server-side fan rotation mechanism controls the fan rotation rate according to the CPU temperature. This means that the server cannot detect FPGA temperature information or increase the fan rotation rate in response to temperature rises. As a result, applications running on the FPGA may stop (e.g., the screen may freeze). As shown in Figure 23, the FPGA core temperature rose from 85°C before application execution to 98°C (symbol α in Figure 23) during application execution, approaching the core temperature threshold. However, the server was unable to detect FPGA board temperature information. Furthermore, because the rotation rates of the three fans (FAN 1, FAN 2, and FAN 3) are controlled according to the CPU temperature, if there is no change in CPU temperature before and during application execution, the fan rotation rate will not be increased (symbol β in Figure 23).

[0010] To address this issue, manufacturers recommend setting the server fan rotation rate to 100%. However, setting the server fan rotation rate to 100% increases the server's power consumption. In particular, in a data center that houses hundreds or thousands of servers, an increase in the server's power consumption significantly impacts the overall server power consumption of the data center. Therefore, it is necessary to relax the manufacturer-recommended fan rotation rate from 100% without exceeding the accelerator board temperature threshold.

[0011] Furthermore, when the intake temperature of a server equipped with an accelerator is set to a high temperature, i.e., when the air conditioning control is set to "weak," the minimum fan rotation rate on the server side that meets the board temperature threshold increases, resulting in increased power consumption on the server side. On the other hand, when the intake temperature is set to a low temperature, i.e., when the air conditioning control is set to "strong," the fan rotation rate on the server side is suppressed, but the power consumption of the air conditioner increases. In other words, there is a trade-off between the fan rotation rate of the accelerator-equipped server and the air conditioning control of the air conditioner in the DC room, and the prior art has not provided a method for determining the optimal settings to reduce the total power consumption in the DC room.

[0012] Furthermore, the power consumption of an FPGA-equipped server is expected to fluctuate depending on the processing characteristics (e.g., APL type, logic utilization, operating frequency) of the running FPGA APL (application), as well as differences in server fan rotation rate and CPU processing. Furthermore, by dividing the FPGA APL, the scale (circuit size) of the program configuration executed on the FPGA can be changed, enabling parallel distributed processing across multiple servers equipped with multiple divided FPGA APL processing circuits. This reduces the power consumption per server, potentially resulting in a reduction in total power consumption. However, prior art did not consider the impact of distributed FPGA APL processing on the power consumption of each FPGA-equipped server.

[0013] The present invention was made in consideration of these points, and its objective is to reduce the total power consumption, consisting of server power consumption and air conditioning power consumption, in a DC room equipped with a server equipped with an accelerator.

[0014] The power consumption reduction control device according to the present invention is a power consumption reduction control device that controls accelerator-equipped servers and air conditioners in a room of a data center, and the room has a server area that represents a group of areas that process loads, and each of the servers is provided with one or more air conditioners, and the power consumption reduction control device controls the board temperature threshold of the accelerator included in the accelerator-equipped server to satisfy an SLA (Service Level Agreement). an air-conditioning situation definition unit that defines an air-conditioning situation by combining each of the divided ranges for the server load information, the outside air temperature, and the initial room temperature with each stage of the air-conditioning control level; an air-conditioning control unit that defines an air-conditioning situation by defining an air-conditioning control turn as a predetermined control time until the air-conditioning control value is updated, and by executing control of the accelerator-equipped servers and air-conditioners in the room based on the classification indicated by each of the air-conditioning situations, an air conditioning control learning data generation unit that acquires information on the change in air intake port temperature of each accelerator-equipped server for each server temperature monitoring turn, which indicates a predetermined time shorter than the air conditioning control turn, and generates, as air conditioning control learning data, the air conditioning power consumption for the air conditioning control turn and the change in air intake port temperature of each accelerator-equipped server for each server temperature monitoring turn, in association with each air conditioning situation; a server operating status monitoring unit that monitors the accelerator board temperature of the accelerator-equipped server; an APL division placement unit that generates APL placement configuration patterns by changing the number of divisions of an APL (Application) to be mounted on the accelerator and the accelerator to which the APL is to be placed;a server situation definition unit that defines a server situation by combining predetermined operating states including the APL arrangement configuration of the accelerator-equipped server that are set in advance; a server control learning data generation unit that searches for a minimum server fan rotation control value within a range that does not exceed a board temperature threshold indicated in the SLA information by operating the accelerator-equipped server for each server temperature monitoring turn in each of the server situations, and generates the minimum server fan rotation control value and the server power consumption of the accelerator-equipped server at that time as server control learning data in association with each server situation; and a server control learning data generation unit that distributes the load indicated in the server load information to the predetermined air conditioning control level and each of the server zones. and an optimal control calculation unit that, in an operation stage when the amount of load to be placed is fixed, calculates the change in the intake port temperature and the air conditioning power consumption for each air conditioning control level using the air conditioning control learning data based on the server load information in the room, the outside air temperature, and the initial room temperature information, and calculates the minimum server fan rotation control value and the server power consumption of each accelerator-equipped server for each server temperature monitoring turn using the server control learning data according to the server operating state including the intake port temperature and the APL placement configuration, thereby calculating a total value of server power consumption within the server area, and determines the APL placement configuration and the minimum server fan rotation control value of each accelerator-equipped server for a pattern that minimizes the total value as an optimal solution.

[0015] According to the present invention, in a DC room equipped with a server equipped with an accelerator, the total power consumption, which is made up of the power consumption of the server and the power consumption of the air conditioning, can be reduced.

[0016] 1 is a diagram showing the minimum fan rotation speed of a server that satisfies an SLA for each air intake temperature.

[0023] FIG. 1 is a diagram showing the percentage of power consumption reduction when the fan rotation speed is set to the minimum fan rotation speed, compared to a fan rotation speed of 100%.

[0024] FIG. 1 is a diagram comparing power consumption according to the number of operating servers (load in a DC room) at different server air intake temperatures.

[0025] FIG. 1 is a diagram showing the relationship between air conditioning power consumption and server power consumption, the total power consumption in a DC room according to the server load in the room.

[0026] FIG. 1 is a diagram showing the difference in power consumption of an FPGA-equipped server when the processing characteristics of the FPGA APL being operated are different.

[0027] FIG. 1 is a diagram showing the total server power consumption when an FPGA APL is divided and installed in different servers.

[0028] FIG. 1 is a diagram explaining the difference in server minimum fan rotation value of each server when the configurations in which the FPGA APL is installed are different.

[0029] FIG. 2 is a diagram explaining the difference in total server power consumption when the configurations in which the FPGA APL is installed are different.

[0029] FIG. 2 is a diagram explaining a framework for dividing an FPGA APL for distributed processing.

[0029] FIG. 3 is a diagram showing the overall configuration of a power amount reduction control system including a power amount reduction control device according to an embodiment of the present invention. FIG. 1 is a functional block diagram showing an example of the configuration of a power amount reduction control device according to the present embodiment. FIG. 2 is a diagram illustrating air conditioning control levels according to the present embodiment. FIG. 3 is a diagram for explaining the contents of air conditioning control learning data according to the present embodiment. FIG. 4 is a diagram showing an example of APL division and placement. FIG. 5 is a diagram showing classification definitions of server situations according to the present embodiment. FIG. 6 is a diagram showing an example of SLA information according to the present embodiment. FIG. 7 is a diagram showing an example of a search for minimum fan rotation control values ​​for each server situation classification according to the present embodiment. FIG. 8 is a diagram for explaining a process for determining an APL placement configuration according to the present embodiment. FIG. 9 is a diagram for explaining a process for calculating an optimal solution for air conditioning control, load placement, fan control, and APL placement configuration according to the present embodiment. FIG. 10 is a flowchart showing the flow of processes executed in a learning phase by a power amount reduction control device according to the present embodiment. FIG. 11 is a flowchart showing the flow of processes executed in an operation phase by a power amount reduction control device according to the present embodiment. FIG. 12 is a hardware configuration diagram showing an example of a computer that realizes the functions of a power amount reduction control device according to the present embodiment. FIG. 13 is a diagram showing an increase in accelerator board temperature in conventional server FAN control based on CPU temperature.

[0017] Next, an embodiment of the present invention (hereinafter referred to as "the present embodiment") will be described. First, as a premise, (1) the relationship between the fan rotation control value (fan rotation rate) of an accelerator-equipped server and the air conditioning control of an air conditioner in a DC room, and (2) the reduction in total power consumption due to the reduction in fan rotation of an offload destination server and the reduction in FPGA-specific power consumption by dividing FPGA APL will be described.

[0018] <Relationship between Fan Rotation Rate and Air Conditioning Control> As mentioned above, in servers equipped with accelerators (hereafter, FPGAs as an example), it is not possible to detect FPGA temperature information or increase the fan rotation rate in response to temperature rises. For this reason, manufacturers recommend setting the server fan rotation rate to 100%. However, based on the technical concept that it may be possible in some cases to run applications at a fan rotation rate below the manufacturer-recommended 100% fan rotation rate without exceeding the FPGA board temperature threshold, we conducted the following measurements to find the minimum fan rotation rate that does not exceed the board temperature threshold (hereinafter, sometimes referred to as "meeting the SLA (Service Level Agreement)").

[0019] We ran the Mandelbrot (534.8 MHz) application on a server equipped with a single Ariia 10 FPGA in environments with different intake temperatures (19°C and 35°C). We measured the minimum fan speed (server minimum fan speed) that met the SLA for each intake temperature and the server's power consumption at that time. In this case, we monitored the FPGA board temperature threshold. If any of the board temperature thresholds (core temperature, board temperature, or supply temperature) was exceeded, the SLA was not met and the server was deemed a failure. The FPGA board temperature monitoring threshold can be the manufacturer-specified upper limit, or it can be set independently based on the manufacturer's upper limit. For example, we set the monitoring threshold to 70% of the manufacturer-specified upper limit.

[0020] As shown in Figure 1, by monitoring the FPGA board temperature threshold, the minimum fan rotation rate of the server that met the SLA was 30% when the air intake temperature was 19°C, and 65% when the air intake temperature was 35°C. In other words, we confirmed that it is possible to set a fan rotation rate lower than the manufacturer's recommended rate (100%) without exceeding the board temperature threshold.

[0021] Furthermore, as shown in Figure 2, when the intake temperature is 19°C and the fan rotation speed is set to the minimum of 30%, power consumption can be reduced by approximately 35% compared to when the fan rotation speed is set to 100%. Also, when the intake temperature is 35°C and the fan rotation speed is set to the minimum of 65%, power consumption can be reduced by approximately 26% compared to when the fan rotation speed is set to 100%.

[0022] Furthermore, as mentioned above, there is a trade-off between the fan rotation rate of a server equipped with an accelerator (FPGA) and air conditioning control. Setting the server's air intake temperature high, i.e., setting the air conditioning to "low," increases the minimum fan rotation rate on the server side that meets the board temperature threshold, thereby increasing the server's power consumption. On the other hand, setting the air intake temperature low, i.e., setting the air conditioning to "high," reduces the server's fan rotation rate but increases the air conditioning's power consumption. When calculating the total power consumption in the DC room based on this, as shown in Figure 3, the relationship between the intake temperature and the total power consumption in the DC room reverses at a boundary of 390 servers (i.e., the number of servers running FPGAs (= DC room load)). In other words, when there are fewer than 390 servers in operation, setting the air conditioning control to "weak" (air intake temperature "high") reduces total power consumption and is therefore desirable, while when there are 390 or more servers in operation, setting the air conditioning control to "strong" (air intake temperature "low") reduces total power consumption and is therefore desirable.

[0023] Figure 4 shows the total power consumption in a DC room as a function of the server load in the room, as a function of the air conditioning power consumption and the server power consumption. When the server load in the DC room is small, as described above, setting the air conditioning control to "weak" to reduce the air conditioning power consumption and increase the fan speed reduces the total power consumption (symbol X in Figure 4). Also, when the server load in the DC room is large, setting the air conditioning control to "strong" to reduce the fan speed reduces the total power consumption (symbol Y in Figure 4). These results suggest that it is necessary to simultaneously control the air conditioning control value (control parameter <1>) and server fan rotation control (control parameter <2>) according to the server load in the DC room, and to derive a control solution that minimizes the total air conditioning power consumption and server power consumption.

[0024] <Reducing Total Power Consumption by Dividing FPGA APLs> Next, we will explain how dividing FPGA APLs reduces total power consumption. The power consumption of an FPGA-equipped server varies depending on the processing characteristics of the FPGA APLs it runs, as well as the server fan rotation rate and CPU processing. For example, as shown in Figure 5, even if the running APL is the same (Mandelbrot), if the operating frequency as a processing characteristic varies, the server's minimum fan rotation value (65%, 50%) will also differ, resulting in different server power consumption (110.04 kW, 96 kW). Furthermore, different running applications can significantly affect the power consumption of a server, depending on the application's processing characteristics. For example, if the running APL is Particles (an OpenCL sample application), the logic utilization (circuit size number (circuit utilization rate)) and operating frequency will vary, resulting in different server minimum fan rotation values ​​and different server power consumption (127.67 W) from the Mandelbrot value. Therefore, when calculating server power consumption, it is necessary to consider the type of running APL and its processing characteristics.

[0025] Furthermore, FPGA APLs can be divided. By dividing FPGA APLs, the scale (circuit size) of the program configuration executed on the FPGA can be changed. Dividing the FPGA APL processing circuit into multiple parts and performing parallel distributed processing on multiple servers reduces the CPU processing load and the heat generated during APL processing due to the smaller circuit size, thereby mitigating the rise in board temperature and lowering the server fan rotation control value. This reduces the amount of power consumed per server, and is expected to ultimately reduce total power consumption.

[0026] 6 shows an example of a two-server configuration in which an FPGA APL is installed on one server and in which the FPGA APL is split and installed on two servers. In the diagram indicated by reference numeral 61, one server is equipped with FPGA circuits (a + b), and the server power is 150 W. On the other hand, the server power of the server without an FPGA circuit is 80 W. Therefore, the total server power is 230 W. In contrast, as shown in the diagram indicated by reference numeral 62, when the FPGA APL is split and FPGA circuit a and FPGA circuit b are installed on each server, the total server power is 95 + 85 = 180 W. Therefore, splitting the FPGA APL reduces the total server power.

[0027] Regarding the division of FPGA APL, we installed the OpenCL sample application Particles on servers in configurations A to C shown in Figure 7 below, measured server power consumption, and confirmed the effectiveness of dividing the FPGA APL. (Configuration A) is a configuration in which the FPGA APL (Particles) is processed on a single FPGA-equipped server (Server 1). (Configuration B) is a configuration in which the offload source server (Server 1) does not have an FPGA, and the FPGA APL (Particles) is processed on a single FPGA-equipped server (Server 2). (Configuration C) is a configuration in which the offload source server (Server 1) does not have an FPGA, and the FPGA APL (Particles) is divided and processed on two servers (Server 1 and Server 2).

[0028] The configuration shown in Figure 7 assumes that the FPGA-equipped server operates at a minimum server fan rotation value that takes into account the FPGA temperature. In the example shown in Figure 7, the minimum server fan rotation value (rate) of the FPGA-equipped server is 30%. It is assumed that the appropriate granularity for splitting FPGA APLs is one that allows them to operate at a fan rotation rate close to the minimum server fan rotation value (rate) when the FPGA is not operating (details will be provided later). It is also assumed that the larger the minimum server fan rotation value of an APL when not split, the greater the power reduction benefit when splitting.

[0029] 7, in configuration A, the minimum fan rotation value was 40% for the FPGA APL processing server (Server 1), and 30% for the servers not equipped with FPGA APL (Server 2, Server 3). In configuration B, the minimum fan rotation value was 10% for the server not equipped with FPGA (Server 1), 40% for the FPGA APL processing server (Server 2), and 30% for the server not equipped with FPGA APL (Server 3). In configuration C, the minimum server fan rotation value of the server without FPGA (Server1) was 10%, the minimum server fan rotation value of the processing server (Server2) for the divided FPGA APL (Particles.aocx) was 40%, and the minimum server fan rotation value of the processing server (Server3) for the similarly divided FPGA APL (BitonicSort_b.aocx) was 30%.

[0030] Figure 8 shows the measurement results of total server power consumption for configurations A, B, and C. The total server power consumption for configuration B, where the offload source server does not have an FPGA, was 243.73 W. This represents a 13.2% reduction compared to the 280.79 W total server power consumption for configuration A, where FPGA applications were run independently on an FPGA-equipped server. Furthermore, the total server power consumption for configuration C, where FPGA APLs were split and executed, was 238.77 W. This represents a 15.0% reduction compared to the total server power consumption for configuration A. Potential factors contributing to this power reduction include the reduction in server fan speed due to the non-FPGA offload source server, the reduction in fan speed on the offload destination server due to APL splitting, and the reduction in FPGA power consumption. Furthermore, further power reductions are possible by using the offload source server as a controller and superimposing it on other services.

[0031] The power consumption reduction control device according to this embodiment reduces the APL function granularity, i.e., divides FPGA APLs, to build a framework for distributed processing in order to reduce the power consumption of each FPGA-equipped server. Specifically, as shown in FIG. 9 , the power consumption reduction control device determines an appropriate number (N) of FPGA APL divisions to prevent an increase in server fan rotation. The power consumption reduction control device then prioritizes deployment of FPGA-divided applications to servers in areas (districts) with good cooling efficiency, for example, by avoiding hot spots. The power consumption reduction control device maintains an appropriate server intake temperature, taking into account the trade-off between server cooling and air conditioning cooling.

[0032] In order for this power reduction control device to reduce the power consumption of the servers and air conditioning in the entire DC, the following parameters must be considered: - Number of kernel divisions of FPGA APL - Server to which the FPGA kernel is deployed (affects server fan rotation and air conditioning cooling efficiency) - Air conditioning control level (affects server air intake temperature) - Fan rotation of each server These parameters must be controlled appropriately. Specifically, the solution that minimizes the total power consumption obtained by the following equation (1) is found.

[0033]

[0034] Here, Psi is the power of server Si, Rsi is the fan rotation rate of server Si while it is running, tsi is the air intake temperature of server Si, Am is the type of FPGA in operation (including FPGA configuration), NAm is the number of divisions of FPGA APL, n(j) is the deployed FPGA kernel, PA is the power consumption of air conditioning, and C is the air conditioning control level.

[0035] In a DC room equipped with servers equipped with accelerators, the power consumption reduction control device according to this embodiment appropriately monitors accelerator board temperature information and controls the air conditioning control value and server fan rotation speed according to the server load in the room, within a range that does not exceed the board temperature threshold (within a range that does not violate the SLA), thereby reducing the total power consumption made up of the air conditioning power consumption and the server power consumption. Furthermore, by dividing and distributing the APL, the power consumption reduction control device can further reduce the total power consumption by appropriately adjusting the strength of the air conditioning control and server fan rotation speed control according to the total load in the room, using the server intake port temperature as an index.

[0036] <Electricity Amount Reduction Control System> Next, an embodiment of the present invention (hereinafter referred to as "the present embodiment") will be described in detail. Fig. 10 is a diagram showing the overall configuration of an electric power amount reduction control system 1 including an electric power amount reduction control device 100 according to the present embodiment.

[0037] As shown in FIG. 10 , the power consumption reduction control system 1 includes a data center (DC 10) having a plurality of servers 3 housed in a predetermined control area (a "placement control area" (server area) described below) and a plurality of air conditioners 4, and a power consumption reduction control device 100. Hereinafter, the plurality of servers 3 housed in the predetermined control area (placement control area) may be referred to as a "server group" or collectively as "servers." The server 3 is an accelerator-equipped server equipped with an accelerator 31 (which is assumed to be an FPGA in this embodiment). The server 3 also has its own CPU (not shown) and a fan 32 for cooling the accelerator 31. Although not shown, the server 3 may not be equipped with the accelerator 31 (FPGA) as an offload source server.

[0038] This power amount reduction control device 100 may be provided inside the DC 10, or may be provided in a location separate from the DC 10. Note that although Fig. 10 illustrates one room in the DC 10, the DC 10 may include multiple rooms, and the power amount reduction control device 100 executes the power amount reduction control described below for each room in the DC 10 (hereinafter referred to as a "DC room").

[0039] The power consumption reduction control device 100 may acquire status information of the air conditioners 4 (air conditioners “1,” “2,” and “3” in FIG. 10 ) provided within the DC 10 and transmit air conditioning control information via an air conditioning management device (not shown), or may be directly connected to and communicate with each air conditioner 4 without going through an air conditioning management device. The power consumption reduction control device 100 may also acquire status information of and transmit control information for servers 3 (accelerator-equipped servers) accommodated as a server group provided within the DC 10 via a server management device (not shown), or may be directly connected to and communicate with each server 3. Each server 3 has a function for monitoring the board temperature of the accelerator 31 provided therein and transmits the measured board temperature to the power consumption reduction control device 100.

[0040] In the DC 10 of this embodiment, the entire server (server group) accommodated therein may be divided into areas where each server 3 is placed, and managed as "placement control areas 30," as shown in FIG. 10. FIG. 10 shows an example in which placement control areas "1" to "6" are provided. The "placement control areas 30" are also referred to as "server areas."

[0041] In this embodiment, "air-conditioning controlled areas 40" are provided in association with the placement-controlled areas 30 of the server groups, as shown in FIG. 10 . The air-conditioning controlled areas 40 are areas in which the room temperature effect of air-conditioning control is measured, and are assumed to face either the inlet side or the outlet side of each server 3. Air blown from the air conditioners 4 is blown out from the air-conditioning controlled areas 40 on the inlet side (air-conditioning controlled areas "3" and "4" in FIG. 10 ) via piping installed, for example, under the floor of the DC 10. Air whose temperature has risen due to the heat from each server 3 is taken in from the suction ports of the piping installed in the air-conditioning controlled areas 40 on the outlet side (air-conditioning controlled areas "1," "2," "5," and "6" in FIG. 10 ), generating an airflow that returns to the air conditioners 4.

[0042] A plurality of sensors (such as temperature sensors) are installed in each of the air-conditioning controlled areas 40. A temperature sensor may also be installed at the air intake of the server 3 (accelerator-equipped server) in each placement controlled area 30 (server area). Furthermore, a sensor (such as a temperature sensor) is also installed outside the DC 10. Information obtained from these sensors (sensor information) can be acquired by the power reduction control device 100 via a communication line or the like.

[0043] It is not essential to provide the air-conditioning controlled area 40, but it is sufficient that a temperature sensor is placed in the DC room in association with each server 3 so that the initial temperature in the room and the temperature at the server air intake port, which will be described later, can be measured. Then, the power amount reduction control device 100 acquires temperature information from the temperature sensor associated with each server 3.

[0044] The power consumption reduction control device 100 according to this embodiment appropriately monitors the server 3's air intake port temperature, accelerator board temperature information, and other information for each server 3 installed in the DC room, and controls the air conditioning control value and server fan rotation speed according to the server load in the room, within a range that does not exceed the board temperature threshold (within a range that does not violate the SLA). In this way, the power consumption reduction control device 100 reduces the total power consumption in the DC room, consisting of the air conditioning power consumption and the server power consumption. Furthermore, by dividing and distributing the APL, the power consumption reduction control device 100 further reduces total power consumption by appropriately adjusting the strength of the air conditioning control and server fan rotation speed control according to the total load in the room, using the server air intake port temperature as an index. The power consumption reduction control device 100 will be described in detail below.

[0045] <Power consumption reduction control device> Fig. 11 is a functional block diagram showing an example of the configuration of a power consumption reduction control device 100 according to this embodiment. The power consumption reduction control device 100 performs control in two stages: a learning phase and an operation phase. In the learning phase (learning stage), the power consumption reduction control device 100 collects learning data related to room air conditioning control and server fan rotation control. Here, the learning data related to room air conditioning control ("air conditioning control learning data 201" described below) is the cooling status (air conditioning power consumption and server intake temperature transition) for one air conditioning control turn for each outside air temperature, room temperature, and air conditioning control level relative to the server load in the room (details will be described later). In addition, the learning data regarding server fan rotation control ("server control learning data 301" described below) is the minimum server fan rotation rate (server minimum fan rotation value) that does not exceed the accelerator 31 board temperature threshold, for each air intake temperature and for each API arrangement configuration within the server area (details described below), in one server temperature monitoring turn of each server 3, and the server power consumption at that time (details described below).

[0046] In the operation phase, the power reduction control device 100 uses air conditioning control learning data 201 to calculate the air conditioning power consumption and server inlet temperature trends for each air conditioning control level based on the server load and initial temperature information in the room. The power reduction control device 100 then uses server control learning data 301 to calculate the power consumption of each server 3 according to the server inlet temperature, the APL layout configuration in the server area, and the like. The power reduction control device 100 then calculates the total value of the air conditioning power consumption and the server power consumption for a pattern (hereinafter referred to as a "control pattern") that combines air conditioning control, server area load layout, server fan control, and APL layout. The power reduction control device 100 then calculates the pattern of air conditioning control, server area load layout, server fan control, and APL layout that minimizes the total value as the optimal solution.

[0047] The power reduction control device 100 is configured by a computer having a control unit, an input / output unit, and a storage unit (all of which are not shown).

[0048] The input unit inputs and outputs information between each device (each air conditioner 4 and each server 3 (accelerator-equipped server)) in the DC 10. This input / output unit is composed of a communication interface that transmits and receives information via a communication line, and an input / output interface that inputs and outputs information between an input device such as a keyboard and an output device such as a monitor (not shown).

[0049] The storage unit is composed of a hard disk, flash memory, RAM (Random Access Memory), etc. This storage unit temporarily stores programs for executing each function of the control unit and information necessary for the control unit's processing. This storage unit also stores learning data related to room air conditioning control (air conditioning control learning data 201), learning data related to server fan rotation control (server control learning data 301), and SLA information 302 indicating information on the board temperature thresholds (e.g., core temperature, board temperature, supply temperature) of each accelerator 31. Information on the server load placed in the room during each air conditioning control turn is acquired from a system management device (not shown) or the like and stored in this storage unit.

[0050] The control unit is responsible for all the processes executed by the power amount reduction control device 100, and as shown in FIG. 11, includes an integrated control unit 200, a server control unit 300, and an air conditioning control unit 400.

[0051] <<Integrated Control Unit>> In the learning phase, the integrated control unit 200 generates air-conditioning control learning data 201, which represents the cooling status (air-conditioning power consumption and server inlet temperature transition) for one air-conditioning control turn for each room server load, including the outside air temperature, room temperature, and air-conditioning control level. In the operation phase, the integrated control unit 200 uses the air-conditioning control learning data 201 and server control learning data 301 generated by the server control unit 300 to determine a pattern of air-conditioning control, load allocation to server zones, server fan rotation control, and APL arrangement that minimizes the sum of air-conditioning power consumption and server power consumption, and causes each server 3 and each air conditioner 4 to execute the air-conditioning control, load allocation to server zones, server fan rotation control, and APL arrangement. The integrated control unit 200 includes a room status acquisition unit 210, an air-conditioning situation definition unit 220, an air-conditioning control learning data generation unit 230, an optimal control calculation unit 240, and a fail-safe control unit 250.

[0052] The room status acquisition unit 210 acquires information about server load, outside air temperature, and initial room temperature. The server load information is, for example, information acquired from the system management device of the DC 10, and is quantified based on information such as the number of processed requests (300 to 400 per minute), total processing traffic volume, the types and number of active APLs, and the number of active virtual resources. The outside air temperature is information acquired by a temperature sensor installed outside the DC 10. The initial room temperature is, for example, the average temperature obtained from a fixed-point sensor in the room, the average sensor temperature in each air-conditioning-controlled zone 40 in the room (see FIG. 10), the average temperature on the intake or exhaust side of the room, etc., and the temperature to be used is set in advance.

[0053] The air conditioning situation definition unit 220 divides the information acquired by the room status acquisition unit 210 into multiple ranges between a minimum value and a maximum value according to the characteristics of the information. For example, if the server load information is the number of requests processed, the range is divided into ranges in increments of 100, such as 0 to 99 requests / minute, 100 to 199 requests / minute, 200 to 299 requests / minute, etc. For the outside air temperature or the initial room temperature, the range is divided into ranges in increments of 3°C, such as 19 to 21°C, 22 to 24°C, 25 to 27°C, etc.

[0054] The air-conditioning situation definition unit 220 also defines the air-conditioning control levels of the air conditioner 4 by dividing the control values ​​(air-conditioning control values) such as the set temperature and air volume into stages within a certain range. FIG. 12 is a diagram illustrating examples of air-conditioning control levels as stages (LV). As shown in FIG. 12, air-conditioning control level stage (LV) "1" has a target set temperature of "32°C" and air volume of "24 Hz." Air-conditioning control level stage (LV) "2" has a target set temperature of "28°C" and air volume of "30 Hz." The following stages (up to stage "5" in this case) are set in advance.

[0055] The air-conditioning situation definition unit 220 then defines a combination of the divided ranges for each piece of information, such as server load information, outside temperature, initial room temperature, and air-conditioning control level, as one air-conditioning situation.

[0056] Returning to FIG. 11 , the air-conditioning control learning data generator 230 defines one cycle (predetermined control time) until the air-conditioning control value is updated as an air-conditioning control turn. The air-conditioning control learning data generator 230 then executes control of each server 3 and air conditioner 4 based on information defined for each air-conditioning situation (server load information, outside temperature, initial room temperature, and air-conditioning control level), thereby acquiring the total power consumption of the air conditioner 4 and the trends in the inlet temperature (server inlet temperature) of each server 3 during the corresponding air-conditioning control turn as learning data (air-conditioning control learning data 201). The trends in the inlet temperature are measured for each server temperature monitoring turn. The temperature monitoring turn is set shorter than the air-conditioning control turn because the SLA for board temperature monitoring is stricter. One air-conditioning control turn consists of multiple server temperature monitoring turns. Furthermore, the control of the server fan rotation can be updated for each server temperature monitoring turn.

[0057] Figure 13 shows the learning results for a certain air-conditioning situation. This example shows three server temperature monitoring turns in one air-conditioning control turn. In Figure 13, the control of each server 3 and air conditioner 4 is performed based on the information defined in the air-conditioning situation (server load information, outdoor temperature, initial room temperature, and air-conditioning control level). As a result, the total power consumption of the air conditioner 4 during the air-conditioning control turn is measured to be 7.5 kWh. Furthermore, at the end of each server temperature monitoring turn, the intake temperature of server "1" is measured to have changed from "19-21°C," "19-21°C," to "22-24°C," the intake temperature of server "2" is measured to have changed from "22-24°C," "22-24°C," to "25-27°C," and the intake temperature of server "3" is measured to have changed from "19-21°C," "22-24°C," to "22-24°C."

[0058] The air conditioning control learning data generation unit 230 stores the air conditioning power consumption of this air conditioning control turn and information on the change in air intake temperature at each server 3 for each server temperature monitoring turn as air conditioning control learning data 201 in the memory unit, in association with the air conditioning situation.

[0059] 11 , in the operation phase, the optimal control calculation unit 240 derives optimal solutions for patterns of air conditioning control, load allocation to server areas, server fan rotation control, and APL layout configuration that minimize the sum of air conditioning power consumption and server power consumption based on the server load in the DC room and initial temperature information (outdoor air temperature, initial temperature in the room), and issues control instructions to the air conditioners 4 and servers 3. This optimal control calculation unit 240 includes a control pattern power consumption calculation unit 241 and an optimal control determination unit 242.

[0060] The control pattern power amount calculation unit 241 uses the air-conditioning control learning data 201 to predict each server inlet temperature (server inlet temperature transition) and air-conditioning power consumption when the server inlet initial temperature and air-conditioning control level are changed for each load allocation pattern of each server area in response to the server load generated within DC 10. The control pattern power amount calculation unit 241 then calculates, as an optimal solution, the server minimum fan rotation value for each server based on the APL allocation configuration that satisfies the SLA and the APL allocation configuration pattern. As a result, the control pattern power amount calculation unit 241 calculates the total value of the air-conditioning power consumption and the server power consumption in the corresponding air-conditioning control turn for each pattern of air-conditioning control, load allocation to server areas, server fan rotation control, and APL allocation configuration (hereinafter referred to as a "predetermined control pattern").

[0061] In addition, when the air conditioning control level is fixed to a predetermined level within DC 10 and the amount of load allocated to each server area is fixed, the control pattern power amount calculation unit 241 determines a pattern that minimizes the total server power consumption within the server area from among the allocation configuration patterns that change the number of divisions and APL placement locations of the APLs that make up the load for the amount of load to be allocated to that server area, i.e., the API placement configuration within the server area, based on the server control learning data 301. Then, the control pattern power amount calculation unit 241 determines the API placement configuration (number of divisions and APL placement locations) of that pattern and the minimum fan rotation speed that satisfies the SLA of each server as the optimal solution (details will be described later with reference to FIG. 18).

[0062] The optimal control determination unit 242 determines the control pattern with the smallest total power consumption from the total values ​​of the air conditioning power consumption and the server power consumption calculated for each predetermined control pattern as the optimal solution.The optimal control determination unit 242 then determines the air conditioning control, load allocation to the server zones, server fan rotation control value (server minimum fan rotation value), and APL layout configuration for the determined control pattern as the optimal solution, and transmits control instructions to the air conditioners 4 and each server 3.Details of the process by the optimal control calculation unit 240 to calculate the optimal solutions for the air conditioning control, load allocation to the server zones, server fan rotation control (server minimum fan rotation value), and APL layout configuration pattern will be described later with reference to FIG. 19, etc.

[0063] When the board temperature of the accelerator 31 mounted on each server 3 in operation exceeds a threshold value, causing an SLA violation, and an alert is received from a failsafe execution unit 370 of the server control unit 300 (described later), the failsafe control unit 250 controls the air conditioner 4 to forcibly strengthen the air conditioning control (strengthen to a predetermined level). The air conditioning control of the air conditioner 4 when an SLA violation occurs and an alert is received is predefined, such as setting the air conditioning control level ( FIG. 12 ) one level higher than the current setting, setting the target temperature to a predetermined temperature lower, or setting the air volume to a predetermined percentage. It may also be predefined that the air conditioning control of the air conditioner 4 is set to the strongest control level. Furthermore, even when the fail-safe control unit 250 does not receive an alert from the fail-safe execution unit 370, if the server operation status monitoring unit 310 acquires information that the board temperature of the accelerator 31 installed in each server 3 currently in operation has exceeded a threshold value, the fail-safe control unit 250 can also forcibly strengthen the air conditioning control of the air conditioner 4.

[0064] <Server Control Unit> The server control unit 300 stores, for each server 3, the minimum server fan rotation rate that does not exceed the accelerator 31 board temperature threshold (SLA information) and the server power consumption at that time, as learning data (server control learning data 301), for each intake port temperature of one server temperature monitoring turn and for each server operating state (number of operating accelerators, operating APL types, APL division number, APL placement destination, etc.). The server control unit 300 includes a server operating status monitoring unit 310, an APL division placement unit 320, a server situation definition unit 330, a server control learning data generation unit 340, an APL placement control unit 350, a server cooling control unit 360, and a fail-safe execution unit 370.

[0065] The server operational status monitoring unit 310 monitors the board temperature of the accelerator 31 at a cycle (predetermined time) of the server temperature monitoring turn. Note that this server temperature monitoring turn is shorter than the air conditioning control turn as described above, and each server temperature monitoring turn is a timing at which the control of the server fan rotation can be updated.

[0066] The APL division placement unit 320 determines the upper limit of the number of divisions for the APL (FPGA APL) installed in the accelerator 31 such as an FPGA. This upper limit of the number of divisions is determined by first checking the server minimum fan rotation value that does not violate the SLA when the server 3 equipped only with the accelerator 31 is in an APL non-operating state, and then increasing the number of divisions by one at a time until the server fan rotation value reaches the server minimum fan rotation value. For example, if the number of divisions for the FPGA APL is "3," that is, when the APL is distributed and allocated to three servers 3, and the fan rotation value of each server becomes the server minimum fan rotation value, the upper limit of the number of divisions will be "3." The adjustment range of the API division number is 1 to N, where N is the upper limit of the number of divisions described above.

[0067] The APL division and placement unit 320 also generates a pattern of "APL placement configuration within a server area" (hereinafter referred to as "APL placement configuration" or "APL placement configuration pattern") in which the number of divisions of the APLs constituting the load and the APL placement destinations are changed. This APL placement configuration pattern indicates in which accelerators 31 the divided APLs are placed.

[0068] Fig. 14 is a diagram showing an APL allocation configuration within a server area (APL allocation configuration) when two applications, APL_A and APL_B, are allocated as loads within the server area. Here, the number of divisions at the time when the APL division allocation unit 320 reaches the minimum server fan rotation value in the APL non-operating state is assumed to be "3" for APL_A and "2" for APL_B. In the example shown in Fig. 14, physical servers A, B, and C are each equipped with one FPGA, and physical server D is equipped with two FPGAs. APL_A is then divided into three (" <3> ") and APL_B is divided into two parts (indicated by " <2> "). Then, the divided APL_A-1 <3> is placed in the FPGA of physical server A and divided into APL_A-2 <3> is placed in the FPGA of physical server D, and the divided APL_A-3 <3> is placed in the FPGA of physical server B. Also, the divided APL_B-1 <2> is placed in the FPGA of physical server C, and the divided APL_B-2 <2> is placed in the other FPGA of physical server D. The APL division and placement unit 320 generates all patterns of APL placement configurations as shown in FIG.

[0069] The server situation definition unit 330 defines the situation of the server 3 according to each range of the server intake temperature and a predetermined server operating state. Here, the predetermined server operating state is information such as the installed server model, accelerator model, number of accelerators, type of operating APL, operating accelerator (APL placement destination), and division block ID (number of divisions).

[0070] FIG. 15 is a diagram illustrating the classification definition of the server situation. As shown in FIG. 15, the server situation corresponds to each range of the server intake port temperature and the server operating state set in the actual operation phase. In the server internal configuration classification (server situation) with classification ID "No. 1," the server intake port temperature range is "19-21°C," and the server operating state at that time (installed server model, accelerator model, number of accelerators, operating APL type, operating accelerator (APL placement destination), and division block ID (number of divisions), etc.) is defined. Although not shown in the figure, the server operating state may also include information such as the APL operating frequency as a parameter.

[0071] Here, among the parameters of the server operating state, the operating accelerator (APL placement destination) and the division block ID (number of divisions) store the ID of the divided APL as the division block ID along with the division number, and indicate which accelerator (in FIG. 15 , FPGA1, FPGA2, etc.) the divided (three-part) APL (Particle-1, Particle-2, Particle-3) is placed on. Similarly, for the server internal placement configuration classification (Server Situation) with classification ID "No. 2," the server operating state is defined when the server intake port temperature range is "19-21°C," there are two accelerators 31, and the divided APL is placed on the two accelerators 31. In this way, different classification IDs are set and defined for each different setting for the parameters of the server intake port temperature and the server operating state.

[0072] In the learning phase, the server control learning data generator 340 operates the server 3 (accelerator-equipped server) for each server situation indicated by the classification defined by the server situation definition unit 330, and searches for the minimum server fan rotation control value (server minimum fan rotation value) that does not violate the SLA during the server temperature monitoring turn.The server control learning data generator 340 then determines the minimum server fan rotation control value that satisfies the SLA and the server power consumption at that time (for each server temperature monitoring turn) in association with the classification ID, and acquires them as server control learning data 301.

[0073] FIG. 16 is a diagram illustrating an example of the SLA information 302. As shown in FIG. 16, for example, the SLA information 302 set as board temperature thresholds for the accelerator model "Ariia10" includes a core temperature threshold of "80°C," a board temperature threshold of "64°C," and a supply temperature threshold of "96°C." Note that the board temperature thresholds set as the SLA information 302 may be arbitrarily set by multiplying the manufacturer's specified threshold temperature by a predetermined percentage (e.g., 80%). When multiple parameters are set in the SLA information 302, the server control learning data generation unit 340 determines that the SLA requirements are not met if any one of the parameters exceeds the threshold.

[0074] 17 is a diagram showing an example of searching for the minimum fan rotation speed (server minimum fan rotation value) for each server situation classification. As shown in FIG. 17, for example, in a server situation with classification ID No. 1, the minimum fan rotation speed is 30%, and the server power consumption at that time is "124W." Also, in a server situation with classification ID No. 2, the minimum fan rotation speed is 65%, and the server power consumption at that time is "148W."

[0075] As described above, the server control learning data generator 340 searches for the minimum fan speed for each server situation classification. Then, the server control learning data generator 340 associates the minimum server fan speed (server minimum fan speed value) that satisfies the SLA with the server situation and stores the server power consumption at that time in the storage unit as server control learning data 301.

[0076] The APL placement control unit 350 places the divided APLs in accordance with the APL placement configuration in the accelerator 31 ( FIG. 10 ) included in each server 3. When searching for the minimum fan rotation control value for each server situation classification in the learning phase, the APL placement control unit 350 cooperates with the server control learning data generation unit 340 to execute placement control of the divided APLs.

[0077] The server cooling control unit 360 controls the fan rotation control value of the fan 32 ( FIG. 10 ) provided in each server 3. When searching for the minimum fan rotation control number (server minimum fan rotation value) for each server situation classification in the learning phase, the server cooling control unit 360 cooperates with the server control learning data generation unit 340 to execute fan control for each server 3.

[0078] In this embodiment, the server cooling control unit 360 performs server cooling by rotating the server fan, i.e., by air cooling. However, in other embodiments, the server 3 may be cooled by water cooling. In such cases, the server cooling control unit 360 uses the cooling water temperature and / or the cooling water flow rate as control parameters. Specifically, the server cooling control unit 360 may, for example, maintain the cooling water temperature constant and perform control to search for a control value that minimizes the cooling water flow rate, or maintain the cooling water flow rate constant and perform control to search for the highest cooling water temperature. Alternatively, the server cooling control unit 360 may set a combination of the cooling water temperature and the cooling water flow rate as a parameter.

[0079] When the server cooling control unit 360 controls the cooling of the server 3 by water cooling, the server control learning data generation unit 340 searches for a cooling water temperature and / or a cooling water flow rate within a range that does not exceed the board temperature threshold indicated in the SLA information 302, instead of searching for the minimum server fan rotation control value within a range that does not exceed the board temperature threshold indicated in the SLA information 302. Then, the server control learning data generation unit 340 generates, in association with each server situation, the highest cooling water temperature and / or the lowest cooling water flow rate within a range that does not exceed the board temperature, and the server power consumption of the server 3 at that time, as server control learning data 301.

[0080] The optimal control calculation unit 240 (FIG. 11) also calculates the server power consumption for the entire room by calculating the highest cooling water temperature and / or the lowest cooling water flow rate and the server power consumption for each server 3 that does not exceed the board temperature according to the intake port temperature and server operating state, using the generated server control learning data 301. The optimal control calculation unit 240 then calculates the air conditioning control level and the total value of the air conditioning power consumption and the server power consumption in the air conditioning control turn for each pattern of the highest cooling water temperature and / or the lowest cooling water flow rate, and determines the control pattern that minimizes this total value.

[0081] During the operation phase, the failsafe execution unit 370 monitors the board temperature of the accelerator 31 of each server 3 via the server operation status monitoring unit 310 at each server temperature monitoring turn. If an SLA violation occurs, the failsafe execution unit 370 outputs instruction information to the server cooling control unit 360 to forcibly increase the server fan rotation speed of that server 3 (to increase it by a predetermined percentage). The server fan rotation speed upon an SLA violation is set in advance, for example, to increase by 10% from the current rotation speed or to 100%. Furthermore, if the SLA violation is not resolved even after a predetermined time has elapsed since increasing the fan rotation speed of the server 3 where the SLA violation occurred, the failsafe execution unit 370 outputs an alert to the failsafe control unit 250 of the integrated control unit 200 and instructs the failsafe control unit 250 to increase the air conditioning control of the air conditioner 4.

[0082] <Air Conditioning Control Unit> The air conditioning control unit 400 controls each air conditioner 4 installed in the DC room, and includes an air conditioning setting unit 410 and an air conditioning operation status measurement unit 420. The air conditioning setting unit 410 sets the air conditioning control for each air conditioner 4. For example, the air conditioning setting unit 410 controls the air conditioners 4 at set temperatures and air volumes according to each air conditioning control level, as shown in Fig. 12. Furthermore, the air conditioning operation status measurement unit 420 measures, for example, the air conditioning power consumption of all air conditioners 4 in the DC room during an air conditioning control turn.

[0083] [Optimal Solution Derivation Process] Next, a detailed description will be given of the optimal solution derivation process for a predetermined control pattern executed by the optimal control calculation unit 240 (FIG. 11). As shown in FIG. 11, this optimal control calculation unit 240 includes a control pattern power amount calculation unit 241 and an optimal control determination unit 242.

[0084] The optimal solution derivation process performed by the control pattern power consumption calculation unit 241 can be either <Process 1> where the air conditioning settings and the load placement amount in the server area are fixed, and the APL placement configuration in the server area and the server minimum fan rotation value are derived as the optimal solution, or <Process 2> where the air conditioning settings and the load placement amount in the server area are not fixed, and the APL placement configuration in the server area and the server minimum fan rotation value are derived as the optimal solution along with the air conditioning control and the load placement amount in the server area.

[0085] <Process 1> First, we will explain <Process 1>, in which the air conditioning settings and the load allocation amount to the server area are fixed. In <Process 1>, the control pattern power amount calculation unit 241 determines a placement configuration pattern that minimizes the total server power consumption in the server area from among placement configuration patterns that can be generated by changing the number of APL divisions that make up the load and the APL placement server in the server area. Then, the control pattern power amount calculation unit 241 determines the optimal solution as the APL division placement in the determined placement configuration pattern and the server minimum fan rotation value that satisfies the SLA of each server.

[0086] Specifically, the control pattern power consumption calculation unit 241 references air-conditioning control learning data 201 ( FIG. 11 ) for a case where the load placement pattern and air-conditioning control level for each server zone are fixed based on input information such as server load information, outside air temperature, and initial room temperature, and acquires information on the change in air intake port temperature for each pattern and information on the power consumption (air-conditioning power consumption) of the air conditioner 4 during the air-conditioning control turn. Then, using input information such as the server load for placement control zone 30 ( FIG. 10 ) within the room, the air intake port temperature at the start of the server temperature monitoring turn for that placement control zone 30, and the server operating status (installed server model, accelerator model, number of accelerators, type of active APL, active accelerator (APL placement destination), and division block ID (number of divisions)), the control pattern power consumption calculation unit 241 references server control learning data 301 and acquires information on the server minimum fan rotation value and server power consumption during the server temperature monitoring turn.

[0087] FIG. 18 is a diagram illustrating the process for determining the APL layout configuration within a server area performed by the control pattern power amount calculation unit 241. As shown in FIG. 18 , for example, when one "APLA" and one "APLB" are to be executed on a group of servers within a server area as loads, the server area inlet temperature at the start of the current server temperature monitoring cycle is "19°C-21°C," and the server models, accelerator models, and number of accelerators installed within the server area are as shown in FIG. 18 , the control pattern power amount calculation unit 241 references the server control learning data 301 to obtain information on the APL layout configuration within the server area, the minimum server fan rotation speed of each server at that time, and the server power consumption of each server. The control pattern power amount calculation unit 241 then calculates the total power consumption of each server in each APL layout configuration and determines the APL layout configuration within the server area with the smallest total power consumption. In FIG. 18 , the APL layout configuration pattern designated by symbol Z, in which the total server power consumption is "350 W," is determined.

[0088] The control pattern power calculation unit 241 can calculate the optimal APL placement configuration within the server area and information on the server power consumption for each temperature monitoring turn based on information on the change in the intake temperature in the corresponding air conditioning situation obtained by referring to the air conditioning control learning data 201.

[0089] If the optimal control calculation unit 240 receives a processing request during operation to change the amount or type of load to be placed in a server area, the optimal control calculation unit 240 reconstructs a new APL placement configuration. In this case, the optimal control calculation unit 240 may reset all APL placement configurations and reconstruct them again, or may update the additional APL placement configuration and the fan rotation control on the additional placement server for only the additional load without changing the existing APL placement configuration.

[0090] <Process 2> Next, we will explain <Process 2>, where the air conditioning settings and load placement amount in the server area are not fixed, and the APL placement configuration in the server area and the server minimum fan rotation value are derived as optimal solutions along with the air conditioning control and load placement amount in the server area.

[0091] The control pattern power amount calculation unit 241 references the air-conditioning control learning data 201 ( FIG. 11 ) when the load placement pattern and air-conditioning control level for each server zone are changed based on input information such as server load information, outside air temperature, and initial room temperature, and acquires information on the change in air intake port temperature for each pattern and information on the power consumption (air-conditioning power consumption) of the air conditioner 4 during the air-conditioning control turn. The control pattern power amount calculation unit 241 also acquires information on the server zone intake air temperature at the start of each server temperature monitoring turn from the information on the change in air intake port temperature for each pattern, and acquires information on the minimum server fan rotation value and server power consumption for each server during each server temperature monitoring turn based on the information in the server control learning data 301. The control pattern power amount calculation unit 241 calculates the total power consumption by adding up the total server power amount and the total air-conditioning power amount for the control pattern represented by the combination of air-conditioning control (air-conditioning control level), load placement to the server zone, server fan rotation control, and APL placement configuration.

[0092] 19 shows an example of a load allocation pattern for each server zone, where [zone 1, zone 2, zone 3, zone 4] is [50%, 50%, 0%, 0%], the APL allocation configuration pattern within the server zone is determined by the control pattern power consumption calculation unit 241, and the air conditioning control level is level 1 out of 3 levels. The control pattern power consumption calculation unit 241 acquires information on the change in air intake temperature and information on air conditioning power consumption by referencing the air conditioning control learning data 201. Furthermore, based on the information on the change in air intake temperature, the control pattern power consumption calculation unit 241 acquires information on the minimum fan rotation value of each server for each server temperature monitoring cycle by referencing the server control learning data 301, and calculates the total server power consumption by summing the power consumption of each server. Then, the control pattern power calculation unit 241 adds up the total server power consumption and total air conditioning power consumption in the load allocation pattern of the server area (load allocation to the server area), the air conditioning control level pattern, the minimum server fan rotation value for each server, and the APL allocation control pattern within each server area, and calculates the total power consumption.

[0093] The optimal control determination unit 242 determines the control pattern that results in the smallest power consumption among the total power consumptions of the air conditioning control, load allocation (load allocation pattern), fan control, and APL allocation configuration patterns calculated by <Process 2> as the optimal solution. If the total power consumption of "7.2 kW" shown in Figure 19 is the smallest total power consumption, the air conditioning control is set to "Level 1," the load allocation pattern is [50%, 50%, 0%, 0%] for each server area, and the fan control in the server temperature management turn of each server is, for example, "60% → 70%" for Server A (an example in which the server temperature monitoring turn is two turns), and the APL allocation configuration shown in Figure 19 is set.

[0094] By doing this, the power consumption reduction control device 100 can divide the APL and place it on the accelerator 31 in a DC room equipped with a server 3 equipped with an accelerator 31, and reduce the total power consumption consisting of the server power consumption and the air conditioning power consumption while satisfying the SLA.

[0095] When the optimal control calculation unit 240 receives a processing request during operation to change the load amount (server load information) of the entire room, the type of load, etc., it updates the load allocation to each room, the control solution for air conditioning control, etc. At that time, the optimal control calculation unit 240 may reset all APL placement configurations and reconfigure them again, or it may update the additional APL placement configuration and the fan rotation control on the additional placement server for only the added load without changing the existing APL placement configuration.

[0096] <Processing Flow> Next, the flow of processing executed by the power amount reduction control device 100 according to this embodiment will be described. Here, the processing for generating air-conditioning control learning data 201 and server control learning data 301, which is performed in the learning phase, will be described with reference to Fig. 20. Furthermore, the processing for deriving an optimal solution for a control pattern, which is performed in the operation phase, will be described with reference to Fig. 21.

[0097] 20 is a flowchart showing the flow of processing executed in the learning phase by the power amount reduction control device 100 according to this embodiment. In this learning phase, the power amount reduction control device 100 generates air conditioning control learning data 201 and server control learning data 301.

[0098] First, the integrated control unit 200 (room status acquisition unit 210) of the power consumption reduction control device 100 acquires information on server load information, outside air temperature, and initial room temperature (step S10). The room status acquisition unit 210 acquires server load information (such as the number of processed requests) from, for example, a system management device of the DC 10. Note that, in the learning phase, the room status acquisition unit 210 may set this server load information based on past data, or may acquire server load information within a predetermined range from the system management device. The room status acquisition unit 210 also acquires the outside air temperature and initial room temperature from temperature sensors installed inside and outside the room.

[0099] Next, the integrated control unit 200 (air-conditioning situation definition unit 220) defines an air-conditioning situation by combining the divided ranges for each piece of information: server load information, outside air temperature, initial room temperature, and air-conditioning control level (step S11). The air-conditioning situation definition unit 220 divides each of the server load information, outside air temperature, and initial room temperature into multiple ranges between the minimum and maximum values ​​according to the characteristics of the information. Furthermore, the air-conditioning control level is set in advance, as shown in FIG. 12. The air-conditioning situation definition unit 220 then defines an air-conditioning situation by combining the divided ranges.

[0100] Next, the integrated control unit 200 (air-conditioning control learning data generation unit 230) executes control of each server 3 and air conditioner 4 based on the information defined in each air-conditioning situation (server load information, outside temperature, initial room temperature, air-conditioning control level), thereby obtaining the total power consumption (air-conditioning power consumption) of the air conditioner 4 for the corresponding air-conditioning control turn and the change in the intake temperature of each server 3 (step S12).

[0101] Here, the air-conditioning control learning data generator 230 outputs an instruction to the air-conditioning control unit 400 to control the air conditioners 4 based on the air-conditioning control level. As a result, the air-conditioning control unit 400 (air-conditioning setting unit 410) instructs the air conditioners 4 in the DC room to set the temperature and airflow. The air-conditioning control unit 400 (air-conditioning operation status measurement unit 420) measures the air-conditioning power consumption during the air-conditioning control turn and outputs the measured air-conditioning power consumption to the integrated control unit 200 (air-conditioning control learning data generator 230). The integrated control unit 200 (air-conditioning control learning data generator 230) instructs the servers in the DC room to apply the load indicated by the server load information defined in the air-conditioning situation, for example, via a server management device (not shown). The air-conditioning control learning data generator 230 then acquires information on the air-inlet temperature (server air-inlet temperature) from a temperature sensor installed at the air-inlet of each server 3 during each server temperature monitoring turn.

[0102] Then, the air conditioning control learning data generation unit 230 generates the air conditioning power consumption for this air conditioning control turn and information on the change in the air intake temperature at each server 3 for each server temperature monitoring turn, corresponding to each air conditioning situation, as air conditioning control learning data 201, and stores it in the memory unit (step S13).

[0103] Next, the server control unit (APL division placement unit 320) checks the minimum server FAN rotation value for the APL (FPGA APL) to be installed in the accelerator 31 so as not to violate the SLA, determines the upper limit of the number of divisions, and then generates a pattern of APL placement configuration within the server area (APL placement configuration pattern) that changes the number of divisions of the APL that constitutes the load and the placement destinations of the divided APLs (step S14).

[0104] Next, the server control unit 300 (server situation definition unit 330) defines a combination of each range of server intake port temperatures and the server operating state (installed server model, accelerator model, number of accelerators, operating APL types, APL placement configuration, etc.) as a server situation (FIG. 15) (step S15). Here, the APL placement configuration is information indicating the division of the APL (number of divisions) and the placement destinations of the divided APLs, and is indicated by the operating accelerators and division block IDs (see FIG. 15).

[0105] Next, the server control unit 300 (server control learning data generator 340) operates server 3 (accelerator-equipped server) in the classification indicated by the server situation, determines the minimum server fan rotation control value (server minimum fan rotation value) within a range that does not exceed the board temperature threshold set in the pre-set SLA information 302, and measures the server power consumption at that time (step S16). For each server situation, the server control learning data generator 340 monitors the board temperature of accelerator 31 using the server operation status monitor 310 at each server temperature monitoring turn, and repeatedly changes and sets the fan control value to determine the minimum server fan rotation control value within a range that does not exceed the board temperature threshold.

[0106] Then, the server control learning data generation unit 340 generates the minimum server fan rotation control value (server minimum fan rotation value) that does not violate the SLA during the server temperature monitoring turn and the server power consumption at that time, corresponding to each server situation, and stores this in the memory unit as server control learning data 301 (step S17), thereby completing the learning phase processing.

[0107] <<Operation Phase>> Figure 21 is a flowchart showing the flow of processing executed by the power consumption reduction control device 100 according to this embodiment in the operation phase. In this operation phase, the power consumption reduction control device 100 executes processing to derive an optimal solution for a control pattern represented by a combination of air conditioning control (air conditioning control level), load allocation to server areas, server fan rotation control, and APL allocation configuration. Furthermore, the power consumption reduction control device 100 performs fail-safe processing if the board temperature of the accelerator 31 violates the SLA during operation.

[0108] First, the integrated control unit 200 (room status acquisition unit 210) of the power amount reduction control device 100 acquires server load information and initial temperature information (outdoor air temperature, initial room temperature) (step S20).

[0109] Next, the optimal control calculation unit 240 (control pattern power calculation unit 241) of the power reduction control device 100 uses the input information of the server load in the room and initial temperature information (outside temperature, initial temperature in the room) to refer to the air conditioning control learning data 201 and obtains information on the air conditioning power consumption and server intake temperature trends for each server corresponding to each air conditioning control level (step S21).

[0110] Next, the control pattern power amount calculation unit 241 determines whether the air conditioning settings (air conditioning control level) and the load allocation amount for the server zones are fixed, that is, whether the air conditioning control level and the load amount for each server zone are set in advance (step S22). If the air conditioning settings and the load allocation amount for the server zones are fixed (step S22 → Yes), the process proceeds to step S23.

[0111] In step S23, the control pattern power consumption calculation unit 241 refers to the server control learning data 301 using the input information of the server load in the server area (placement control area 30 in Figure 10) in the room, the intake port temperature at the start of the server temperature monitoring turn in that server area, and the server operating status (installed server model, accelerator model, number of accelerators, operating APL type, operating accelerator (APL placement destination), and division block ID (number of divisions)), and obtains information on the server minimum fan rotation value and server power consumption for each APL placement configuration pattern during that server temperature monitoring turn.

[0112] Then, the control pattern power amount calculation unit 241 determines, as the optimal solution, the pattern that minimizes the total server power consumption in the server area for each server temperature monitoring turn in each APL placement configuration pattern (step S24).

[0113] Next, the optimal control determination unit 242 causes each server 3 to control the server minimum fan rotation value in the server temperature monitoring cycle in the determined APL placement configuration pattern (APL division number and APL placement) with the fixed air conditioning settings and load placement amount for the server area (step S25).Then, the process proceeds to step S29.

[0114] On the other hand, if the air conditioning settings and the load allocation amount to the server zone are not fixed in step S22 (step S22 → No), the process proceeds to step S26. In step S26, the control pattern power amount calculation unit 241 uses the information on the server zone inlet temperatures at the start of each server temperature monitoring turn acquired in step S21 to refer to the server control learning data 301 and acquires information on the server minimum fan rotation value and the server power consumption of each server for each server temperature monitoring turn. The control pattern power amount calculation unit 241 then calculates the server power consumption during the air conditioning control turn for each server 3, and further calculates the total server power consumption of all servers 3, thereby calculating the sum of the air conditioning power consumption and the server power consumption for each control pattern.

[0115] The optimal control calculation unit 240 (optimal control determination unit 242) determines the smallest total power consumption as the optimal solution from the total values ​​of the air conditioning power consumption and the server power consumption calculated for each control pattern (step S27).

[0116] Then, based on the determined control pattern, the optimal control determination unit 242 sets the air conditioning control level in the air conditioning control turn to the air conditioner 4, allocates the load to each server area, divides the APL via the APL allocation control unit 350 and allocates it to the accelerator 31, and then has each server 3 execute the server minimum fan rotation value in the server temperature monitoring turn (step S28).

[0117] During operation, the server operational status monitoring unit 310 of the power consumption reduction control device 100 monitors the board temperature of the accelerator 31. The server operational status monitoring unit 310 then determines whether the accelerator board temperature threshold has been exceeded, resulting in an SLA violation (step S29). If no SLA violation has occurred (step S29 → No), the server operational status monitoring unit 310 continues monitoring the board temperature. On the other hand, if it is determined that an SLA violation has occurred (step S29 → Yes), the server operational status monitoring unit 310 executes fail-safe processing (step S30).

[0118] In this failsafe processing, first, the failsafe execution unit 370 of the server control unit 300 outputs instruction information to the server cooling control unit 360 to forcibly increase the server fan rotation speed of that server 3. If the SLA violation is not resolved even after a predetermined time has passed since the failsafe execution unit 370 increased the fan rotation speed of the server 3 where the SLA violation occurred, the failsafe execution unit 370 outputs an alert to the failsafe control unit 250 of the integrated control unit 200. When the failsafe control unit 250 receives the alert, it controls the air conditioner 4 to forcibly increase the air conditioning control.

[0119] In this way, in a DC room equipped with a server 3 equipped with an accelerator 31, the power consumption reduction control device 100 can reduce the total power consumption consisting of the server power consumption and the air conditioning power consumption by dividing the APL and placing it on the accelerator 31 while satisfying the SLA.

[0120] In this embodiment, the SLA information 302 is the board temperature threshold of the accelerator 31. As a variation thereof, the SLA information 302 may include a CPU temperature threshold in addition to the board temperature threshold of the accelerator 31. When the CPU temperature threshold is included in the SLA information 302 in this manner, in the learning phase, the server control learning data generation unit 340 determines the minimum server fan rotation control value within a range that does not exceed the board temperature threshold and the CPU temperature threshold. Furthermore, in the operation phase, the server operational status monitoring unit 310 monitors the CPU temperature in addition to the accelerator board temperature of each server 3. Then, even if an SLA violation occurs regarding the CPU temperature threshold, the failsafe execution unit 370 executes failsafe processing and outputs instruction information to the server cooling control unit 360 to forcibly increase the server fan rotation speed of that server 3.

[0121] In this way, the power consumption reduction control device 100 can reduce the total power consumption consisting of air conditioning power consumption and server power consumption by controlling the air conditioning control value, server fan rotation control, and APL layout configuration according to the server load in the room, within a range that does not exceed the board temperature threshold and the CPU temperature threshold (within a range that does not violate the SLA).

[0122] <Hardware Configuration> The power amount reduction control device 100 according to this embodiment is realized by a computer 900 having a configuration such as that shown in Fig. 22. Fig. 22 is a hardware configuration diagram showing an example of the computer 900 that realizes the functions of the power amount reduction control device 100 according to this embodiment. The computer 900 has a CPU (Central Processing Unit) 901, a ROM (Read Only Memory) 902, a RAM 903, an HDD (Hard Disk Drive) 904, an input / output I / F (Interface) 905, a communication I / F 906, and a media I / F 907.

[0123] The CPU 901 operates based on programs stored in the ROM 902 or the HDD 904, and performs control by the control unit. The ROM 902 stores a boot program executed by the CPU 901 when the computer 900 is started up, programs related to the hardware of the computer 900, and the like.

[0124] The CPU 901 controls an input device 910 such as a mouse or keyboard, and an output device 911 such as a display or printer, via an input / output I / F 905. The CPU 901 acquires data from the input device 910 via the input / output I / F 905, and outputs generated data to the output device 911. Note that a GPU (Graphics Processing Unit) or the like may be used as a processor together with the CPU 901.

[0125] The HDD 904 stores programs executed by the CPU 901 and data used by the programs. The communication I / F 906 receives data from other devices via a communication network (e.g., NW (Network) 920) and outputs the data to the CPU 901, and also transmits data generated by the CPU 901 to other devices via the communication network.

[0126] The media I / F 907 reads a program or data stored in the recording medium 912 and outputs it to the CPU 901 via the RAM 903. The CPU 901 loads a program related to a target process from the recording medium 912 onto the RAM 903 via the media I / F 907, and executes the loaded program. The recording medium 912 is an optical recording medium such as a DVD (Digital Versatile Disc) or a PD (Phase Change Rewritable Disc), a magneto-optical recording medium such as an MO (Magneto Optical Disc), a magnetic recording medium, a semiconductor memory, or the like.

[0127] For example, when a computer 900 functions as the power amount reduction control device 100 of the present invention, a CPU 901 of the computer 900 executes a program loaded onto a RAM 903 to realize the functions of the power amount reduction control device 100. Furthermore, data stored in the RAM 903 is stored in an HDD 904. The CPU 901 reads and executes a program related to a target process from a recording medium 912. Alternatively, the CPU 901 may read a program related to a target process from another device via a communication network (NW 920).

[0128] <Effects> The effects of the power consumption reduction control device 100 and the like according to the present invention will be described below. The power consumption reduction control device according to the present invention is a power consumption reduction control device 100 that controls accelerator-equipped servers and air conditioners 4 in a room in a data center 10, and in the room, a plurality of servers 3 are arranged in each server area that indicates a concentrated area for processing a load, and one or more air conditioners 4 are also arranged. The power consumption reduction control device 100 sets the board temperature threshold of the accelerator 31 provided in the accelerator-equipped server in accordance with an SLA (Service Level Agreement), an air-conditioning situation definition unit 220 that defines an air-conditioning situation by combining each of the divided ranges for the server load information, the outside air temperature, and the initial room temperature, and each stage of the air-conditioning control level, and defines an air-conditioning situation as an air-conditioning situation; an air-conditioning control turn is a predetermined control time until the air-conditioning control value is updated, and by executing control of the accelerator-equipped servers and the air-conditioners in the room based on the classification indicated by each air-conditioning situation, the air-conditioning power consumption in the air-conditioning control turn and the time shorter than the air-conditioning control turn are calculated; an air conditioning control learning data generation unit (230) that acquires information on the change in air intake port temperature of each accelerator-equipped server for each server temperature monitoring turn indicating a predetermined time, and generates, as air conditioning control learning data, the air conditioning power consumption for the air conditioning control turn and the change in air intake port temperature of each accelerator-equipped server for each server temperature monitoring turn, in association with each air conditioning situation; a server operating status monitoring unit (310) that monitors the board temperature of an accelerator (31) provided in the accelerator-equipped server; an APL division placement unit (320) that generates an APL placement configuration pattern by changing the number of divisions of an APL (Application) to be mounted on the accelerator (31) and the accelerator (31) to which the APL is to be placed;a server control learning data generation unit that searches for the minimum server fan rotation control value within a range that does not exceed the board temperature threshold indicated in the SLA information by operating the accelerator-equipped server for each server temperature monitoring turn in each server situation, and generates the minimum server fan rotation control value and the server power consumption of the accelerator-equipped server at that time as server control learning data in association with each server situation; and a server control learning data generation unit that generates the minimum server fan rotation control value and the server power consumption of the accelerator-equipped server at that time as server control learning data in accordance with a predetermined air conditioning control level and a load indicated in the server load information for each server area in each server situation. and an optimal control calculation unit 240 that, in the operation stage when the load allocation amount for distributing and arranging the servers is fixed, calculates the intake port temperature transition and air conditioning power consumption for each air conditioning control level using air conditioning control learning data based on information about the server load in the room, the outside air temperature, and the initial temperature in the room, and calculates the minimum server fan rotation control value and server power consumption for each accelerator-equipped server according to the server operating state including the intake port temperature and APL arrangement configuration for each server temperature monitoring turn using the server control learning data, thereby calculating the total value of server power consumption within the server area, and determines the APL arrangement configuration of the pattern that minimizes the total value and the minimum server fan rotation control value for each accelerator-equipped server as the optimal solution.

[0129] In this way, the power reduction control device 100 can calculate, as an optimal solution, an APL placement configuration (number of APL divisions and placement locations) that minimizes the total amount of server power consumption in the room when the air conditioning control level and the load allocation amount to each server area are fixed for accelerator-equipped servers in the DC room, and the minimum server fan rotation control value that satisfies the SLA of the accelerator-equipped servers. Thus, in a DC room equipped with accelerator-equipped servers, the total amount of power consumption made up of server power consumption and air conditioning power consumption can be reduced.

[0130] Furthermore, in the power reduction control device, in the operational stage, when a predetermined air conditioning control level and load allocation amount to the server area are not fixed, the optimal control calculation unit 240 calculates the intake port temperature transition and air conditioning power consumption for each air conditioning control level using air conditioning control learning data 201 based on information on server load information in the room, outside air temperature, and initial temperature in the room, and refers to server control learning data 301 based on the intake port temperature at the start of each server temperature monitoring turn indicated by the intake port temperature transition, calculates and adds up the server power consumption and air conditioning power consumption for the entire room in the control pattern indicated by the combination of the air conditioning control level, load allocation amount to the server area, minimum server FAN rotation control value of each accelerator-equipped server, and APL placement configuration, and determines the control pattern that results in the smallest total power consumption as the optimal solution.

[0131] In this way, when the air conditioning level and the amount of load placed on each server area are not fixed for an accelerator-equipped server installed in a DC room, the power reduction control device 100 can optimize the air conditioning control level, the amount of load placed on the server area, the minimum server FAN rotation control value that satisfies the SLA of each accelerator-equipped server, and the APL placement configuration (number of APL divisions and placement location), thereby reducing the total power consumption consisting of server power consumption and air conditioning power consumption.

[0132] The present invention is not limited to the above-described embodiments, and many modifications can be made by a person having ordinary skill in the art within the technical concept of the present invention.

[0133] REFERENCE SIGNS LIST 1 Power consumption reduction control system 3 Server (accelerator-equipped server) 4 Air conditioner 10 Data center (DC) 30 Placement control area 31 Accelerator 32 FAN 40 Air conditioning control area 100 Power consumption reduction control device 200 Integrated control unit 201 Air conditioning control learning data 210 Room status acquisition unit 220 Air conditioning situation definition unit 230 Air conditioning control learning data generation unit 240 Optimal control calculation unit 241 Control pattern power consumption calculation unit 242 Optimal control determination unit 250 Fail-safe control unit 300 Server control unit 301 Server control learning data 302 SLA information 310 Server operation status monitoring unit 320 APL division placement unit 330 Server situation definition unit 340 Server control learning data generation unit 350 APL placement control unit 360 Server cooling control unit 370 Fail-safe execution unit 400 Air conditioning control unit 410 Air conditioning setting unit 420 Air conditioning operation status measurement unit

Claims

1. A power reduction control device that controls accelerator-equipped servers and air conditioners in a room of a data center, wherein the room has a plurality of servers arranged in each server area indicating a group of areas for processing loads, and one or more air conditioners arranged therein, the power reduction control device comprising: a storage unit that stores an accelerator board temperature threshold value of the accelerator-equipped server as SLA (Service Level Agreement) information; an air conditioning situation definition unit that divides server load information indicating the processing load on the entire accelerator-equipped servers in the room, an outside air temperature, and an initial room temperature into a predetermined range width, divides the air conditioning control value of the air conditioner into predetermined air conditioning control levels, and defines an air conditioning situation by combining each divided range of the server load information, the outside air temperature, and the initial room temperature with each stage of the air conditioning control level; an air-conditioning control learning data generation unit that defines a predetermined control time until the air-conditioning control value is updated as an air-conditioning control turn, and executes control of the accelerator-equipped server and the air conditioner in the room based on the classification indicated by each of the air-conditioning situations, thereby acquiring information on the air-conditioning power consumption in the air-conditioning control turn and the change in the air-inlet temperature of each of the accelerator-equipped servers for each server temperature monitoring turn indicating a predetermined time shorter than the air-conditioning control turn, and generates the air-conditioning power consumption in the air-conditioning control turn and the change in the air-inlet temperature of each of the accelerator-equipped servers for each server temperature monitoring turn as air-conditioning control learning data in association with each air-conditioning situation; a server operation status monitoring unit that monitors the board temperature of the accelerator provided in the accelerator-equipped server; an APL division arrangement unit that generates an APL arrangement configuration pattern in which the number of divisions of an APL (Application) mounted on the accelerator and the accelerator to which the APL is to be arranged are changed;a server situation definition unit that divides the intake port temperature of the accelerator-equipped server into a predetermined range width, and defines each divided range of the intake port temperature as a server situation by combining a predetermined operating state including the APL arrangement configuration of the accelerator-equipped server that is set in advance; a server control learning data generation unit that searches for a minimum server fan rotation control value that does not exceed a board temperature threshold value indicated in the SLA information by operating the accelerator-equipped server for each server temperature monitoring turn in each of the server situations, and generates, as server control learning data, the minimum server fan rotation control value and the server power consumption of the accelerator-equipped server at that time for each server temperature monitoring turn in association with each server situation; and an optimal control calculation unit that, in an operation stage when the predetermined air-conditioning control level and a load allocation amount for distributing and allocating the load indicated by the server load information to each of the server zones are fixed, calculates the intake port temperature transition and the air-conditioning power consumption for each air-conditioning control level using the air-conditioning control learning data based on information on the server load information in the room, an outside air temperature, and an initial room temperature, and calculates the minimum server FAN rotation control value and the server power consumption of each of the accelerator-equipped servers according to the server operating state including the intake port temperature and the APL arrangement configuration for each server temperature monitoring turn using the server control learning data, thereby calculating a total value of the server power consumption within the server zone, and determines the APL arrangement configuration and the minimum server FAN rotation control value of each of the accelerator-equipped servers for a pattern that minimizes the total value as an optimal solution.

2. The power reduction control device according to claim 1, characterized in that during the operation phase, when the specified air-conditioning control level and the load allocation amount to the server area are not fixed, the optimal control calculation unit calculates the intake port temperature trend and the air conditioning power consumption for each air-conditioning control level using the air-conditioning control learning data based on the server load information in the room, the outside air temperature, and information on the initial temperature in the room, and refers to the server control learning data based on the intake port temperature at the start of each server temperature monitoring turn indicated by the intake port temperature trend, calculates and adds up the server power consumption and air conditioning power consumption for the entire room in a control pattern indicated by a combination of the air-conditioning control level, the load allocation amount to the server area, the minimum server FAN rotation control value of each accelerator-equipped server, and APL arrangement configuration, and determines the control pattern that results in the smallest total power consumption as the optimal solution.

3. A power reduction control method of a power reduction control device that controls accelerator-equipped servers and air conditioners in a room of a data center, wherein the room has a plurality of servers arranged in each server area indicating a consolidated area for processing a load, and one or more air conditioners arranged therein, the power reduction control device has a storage unit that stores a board temperature threshold value of an accelerator equipped in the accelerator-equipped server as SLA information, and an air-conditioning situation definition step of dividing server load information indicating the processing load on the entire accelerator-equipped servers in the room, an outside air temperature, and an initial room temperature into predetermined range widths, dividing the air-conditioning control value of the air conditioner into predetermined air-conditioning control levels, and combining each divided range of the server load information, the outside air temperature, and the initial room temperature with each stage of the air-conditioning control level to define an air-conditioning situation; an air-conditioning control learning data generating step of: defining a predetermined control time until the air-conditioning control value is updated as an air-conditioning control turn, and executing control of the accelerator-equipped server and the air conditioner in the room based on the classification indicated by each of the air-conditioning situations, acquiring information on the air-conditioning power consumption in the air-conditioning control turn and the change in the air-inlet temperature of each of the accelerator-equipped servers for each server temperature monitoring turn indicating a predetermined time shorter than the air-conditioning control turn, and generating the air-conditioning power consumption in the air-conditioning control turn and the change in the air-inlet temperature of each of the accelerator-equipped servers for each server temperature monitoring turn as air-conditioning control learning data in association with each air-conditioning situation; a server operation status monitoring step of monitoring the board temperature of the accelerator provided in the accelerator-equipped server; an APL division arrangement step of generating an APL arrangement configuration pattern in which the number of divisions of the APL mounted on the accelerator and the accelerator to which the APL is to be arranged are changed; a server situation definition step of dividing an intake port temperature of the accelerator-equipped server into a predetermined range width, and combining each range of the divided intake port temperature with a predetermined operating state including the APL arrangement configuration of the accelerator-equipped server that is set in advance, to define the server situation;a server control learning data generating step of searching for a minimum server FAN rotation control value within a range not exceeding a board temperature threshold value indicated in the SLA information by operating the accelerator-equipped server for each of the server temperature monitoring turns in each of the server situations, and generating, as server control learning data, the minimum server FAN rotation control value for each of the server temperature monitoring turns and the server power consumption of the accelerator-equipped server at that time in association with each of the server situations; and an optimal control calculation step of: calculating, in an operation stage when the predetermined air-conditioning control level and the load allocation amount for distributing and allocating the load indicated by the server load information to each of the server zones are fixed, the intake port temperature transition and the air-conditioning power consumption for each air-conditioning control level using the air-conditioning control learning data based on information on the server load information in the room, the outside air temperature, and the initial temperature in the room, and calculating, for each server temperature monitoring turn, the minimum server FAN rotation control value and the server power consumption of each of the accelerator-equipped servers according to the server operating state including the intake port temperature and the APL arrangement configuration using the server control learning data, thereby calculating a total value of the server power consumption within the server zone, and determining, as an optimal solution, the APL arrangement configuration and the minimum server FAN rotation control value of each of the accelerator-equipped servers for a pattern that minimizes the total value.

4. The power reduction control method according to claim 3, characterized in that when executing the optimal control calculation step in the operation stage, if the specified air-conditioning control level and the load allocation amount to the server area are not fixed, the air-conditioning control learning data is used to calculate the intake port temperature trend and the air-conditioning power consumption for each air-conditioning control level based on the server load information in the room, the outside air temperature, and information on the initial temperature in the room, and the server control learning data is used based on the intake port temperature at the start of each server temperature monitoring turn indicated by the intake port temperature trend to calculate and sum up the server power consumption and air-conditioning power consumption of the entire room in a control pattern indicated by a combination of the air-conditioning control level, the load allocation amount to the server area, the minimum server FAN rotation control value of each accelerator-equipped server, and APL arrangement configuration, and the control pattern which results in the smallest total power consumption is determined as the optimal solution.

Images