METHOD FOR OPTIMIZING THE POWER CONSUMPTION OF AN IT INFRASTRUCTURE AND CORRESPONDING SYSTEM

Description

TECHNICAL FIELD OF THE INVENTION

[0001] The technical field of the invention is that of optimizing the electrical consumption of an IT infrastructure. TECHNOLOGICAL BACKGROUND OF THE INVENTION

[0002] To perform tasks that require a computing infrastructure—for example, a supercomputer performing numerical simulations—that infrastructure must be powered by electricity. The more tasks the computing infrastructure performs at any given time, the greater the computing resources required to complete those tasks. This results in a greater amount of electricity being consumed. Electrical energy is supplied by an electrical grid, which significantly increases the demand for electricity from that distribution network.

[0003] Furthermore, during task execution, the use of computing resources also increases the heat dissipated by the IT infrastructure. To ensure optimal operation of this infrastructure, it is coupled with a thermal management system. Thermal management can also increase the demand for electrical energy from the power grid when the infrastructure temperature rises.

[0004] In a context of increasing numbers of IT infrastructures, there is therefore a need to limit the use of electrical energy from the electrical distribution network by limiting the impact on their performance.

[0005] Document WO 2016 / 105625 A1 describes a system and method for optimizing the power consumption of an IT infrastructure. SUMMARY OF THE INVENTION

[0006] The invention offers a solution to the problems mentioned above, by optimizing the electrical consumption of an infrastructure from an electrical network.

[0007] The invention relates to a method for optimizing the electrical energy consumption of a computer infrastructure comprising a plurality of computing resources and a first thermal regulation system, the computer infrastructure being powered by an electrical network and at least one local electrical energy source, the computer infrastructure having to perform a set of computing tasks by means of the plurality of computing resources, the method comprising the following steps: calculation of an estimate of electrical power deliverable by the electrical grid and an estimate of electrical power deliverable by the local electrical energy source over a time interval; estimation of electrical power consumed by the IT infrastructure over the time interval, based on the estimate of electrical power deliverable by the electrical grid over the time interval and the estimate of electrical power deliverable by the local electrical energy source over the time interval; based on the estimated electrical power consumed, determination of a subset of computing tasks from the set of computing tasks to be executed over the time interval; and estimation of a scheduling of the subset of computing tasks over the plurality of computing resources and a control of the plurality of computing resources and the first thermal regulation system over the time interval;based on the estimated scheduling and control, estimate of the electrical power required for the IT infrastructure over the time interval; based on the estimated electrical power required, determination of a parameterization of the electrical network and the local electrical energy source for the time interval; characterized in that the calculation of the estimate of electrical power deliverable by the electrical network is a function of an index K proportional to a quantity of CO2 emitted by the electrical network to deliver electrical energy over the time interval.

[0008] Thanks to the invention, the IT infrastructure is partially powered by at least one local electrical energy source, specific to the IT infrastructure. The use of electrical energy from the power grid is therefore minimized.

[0009] The configuration of the electrical network, the local electrical energy source and the first thermal regulation system makes it possible to adjust the electrical energy consumed to cover at least the electrical power required for the operation of the IT infrastructure, i.e. the energy needed to perform certain computing tasks and to regulate the temperature, while minimizing the share of electrical energy used from the electrical network.

[0010] Calculating the electrical power required allows the IT infrastructure to be maintained in optimal operating conditions, i.e. with optimal computing performance and temperature.

[0011] The electrical power consumed by the first control system is a function of the heat dissipated by the IT infrastructure performing the computational tasks. This first thermal control system dissipates this heat to maintain the IT infrastructure's temperature within the range required for optimal operation. Determining the subset of computational tasks to be performed over a given time interval allows us to determine the level of utilization of the IT infrastructure's computing resources and thus estimate the amount of heat to be dissipated over that time interval. Knowing the amount of heat to be dissipated allows us to estimate the initial electrical power required for control to maintain the IT infrastructure within an initial optimal temperature range.

[0012] Advantageously, the calculation of the estimated electrical power deliverable by the electrical network takes into account a demand for reduction of the electrical consumption of the IT infrastructure.

[0013] The request to reduce electricity consumption can be sent by the electricity energy supplier operating the electricity distribution network to force a decrease in the electricity consumption of the IT infrastructure.

[0014] Advantageously, the calculation of the estimated electrical power deliverable by the local electrical energy source takes into account at least one environmental factor.

[0015] The process thus takes advantage of environmental data such as temperature, sunshine or wind forecasts that can influence the production of electrical energy by the local electrical energy source and thermal regulation.

[0016] Advantageously, the local source of electrical energy is advantageously a renewable energy source.

[0017] Thus, it makes it possible to reduce the carbon footprint of the IT infrastructure by diluting the electrical energy from the electrical grid, potentially carbon-based, with electrical energy from the renewable energy source.

[0018] The process according to the invention is particularly relevant for reducing the carbon footprint of IT infrastructure. Indeed, in a context where part of the electricity production supplied by the grid may include a so-called carbon-based component, i.e., one that generates CO2, the carbon footprint of the infrastructure can be substantial.

[0019] Advantageously, the local electrical power source includes an electrical energy storage unit configured to store a quantity of energy, and the calculation of the estimate of the electrical power deliverable by the local power source is a function of the quantity of energy stored in the electrical energy storage unit.

[0020] In this way, the stored energy can be used to power the IT infrastructure, further reducing electricity consumption from the grid. The electrical energy storage unit also absorbs brief energy fluctuations from the grid and the local power source, thus relieving the burden of real-time monitoring of their production levels.

[0021] Advantageously, surplus production from the electrical grid and / or local electrical energy source, not consumed by the IT infrastructure over the time interval, is stored in the electrical energy storage unit.

[0022] Advantageously, the local electrical power source includes a second thermal regulation system coupled to the electrical energy storage unit, the parameter setting determination also including the parameter setting of the second thermal regulation system.

[0023] Advantageously, a surplus of production from the local electrical energy source, not consumed by the IT infrastructure over the time interval, is sent to the electrical grid.

[0024] Advantageously, each computational task in the set of computational tasks has a criticality, and the determination of the subset of computational tasks is a function of the criticality of each computational task.

[0025] Thus, when the available power consumption does not allow the execution of all computing tasks on the IT infrastructure, a criticality prioritization can be implemented to abandon or defer the execution of the least critical computing tasks.

[0026] Advantageously, each computational task in the set of computational tasks presents an estimate of the number of resources required, and the determination of the subset of computational tasks to be executed over the time interval is a function of the estimate of the number of resources required for each computational task.

[0027] Thus, when the available power consumption does not allow the execution of all tasks on the IT infrastructure, a prioritization by the number of computing resources required, and therefore by power consumption, can be implemented to abandon or defer the execution of the most resource-intensive computing tasks.

[0028] Advantageously, the process includes: the parameterization of the electrical network and the local electrical energy source according to the estimated parameterization; and the scheduling of the subset of computing tasks on the plurality of computing resources and the control of the plurality of computing resources and the first thermal regulation system over the time interval.

[0029] Advantageously, the IT infrastructure is a supercomputer.

[0030] The invention also relates to a system for optimizing the electrical consumption of a computer infrastructure comprising a plurality of computing resources and a first thermal regulation system, the computer infrastructure being powered by an electrical network and at least one local electrical energy source, the computer infrastructure having to perform a set of computing tasks by means of the plurality of computing resources, the system comprising: an optimization module configured to: based on an estimated electrical power consumption, determine a subset of computing tasks from the set of computing tasks to be executed over a time interval and estimate a scheduling of the subset of computing tasks on the plurality of computing resources and a control of the plurality of computing resources and the first thermal regulation system over a time interval; based on the estimated scheduling and control, estimate the electrical power consumed by the IT infrastructure over the time interval; an energy management device configured to: calculate an estimate of the electrical power deliverable by the electrical grid and an estimate of the electrical power deliverable by the local electrical energy source over the time interval;estimate the electrical power consumed by the IT infrastructure over the time interval, based on the estimated electrical power deliverable by the electrical network over the time interval and the estimated electrical power deliverable by the local electrical energy source over the time interval; and based on the estimated electrical power required, estimate a parameterization of the electrical network and the local electrical energy source over the time interval; characterized in that the calculation of the estimate of electrical power deliverable by the electrical network is a function of an index K proportional to a quantity of CO2 emitted by the electrical network to deliver electrical energy over the time interval.

[0031] The invention also relates to a computer program product comprising instructions which, when the program is executed on a computer, cause the computer to implement the steps of the process according to the invention.

[0032] Finally, the invention relates to a computer-readable recording medium comprising instructions which, when executed by a computer, lead the computer to implement the steps of the process according to the invention.

[0033] The invention and its various applications will be better understood by reading the following description and examining the accompanying figures. BRIEF DESCRIPTION OF THE FIGURES

[0034] The figures are shown for illustrative purposes only and are not intended to limit the invention. Unless otherwise specified, the same element appearing in different figures has a unique reference numeral. [ FIG.1[ ] schematically represents one embodiment of an optimization system according to the invention. ] FIG.2 ] schematically represents an example of the implementation of an optimization process according to the invention. DETAILED DESCRIPTION

[0035] The invention will be described with reference to the [ [FIG.1] and [FIG.2] The invention relates to a method 100 for optimizing the electrical consumption of a computer infrastructure CU required to perform a set W of computing tasks.

[0036] CU IT infrastructure is, for example, a supercomputer or a data center.

[0037] The CU computing infrastructure comprises a plurality of computing resources (RES). By computing resources (RES), we mean, for example, a plurality of processors and memories, for example interconnected in a cluster, and configured to perform computing tasks, such as numerical simulations.

[0038] The CU IT infrastructure also includes a first thermal control system, TH1. This first thermal control system is advantageously configured to maintain the CU IT infrastructure within an optimal temperature range, in which its operation is optimal. Indeed, the use of computing resources (RES) generates heat. To regulate the temperature of the CU IT infrastructure, the first control system, TH1, consumes a first electrical power (PTH1). The first thermal control system, TH1, can also be configured to regulate the temperature of a building housing the CU IT infrastructure.

[0039] The CU IT infrastructure is powered by electrical energy from an electrical grid and by at least one local renewable energy source.

[0040] By electrical network GRID, we mean an electrical energy distribution network, for example regional, managed by an electrical energy supplier.

[0041] The electricity grid is advantageously configured to communicate a PDGRID estimate of the electricity production it is able to deliver over a given time interval. This PDGRID estimate can, for example, be provided by the electricity supplier itself, or by a third-party electricity management service commonly known as an aggregator.

[0042] A local renewable energy (RE) power source is a separate source of electricity from the grid and specific to the local IT infrastructure (LUI). It can be a renewable energy source utilizing wind, solar, geothermal, tidal, or dry combustion energy such as wood. Ideally, a local RE power source is carbon-free, meaning it does not emit greenhouse gases, particularly CO2, to produce electricity. Examples of such sources include a wind turbine, photovoltaic panels, or a geothermal heat pump.

[0043] A local renewable energy source (RES) advantageously comprises multiple means of electricity generation, considered as a whole and centrally controlled, for example, by a programmable logic controller (PLC). Thus, an electricity generation target sent to the RES is distributed among the different means of energy generation. For example, the RES may include an electrical energy storage unit (STR) configured to store a quantity of energy, such as a battery.

[0044] The local renewable energy source may also include a combined heat and power (CHP) unit.

[0045] The method 100 according to the invention can also implement a PRED prediction service for WND environmental data. WND environmental data includes, for example, weather forecasts.

[0046] The optimization method 100 according to the invention is implemented by an optimization system SYS according to the invention. The SYS system includes, in particular: an OPT optimization module; and an EMS energy management device.

[0047] The OPT optimization module and the EMS energy management system can be separate PLCs or computing components running on the same computer. The OPT optimization module and the EMS energy management system are connected, for example, via a telecommunications network, to exchange information. The OPT optimization module and / or the EMS energy management system can also be advantageously connected to the WND environmental data prediction service (PRED).

[0048] The EMS energy management device is also connected to the electrical grid and the local renewable energy source so that it can send them REP settings.

[0049] The process 100 includes a first calculation step 110 of an estimate of electrical power deliverable PDGRID by the electrical network GRID and an estimate of electrical power deliverable PDENR by the local electrical energy source ENR over a time interval t0.

[0050] According to the example of implementation of the [ FIG.2The energy management system (EMS) initially receives information from the electrical grid (GRID). This information may include a time interval t0 and the electrical power that the grid is capable of supplying during that time interval. It may also include a request to reduce the electrical consumption of the IT infrastructure (CU) during that time interval t0. This reception may follow a request sent by the EMS to the electrical grid.

[0051] The energy management system (EMS) then calculates an estimate of the power output available (PDGRID) from the grid over the time interval t0. This calculation advantageously incorporates information received from the grid, such as requests to reduce electricity consumption. Alternatively, the grid directly sends an estimate of the power it can deliver (PDGRID) and the time interval t0 during which it can deliver it.

[0052] The reception of information from the electrical grid (GRID) can occur periodically. The period can range from one minute to several hours, for example, ten minutes. The time interval t0 advantageously corresponds to the period of information reception from the electrical grid (GRID).

[0053] The time interval t0 can also start each time information is received from the electrical grid (GRID), even if such information is not received periodically. Receiving this information advantageously re-triggers the steps of process 100 to incorporate a new estimate of the power output available from the electrical grid (GRID). Thus, when the electricity provider managing the distribution network is unable to absorb a decrease or surplus in electricity production during this time interval t0, it can send new information to the SYS so that the latter can optimize the power consumption of the CU IT infrastructure by taking into account the production variations.

[0054] The EMS energy management system also calculates the estimated electrical power output (PDENR) from the local renewable energy source (ENR) over the time interval t0. To do this, the EMS system advantageously takes into account the receipt of at least one environmental data point (WND) provided by the PRED prediction service. Each WND environmental data point includes, for example, the predicted sunshine or wind speed over the time interval t0 at the location of the local renewable energy source (ENR) and therefore the CU IT infrastructure.

[0055] When the local renewable energy source (RES) includes an electrical energy storage unit (STR), it is advantageous for the energy management system (EMS) to also receive an estimate of the amount of energy Q stored by said STR. In this way, the EMS can determine whether some of the stored electrical energy can be used to power the IT infrastructure (CU) or whether, on the contrary, it is preferable to replenish the electrical energy stock Q.

[0056] Once the estimates of the deliverable power PDGRID and PDENR from the electrical grid and the local renewable energy source (ENR) have been calculated, the energy management system (EMS) calculates an estimate of the electrical power consumption (PC) over the time interval t0 by the IT infrastructure (CU). The electrical power consumption (PC) is, for example, the maximum total power deliverable by the electrical grid and the local renewable energy source (ENR) that can supply the IT infrastructure (CU) during the execution of tasks W, taking into account the consumption of the first control system (TH1).

[0057] The estimate of the electrical power consumption PC over the time interval t0 and advantageously the time interval t0, are then sent 220 by the energy management device EMS to the optimization module OPT.

[0058] In the example of implementation of the [ FIG.2The OPT optimization module determines a subset W' of computational tasks to be executed over the time interval t0 from the set W of computational tasks. The set W of computational tasks advantageously includes computational tasks currently being executed by the IT infrastructure CU but not yet completed, as well as computational tasks recently received or whose execution is pending.

[0059] For example, when a computing task from the set W of computing tasks is added to the computing task subset, its execution is scheduled for the time interval t0. If the computing task is already running on the computing infrastructure CU, it continues its execution without interruption. A computing task waiting to be executed is then scheduled for the time interval t0.

[0060] When a computational task from set W is not added to the subset of computational tasks, its execution is not scheduled for the time interval t0. If the computational task is already running on the computing infrastructure CU, it is stopped or hibernated. A computational task awaiting execution is postponed to a later time interval or abandoned.

[0061] A stop indicates that the computing task is being executed by the CU computing infrastructure but must be stopped at the time interval t0.

[0062] A hibernation state indicates that the computing task is running but needs to be stopped, and that a backup of that computing task at the time of stopping is stored to serve as a starting point when it is restarted. A computing task that is simply stopped does not benefit from a backup, and its restart, if necessary, is equivalent to an initial execution of that task.

[0063] A computing task can potentially be executed in power-saving mode, meaning that the number of computing resources allocated to it is reduced or that the energy consumed by the allocated computing resources is limited. In the latter case, this might involve limiting the power used, known as "power capping." It could also involve reducing the operating frequency of these resources, such as the processor frequency.

[0064] The determination 230 of the subset of computational tasks is advantageously a function of an estimate of a gain G, associated with each action relating to the computational tasks of the set W of computational tasks.

[0065] For example, running in power-saving mode reduces the instantaneous electrical power consumed by the task in question. The gain G is therefore advantageously positive and preferentially opposed to the variation in the number of resources allocated to said task. It can be formulated as G = C 1 − C 2 ∝ N 1 − N 2 Or C 1 and C 2 are estimates of the average consumption of said computing task respectively outside energy-saving mode and in energy-saving mode and N 1 and N 2 are, for example, the number of computing resources allocated to said computing task respectively outside of energy saving mode and in energy saving mode.

[0066] Putting a computational task into hibernation can offer a gain G zero, allowing it to be considered without negative effect.

[0067] Stopping a computational task after a given time J , can exhibit a gain G proportional to: G = − t − J × C with t time and C the average consumption of said task.

[0068] It is advantageous to correct the gain Gby a future non-fulfillment risk factor R, allowing for consideration of contingencies that result in a subset W' of unfeasible computational tasks within the time interval t0. Indeed, the IT infrastructure CU may, in addition to power consumption constraints, be subject to usage constraints. It is therefore not feasible to systematically abandon or stop computational tasks of low or moderate criticality solely to satisfy power consumption constraints. Similarly, the execution of a computational task cannot be postponed indefinitely. To avoid these pitfalls, the risk R The risk of future non-fulfillment can be defined according to criteria set by users of the CU IT infrastructure. RThe risk of future non-occurrence can be defined at a given time and adapted during the operation of the CU IT infrastructure. For example, artificial intelligence could be used to estimate the risk. R.

[0069] A total gain GT associated with N Calculation tasks for the set W allow us to determine the impact of the risk R of future non-realization on the operation of the CU IT infrastructure and can be calculated as follows: G T = ∑ i N g a i − R a 1 , … , a n Or a 1, ..., year represents all the actions considered.

[0070] The determination 230 of the subset W' of the computing task is a function of the electrical power consumption PC estimated by the energy management device EMS.

[0071] To improve the determination of the subset W' of computing tasks, each computing task in set W advantageously provides an estimate of the number of computing resources (RES) required. For example, for each computing task, this might be the number of processors or compute nodes needed to perform that task. The estimate of the number of computing resources required is, for example, provided by a user when submitting the computing task to the digital infrastructure (CU). The estimate of the number of computing resources required can also be calculated by a specific automated system, for example, one based on artificial intelligence.

[0072] The determination 230 of the subset W' of computing tasks is then advantageously based on the estimation of the number of RES resources needed for each computing task in the set W of computing tasks.

[0073] The determination of the subset W' of computational tasks can also be based on the type of computational resources (RES) to be allocated to each computational task. Indeed, the execution time of certain computational tasks may be limited by the number of processor-type RES or memory-type RES resources. In the first case, we speak of CPU-bound in English and in the second case of memory-bound in English.

[0074] Determining the subset W' of computational tasks can also estimate an energy consumption for each computational task in set W. Estimating the energy consumption thus allows us to determine the total energy consumption required to carry out the subset W' of computational tasks.

[0075] The energy consumption estimate for each computing task is based, for example, on the number of computing resources (RES) required and the estimated computing time. This number of computing resources and the estimated time can be specified by a user when submitting a computing task to the computing infrastructure (CU).

[0076] The estimation of energy consumption for each computational task can also be based on measuring, after execution, the energy actually consumed by similar computational tasks. Furthermore, this estimation advantageously employs artificial intelligence to, firstly, classify solved tasks according to their energy consumption and, secondly, determine the energy consumption for each computational task within the set W of computational tasks, based on the classified solved tasks.

[0077] Each computational task in the set W of computational tasks can also have a criticality rating. Thus, the determination of the subset W' of computational tasks can take into account the criticality of each computational task in the set W. For example, a high criticality rating might indicate that a task should complete successfully before a given time. Therefore, a computational task with a high criticality rating is preferentially added to the subset W' of computational tasks to be executed over the time interval. A computational task with a medium criticality rating can, for example, be put into hibernation or stopped and restarted in less than a day. A computational task with a low criticality rating can, for example, be stopped, without any constraint on the duration before restarting said task, or even abandoned, that is to say, removed from the set W of computational tasks.

[0078] The W' subset of computing tasks is sent 240 to the CU computing infrastructure to be scheduled on the RES computing resources.

[0079] According to the example of implementation of the [ FIG.2 The OPT optimization module estimates an ORD schedule for the subset W' of computational tasks across the plurality of computing resources (RES). This involves assigning each computing resource (RES) a set of computational tasks ordered according to time. The ORD scheduling estimation advantageously takes into account the actions associated with each computational task, and in particular, those tasks that can be executed in energy-saving mode.

[0080] The OPT optimization module also estimates 250 DRV control of the plurality of RES computing resources. This involves, for example, determining the state of each RES computing resource within the plurality of RES computing resources of the CU IT infrastructure.

[0081] DRV management of RES computing resources may include at least one of the following actions: start a computing resource; hibernate a computing resource; stop a computing resource.

[0082] Thus, each RES computing resource not allocated to the execution of computing tasks in the subset W' of computing tasks can be hibernated or stopped, thereby reducing the power consumed by the CU computing infrastructure.

[0083] The DRV pilot's 250 estimate may also be a function of a gain G associated with each DRV control action. For example, stopping a computing resource can be defined as: G = C actif × d − P Or Cactive is the consumption of the active or started RES computing resource, and d is the estimated duration of the expected downtime. P is a penalty, for example a positive constant, accounting for the restart time of said RES computing resource and an associated probability of non-restart.

[0084] The gain from putting a computing resource into hibernation can be defined as: G = C actif − C veille × d − P Or C standby is the consumption of the RES computing resource in standby mode.

[0085] The winning GThe DRV control estimation of 250 can also be corrected using a risk-based approach. Indeed, the DRV control estimation may involve very frequent, even aggressive, starting and stopping of RES computing resources to reduce associated energy consumption. However, there is a probability that the computing resources will not restart correctly. Correcting the gain using a risk-based approach thus reduces the number of start / stop cycles and therefore preserves the RES computing resources.

[0086] The process also includes estimation 250 of the DRV control of the first TH1 thermal regulation system. Indeed, the electrical power consumed by each computing resource (RES) is partly dissipated as heat. The first TH1 regulation system thus makes it possible to regulate the temperature of the IT infrastructure (CU), and more specifically the temperature of each computing resource (RES), so that it is maintained within the first optimal temperature range.

[0087] Estimate 250 of the DRV control of the first TH1 thermal regulation system, for example, takes into account the energy dissipation of each computing resource RES depending on its state, such as active, standby or hibernation, and shut down. The DRV control of the first TH1 thermal regulation system thus makes it possible to determine the total heat to be dissipated and therefore an initial electrical power requirement for regulation PTH1.

[0088] Furthermore, the DRV control estimation 250 can advantageously incorporate at least one environmental WND data point provided by the PRED prediction service. For example, if the first control system TH1 is a heat exchanger transferring heat from the outside air to the IT infrastructure CU, then the environmental data considered advantageously includes the outside air temperature and humidity. Thus, if the outside air is hot and dry, the first electrical power consumption of the PTH1 control system will be greater than when the outside air is cool and humid.

[0089] Process 100 also includes, based on the previously estimated ORD scheduling and DRV control, an estimate 270 of the minimum electrical power required (PMIN) to ensure the operation of the computing infrastructure (CU) over the time interval t0 and to enable the execution of the subset W' of computing tasks. The minimum electrical power required (PMIN) specifically ensures the power supply to each computing resource (RES) and the first control system (TH1) over the time interval t0.

[0090] The estimate 270 of said required power PMIN is advantageously a function of the subset W' of computing tasks, and of the previously estimated scheduling ORD and DRV control. Said required power PMIN corresponds, for example, to the sum of the estimated power for each computing resource RES as a function of its state over the time interval t0 and the first thermal regulation power PHT1.

[0091] The required electrical power PMIN is then communicated to the EMS energy management device in order to determine a REP parameter setting. The REP parameter setting corresponds, for example, to: the electrical power to be delivered by the electrical network GRID over the time interval t0; and the electrical power to be delivered by the local electrical energy source ENR over the time interval t0.

[0092] This includes, for example, a production instruction sent to the grid and the local renewable energy source. The PWR configuration advantageously includes a management instruction sent to the energy storage unit (ESU) when the local renewable energy source has one. This instruction might be, for example, to release stored energy, either from the local renewable energy source and the grid, or conversely, to store a portion of the energy produced by the local renewable energy source and / or the grid.

[0093] The REP parameter setting is determined based on the previously estimated required electrical power PMIN. According to the implementation example of the [ FIG.2[ ], the determination 290 of the REP parameterization is carried out by the EMS energy management device. It advantageously takes into account the consumable power PC and / or the required power PMIN over the time interval t0. It can also take into account the amount of energy stored in the electrical energy storage unit STR of the local renewable energy source ENR.

[0094] When the estimated PWR parameterization allows the estimated PMIN electrical power requirement to be provided, the PWR parameterization is advantageously sent to the electrical grid and to the local renewable energy source.

[0095] However, process 100 advantageously includes a control step 310 to determine whether the estimated REP parameterization provides the estimated minimum electrical power requirement (PMIN). If not, process 100 advantageously includes updating the estimated electrical power consumption (PC) over the time interval t0. This update includes, for example, a repetition of: the first calculation step 110 of the estimates of the deliverable electrical power PDGRID, PDENR by the electrical network GRID and the local electrical energy source ENR over a time interval t0; the calculation 210 of the estimate of the consumable electrical power PC over the time interval t0; and the determination 290 of the REP parameterization; as described previously. Indeed, energy production conditions may have changed since the last execution of the first calculation step 110, and a new determination 290 of the PWR parameterization may allow the control step 310 to be satisfied.

[0096] When the controlled step 310 is not satisfied with the updated estimates, the process 100 advantageously includes a triggering of steps 230, 250, 270 allowing an update of the required electrical power PMIN to be obtained.

[0097] Obtaining compatible REP parameters and a minimum electrical power requirement (PMIN) can present a combinatorial optimization problem. It is then advantageous to solve this combinatorial optimization problem using, for example, a heuristic method, such as a trial-and-error approach based on random propositions.

[0098] For example, a plurality of subsets W' of computing tasks can be determined, each subset W' being obtained randomly. An estimate of the required electrical power PMIN is advantageously estimated from each subset W' of computing tasks. A scheduling ORD and a control DRV are advantageously determined for each subset W'. A comparison of each required electrical power PMIN with a previously estimated REP parameterization allows at least one subset W' to be determined that satisfies said REP parameterization.

[0099] It is also possible to determine a plurality of REP parameterizations, for example, randomly. A comparison of each REP parameterization with respect to each required electrical power PMIN thus makes it possible to determine at least one subset W' satisfying at least one REP parameterization.

[0100] It can also be advantageous to solve the combinatorial optimization problem using artificial intelligence, for example by implementing reinforcement learning.

[0101] When the REP parameter allows for the provision of the estimated electrical power required (PMIN), the electrical grid (GRID) and the local renewable energy source (ENR) receive the REP parameter. They then each deliver electrical power according to the previously determined REP parameter, thus supplying the IT infrastructure (CU).

[0102] The estimated ORD scheduling is sent to the CU IT infrastructure to schedule the computing tasks of the W' subset of computing tasks.

[0103] The estimated DRV control is also sent 260 to the CU IT infrastructure in order to implement 360, 370 the estimated DRV control on the plurality of RES computing resources and the first TH1 control system.

[0104] The PWR settings determined by the energy management device may include the storage of surplus energy from the grid and / or the local renewable energy source (RES) by the STR electrical energy storage unit. The STR electrical energy storage unit may also be configured to store surplus electrical energy from the local renewable energy source (RES).

[0105] Production surplus, or simply surplus, refers to the difference between the instantaneous power produced by the grid and / or the local electrical energy source and the instantaneous power consumed by the IT infrastructure. When the surplus is to be stored by the electrical energy storage unit, it advantageously excludes the energy originating from said electrical energy storage unit.

[0106] The local renewable energy (RE) power source can also include a second thermal control system coupled to the electrical energy storage unit (STR). This second thermal control system allows, for example, maintaining the temperature of the STR storage unit within a second optimal temperature range over the time interval t0. For a battery, the optimal temperature range extends, for example, from 10°C to 35°C. The second thermal control system therefore consumes a second electrical control power to maintain the STR storage unit's temperature within a second optimal temperature range, thus optimizing the STR storage unit's operation. Estimate 290 of the REP parameterization advantageously incorporates a second control power PTH2 to maintain the STR electrical energy storage unit within a second optimal temperature range.

[0107] According to the invention, the calculation 110 of the estimated electrical power deliverable by the electrical grid is a function of an index K. This index K relates to the electrical energy produced by the electrical grid. The index K is, for example, proportional to the amount of CO2 emitted by the electrical grid to deliver electrical energy over a given time interval. It can be expressed in grams of CO2 / kWh. This can then be referred to as the carbon footprint of electricity production. This index K can also be proportional to the difficulty of supplying the electrical energy delivered by the electrical grid; for example, it corresponds to an electricity price indicated in a monetary unit. The index K is obtained, for example, when the energy management system (EMS) receives information 120 from the electrical grid.

[0108] A change in the K index, or a K index value reaching a predetermined threshold, can trigger a new transmission of information from the electricity grid (GRID) to the energy management system (EMS). This new transmission might include, for example, a new request to reduce electricity consumption over the time interval t0. In this way, the estimated power consumption (PC) by the IT infrastructure (CU) can be adjusted based on changes in the K index. For example, PC power consumption can be reduced when the carbon footprint of electricity production is high. Conversely, PC power consumption can be increased, for example, by prioritizing energy-intensive tasks or storing energy, when the carbon footprint is low. Following the same logic, PC power consumption can be adjusted based on the electricity price.

[0109] Calculation 110 of the PDGRID deliverable electrical power estimate by the electrical grid can advantageously take into account the amount of CO2 that could be produced over the time interval t0 to power the CU computing infrastructure. This makes it possible to limit the amount of CO2 emitted to perform computing tasks, thereby reducing the carbon footprint of the CU computing infrastructure.

[0110] It also becomes possible to run energy-intensive computing tasks when the K index is low. Thus, energy-intensive tasks have a controlled carbon footprint. The same reasoning can be applied to the price of electricity, allowing energy-intensive computing tasks to be run when the price is low.

[0111] The index K can be taken into account in determining the subset W' of computational tasks to limit the number of computational tasks to be executed over the time interval t0, for example when the amount of CO2 associated with their execution is too high, i.e. above a threshold.

[0112] The K index can also be taken into account in determining the PWR parameterization of the electrical grid and the local renewable energy source. For example, a PWR parameterization determined for a high K index may preferentially use some of the energy stored in the electrical energy storage unit (STR). Conversely, a PWR parameterization determined for a low K index may preferentially store some of the energy produced by the electrical grid in the electrical energy storage unit (STR).

[0113] To return to a previously discussed example, a variation in the K index, or a K index value reaching a predetermined threshold, can trigger a PWR configuration such that the electrical power produced from the grid and / or the local renewable energy source exceeds the minimum required electrical power (MRP). In this way, a surplus of electricity production is generated and can be stored for later use. Electrical energy can thus be stored when it is inexpensive or when its carbon footprint is low.

[0114] The determination of the subset of computational tasks 230 is advantageously a function of an estimate of a gain G , associated with each action related to the computational tasks added or not to the subset W' of tasks to be executed over the time interval t0. The gain Gcan also be proportional to the index K. For example, launching an energy-intensive computing task when the index K is below a threshold index Ks corresponds to a gain G : G = C max × d × K S − K Or d is the estimate of the maximum consumption time C max,

Claims

1. Method (100) for optimizing the electrical energy consumption of a computer infrastructure (CU) comprising a plurality of computing resources (RES) and a first temperature control system (TH1), the computer infrastructure (CU) being powered by an electrical grid (GRID) and at least one local electrical energy source (ENR), the computer infrastructure (CU) having to execute a set (W) of computing tasks using the plurality of computing resources (RES), the method (100) comprising the following steps: - calculating (110) an estimate of electrical power (PDGRID) that can be supplied by the electrical grid (GRID) and an estimate of electrical power (PDENR) that can be supplied by the local electrical energy source (ENR) over a time interval (t0); - estimating (210) an electrical power (PC) that can be consumed by the computer infrastructure (CU) over the time interval (t0), on the basis of the estimate of electrical power (PDGRID) that can be supplied by the electrical grid (GRID) over the time interval (t0) and of the estimate of the electrical power (PDENR) that can be supplied by the local electrical energy source (ENR) over the time interval (t0); - on the basis of the estimated consumable electrical power (PC), determining (230) a subset (W') of computing tasks from the set (W) of computing tasks to be executed over the time interval (t0), and estimating (250) a scheduling (ORD) of the subset (W') of computing tasks on the plurality of computing resources (RES) and a control (DRV) of the plurality of computing resources (RES) and the first temperature control system (TH1) over the time interval (t0); - on the basis of the estimated scheduling (ORD) and control (DRV), estimating (270) an electrical power (PMIN) required by the computer infrastructure (CU) over the time interval (t0); - on the basis of the estimated required electrical power (PMIN), determining (290) a parameterization (REP) of the electrical grid (GRID) and of the local electrical energy source (ENR) for the time interval, characterized in that the calculation of the estimate of electrical power (PDGRID) that can be supplied by the electrical grid (GRID) is a function of an index K proportional to a quantity of CO2 emitted by the electrical grid (GRID) to supply electrical energy over the time interval (t0).

2. Method (100) according to the preceding claim, wherein the calculation (110) of the estimate of electrical power (PDGRID) that can be supplied by the electrical grid (GRID) takes into account a request to reduce the electrical consumption of the computer infrastructure (CU).

3. Method (100) according to one of the preceding claims, wherein the calculation (110) of the estimate of electrical power (PDENR) that can be supplied by the local electrical energy source (ENR) takes into account at least one piece of environmental data (WND).

4. Method (100) according to one of the preceding claims, wherein the local electrical energy source (ENR) is a renewable energy source.

5. Method (100) according to one of the preceding claims, wherein the local electrical energy source (ENR) comprises an electrical energy storage unit (STR) which is configured to store a quantity of energy, and wherein the calculation (110) of the estimate of the electrical power (PDENR) that can be supplied by the local energy source (ENR) is a function of the quantity (Q) of energy stored in the electrical energy storage unit (STR).

6. Method (100) according to the preceding claim, wherein excess production from the electrical grid (GRID) and / or from the local electrical energy source (ENR), not consumed by the computer infrastructure (CU) over the time interval (t0), is stored in the electrical energy storage unit (STR).

7. Method (100) according to one of the two preceding claims, wherein the local electrical energy source (ENR) comprises a second temperature control system (TH2) coupled to the electrical energy storage unit (STR), determining (290) the parameterization (REP) also comprises parameterizing the second temperature control system (PTH2).

8. Method (100) according to one of the preceding claims, wherein excess production from the local electrical energy source (ENR), not consumed by the computer infrastructure (CU) over the time interval (tG), is sent to the electrical grid (GRID).

9. Method (100) according to one of the preceding claims, wherein each computing task of the set (W) of computing tasks has a criticality and wherein determining (130) the subset (W') of computing tasks is a function of the criticality of each computing task.

10. Method (100) according to one of the preceding claims, wherein each computing task of the set (W) of computing tasks has an estimate of a required number of resources, and wherein determining (130) the subset (W') of computing tasks to be executed over the time interval (t0) is also a function of the estimate of the number of resources required for each computing task.

11. Method (100) according to one of the preceding claims, further comprising: - parameterizing (320) the electrical grid (GRID), the local electrical energy source (ENR) and the temperature control system (TH) according to the estimated parameterization; and - scheduling the subset (W') of computing tasks on the plurality of computing resources (RES) and controlling the plurality of computing resources (RES) and the first temperature control system over the time interval (t0).

12. Method (100) according to one of the preceding claims, characterized in that the computer infrastructure (CU) is a supercomputer.

13. System (SYS) for optimizing the power consumption of a computer infrastructure (CU) comprising a plurality of computing resources (RES) and a first temperature control system (TH1), the computer infrastructure (CU) being powered by an electrical grid (GRID) and a local electrical energy source (ENR), and having to execute a set of computing tasks (W) using the plurality of computing resources (RES), the system (SYS) comprising: - an optimization module (OPT), configured to: - on the basis of an estimated consumable electrical power (PC), determine (130) a subset (W') of computing tasks from the set (W) of computing tasks to be executed over a time interval and estimate (140) a scheduling of the subset (W') of computing tasks on the plurality of computing resources (RES) and a control of the plurality of computing resources (RES) and of the first temperature control system (TH1) over the time interval; - on the basis of the estimated scheduling and control, estimate (150) an electrical power (PMIN) consumed by the computer infrastructure (CU) over the time interval; - an energy management device (EMS) configured to: - calculate (110) an estimate of electrical power (PDGRID) that can be supplied by the electrical grid (GRID) and an estimate of electrical power (PDENR) that can be supplied by the local electrical energy source (ENR) over the time interval; - estimate (120) the electrical power (PC) that can be consumed by the computer infrastructure (CU) over the time interval, on the basis of the estimate of electrical power (PDGRID) that can be supplied by the electrical grid (GRID) over the time interval and the estimate of electrical power (PDENR) that can be supplied by the local electrical energy source (ENR) over the time interval; and - on the basis of the estimated required electrical power (PMIN), estimate (160) a parameterization (REP) of the electrical grid (GRID) and the local electrical energy source (ENR) over the time interval, characterized in that the calculation of the estimate of electrical power (PDGRID) by the electrical grid (GRID) is a function of an index K proportional to a quantity of CO2 emitted by the electrical grid (GRID) to supply electrical energy over the time interval (t0)14. Computer program product comprising instructions that, when the program is executed on a computer, cause the computer to implement the steps of the method (100) according to one of claims 1 to 12.

15. Computer-readable recording medium comprising instructions that, when executed by a computer, cause the computer to implement the steps of the method according to one of claims 1 to 12.