A server cabinet power distribution method

Abstract

This invention discloses a server rack power allocation method. The method includes the following steps: Pre-setting all racks to an "on" state, and evenly distributing the total power required for the overall computing tasks of the data center across all racks as the initial state for iteration; quickly estimating the temperature of each rack using a linear influence coefficient matrix, and using this initial hotspot temperature as the initial temperature limit; traversing all available on / off state combinations for each server rack, and for each on / off state, solving a linear programming problem for maximizing the total effective power under a given temperature limit, to obtain the on / off state combination and power allocation vector that yields the maximum total effective power; if the obtained total effective power is close to the actual required total effective power, the iteration ends. This invention offers good versatility, low computational complexity, and fast computation speed, enabling the determination of the optimal rack on / off state combination and achieving lower hotspot temperatures.

Description

Technical Field

[0001] This invention relates to the field of data center thermal management technology, and specifically to a server rack power allocation method. Background Technology

[0002] Data centers are the infrastructure supporting modern internet information services and scientific computing. Electronic components within data center server rooms have high heat density, requiring cooling systems to dissipate heat and ensure safe and stable operation. The air within the server room, through the combined action of server racks, air conditioners, and other facilities, forms a complex airflow pattern. The heat emitted by each server rack can cause temperature increases in other racks to varying degrees. The hottest points are called hotspots, which can easily lead to overheating failures and damage to equipment. By optimizing the distribution of total power among the server racks, heat concentration can be effectively reduced, hotspot temperatures lowered, and thus equipment lifespan extended, while cooling system efficiency improved. Traditional power allocation schemes include: allocating power in rotation according to server rack number or prioritizing more idle racks. This approach completely ignores the impact on the thermal environment and cannot reduce hotspot temperatures; allocating power based on simple empirical formulas cannot guarantee an optimal solution that adapts to the airflow pattern of the data center and lacks versatility; using combined optimization algorithms such as genetic algorithms for allocation, which involves a large amount of computation and is random, and cannot guarantee an optimal solution; and using greedy algorithms for allocation, which also involves a large amount of computation and does not consider practical constraints such as rack on / off status, idle power, and power limits. Summary of the Invention

[0003] To overcome the shortcomings of existing data center power allocation schemes, this invention provides a rapid power allocation method for server racks that considers practical constraints such as rack on / off status, rack idle power, and rack power limits. This method is versatile, computationally efficient, and fast, enabling it to determine the optimal combination of rack on / off states and achieve lower hotspot temperatures.

[0004] The technical solution adopted by this invention to solve its technical problem is:

[0005] This invention provides a server rack power allocation method for reducing hot spot temperatures in a data center thermal management system. The method includes the following steps:

[0006] S1. Preset all server racks to the "on" state, and set the total power required for the overall computing tasks of the data center to P. demand The components are evenly distributed across the racks as the initial state for the iteration.

[0007] S2. Quickly estimate the temperature of each rack by using a linear influence coefficient matrix describing the mutual influence between racks, and then calculate the current hotspot temperature T. maxAs the initial temperature limit T limit ;

[0008] S3. Iterate through all available switch state combinations for each server rack, and for each switch state, solve for the given temperature limit T. limit Under the condition of maximizing the total effective power P eff The linear programming problem yields the switching state combination s and power allocation vector P that maximize the total effective power; where: the total effective power P eff The sum of the power of all open racks after removing the power of idle racks;

[0009] S4, if the total effective power P is obtained eff Total effective power P close to actual demand demand If the result is positive, the iteration ends; otherwise, the temperature limit T is updated based on the calculation result of the previous step. limit Return to step S2 and continue to the next iteration.

[0010] In this invention, in step S2, the linear influence coefficient matrix is ​​obtained by linearizing the computational fluid dynamics model and energy equation, or by fitting actual measured data from a computer room temperature sensor.

[0011] In this invention, the linear programming problem to be solved in step S3 includes the idle power P of each rack. idle With maximum power constraints.

[0012] In this invention, in step S3, the HiGHS, Interior-Point, or Simplex algorithm is used to solve the linear programming problem.

[0013] In this invention, in step S3, if the obtained total effective power P eff Total effective power P required by actual needs demand The difference is less than P demand If the value is one ten-thousandth, the iteration ends.

[0014] In this invention, the formula for updating the temperature limit during iteration is:

[0015]

[0016] in: This indicates an updated temperature limit.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] 1. High versatility: This invention is not targeted at a specific data center layout. The linear influence coefficient matrix can express the impact of the heat generated by each rack on the temperature of all other racks, thus making this invention applicable to data centers of any layout.

[0019] 2. Fast computation speed: This invention transforms the computer room power allocation problem into a linear programming problem and uses HiGHS, Interior-Point, or Simplex algorithms to solve it. This method has a faster computation speed than greedy algorithms, genetic algorithms, and other methods, and can obtain the optimal solution.

[0020] 3. Able to simultaneously determine the optimal rack switch state combination: The present invention can simultaneously obtain the optimal power distribution ratio vector P and the rack switch state combination s, and can shut down some racks to reduce heat generation when necessary, or open more racks to disperse heat sources and achieve lower hot spot temperatures.

[0021] 4. Good scalability: Although the influence coefficient matrix used in this invention is obtained by approximation through computational fluid dynamics and linearization of the energy equation, it can also be obtained directly from real data centers through experimental measurements and other means, thus enabling the method to be extended to various practical situations. Attached Figure Description

[0022] Figure 1 This is a flowchart of the steps of the method of the present invention.

[0023] Figure 2 This is a schematic diagram of the data center layout in an embodiment of the present invention. Detailed Implementation

[0024] This invention provides a server rack power allocation method for reducing hot spot temperatures in a data center thermal management system. The specific steps are as follows:

[0025] S1. The coefficient matrix describing the mutual influence between the cabinets is obtained by linearizing the computational fluid dynamics model and the energy equation.

[0026] S2. Preset all server racks to the "on" state, and set the total power required for the overall computing tasks of the data center to P. demand The components are evenly distributed across the racks as the initial state for the iteration.

[0027] S3. Quickly estimate the temperature of each rack using a linear influence coefficient matrix, and then calculate the current hotspot temperature T. max As the initial temperature limit T limit ;

[0028] S4. Iterate through all available switch state combinations for each server rack. For each switch state, consider the idle power P of each rack.idle Given a maximum power constraint, the HiGHS, Interior-Point, or Simplex algorithm is used to solve for a given temperature limit T. limit Under the condition of maximizing the total effective power P eff The linear programming problem yields the switching state combination s and power allocation vector P that can achieve the maximum total effective power;

[0029] S5, if the total effective power P is obtained eff Total effective power P close to actual demand demand If the iteration ends, then the iteration ends; otherwise, follow the formula. Update temperature limit T limit Return to step S3 and continue to the next iteration.

[0030] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0031] Example 1

[0032] A server rack power allocation method for reducing hotspot temperatures in data centers includes calculation steps such as calculating the influence coefficient matrix, setting initial states, estimating rack temperatures, setting temperature limits, solving for the power allocation vector, and updating temperature limits. This method optimizes the on / off status and power allocation ratio of each server rack based on the data center layout and airflow organization pattern to achieve a reduction in hotspot temperatures.

[0033] The data center layout plan used in this embodiment is as follows: Figure 2 As shown, the server room contains 2 air conditioners, 7 server racks, and 1 piece of furniture with no heat output (shown as a blank square). The room has a raised floor, and the air conditioners supply air at a temperature of 300K beneath the raised floor. The cool air rises into the room through perforated floor tiles (shown by dashed boxes in the diagram) and enters the server racks in the direction indicated by the arrows. The exhaust air from the server racks returns to the air conditioners after passing through the room. Each server rack has an idle power P when powered on. idle = 1kW, with a power limit of 8kW per rack, meaning each rack can provide a maximum of 7kW of effective power. The current total effective power requirement P demand =19.4kW. The optimal cabinet switch state combination s and power allocation vector P are solved using the method of this invention. The steps are as follows:

[0034] S1. The coefficient matrix Θ describing the mutual influence between the racks is obtained by linearizing the computational fluid dynamics model and energy equation. Θ is a 7x7 matrix, where the element in the i-th row and j-th column represents the temperature rise of the j-th rack caused by the i-th rack increasing its heat output by 1kW, with the dimension K / kW.

[0035] S2. Preset all server racks to the open state, i.e., s = (1,1,1,1,1,1,1), and calculate the total power required for the overall computing tasks of the data center, P. demand The power is evenly distributed across the racks, i.e., the power distribution vector is P = (P idle +P demand / 7,P idle +P demand / 7,P idle +P demand / 7,P idle +P demand / 7,P idle +P demand / 7,P idle +P demand / 7,P idle +P demand / 7), as the initial state for iteration;

[0036] S3. Calculate the temperature of each rack, T = P·Θ, based on the influence coefficient matrix and power distribution vector. T is the rack temperature vector, containing 7 elements, each representing the current temperature of the rack. The maximum value of each element in T is the hotspot temperature Thotspot. max This value is used as the initial temperature limit T. limit ;

[0037] S4. Iterate through all available switch state combinations for each server rack, meaning all elements in vector s can be either 0 or 1. Excluding the case of all servers being off, there are a total of 2... 7 -1 state combinations. For each switching state combination, perform the following steps: For each cabinet, if the corresponding value in the s vector is 0, the cabinet is off and does not participate in linear programming; if it is 1, its power upper and lower limits are 8kW and 1kW, respectively. Use the HiGHS, Interior-Point, or Simplex algorithm to solve for the given upper temperature limit T. limit Under the condition of maximizing the total effective power P eff This is a linear programming problem. The total effective power is the sum of the power of all active racks minus the idle power. This step yields the switch state combination s and the power allocation vector P that maximizes the total effective power.

[0038] S5, if the total effective power P is obtained eff Total effective power P required by actual needs demand The difference is less than P demand If the result is one ten-thousandth, the iteration ends; otherwise, follow the formula. Update temperature limit T limit Return to step S3 and continue to the next iteration.

[0039] After calculation, the resulting rack switch state combination is s=(1,1,1,0,1,1,0), the power allocation vector is P=(4.23,4.67,6.57,0.00,3.68,5.25,0.00), the temperature of each rack is T=(303.62,303.62,303.62,300.33,303.62,303.62,300.51), and the hot spot temperature is 303.62K. This is a reduction of 0.60K compared to the minimum rack strategy, a reduction of 0.17K compared to the maximum rack strategy, and an average reduction of 1.38K compared to the random allocation scheme.

[0040] The above embodiments are typical implementations of the present invention, but the implementation of the present invention is not limited to the embodiments described above. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the principles and essence of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A server rack power allocation method for reducing hotspot temperatures in a data center thermal management system, characterized in that... The method includes the following steps: S1. Preset all server racks to the "on" state, and set the total power required for the overall computing tasks of the data center to P. demand The components are evenly distributed among the racks as the initial state for the iteration. S2. The temperature of each rack can be quickly estimated by using the linear influence coefficient matrix and power distribution vector describing the mutual influence between racks, T=P·Θ, where T is the rack temperature vector, and the maximum value of each element in the vector is the hot spot temperature T. max P is the power allocation vector, and Θ is the linear influence coefficient matrix; the hotspot temperature T at this time... max As the initial temperature limit T limit ; S3. Iterate through all available switch state combinations for each server rack, and for each switch state, solve for the given temperature limit T. limit Under the condition of maximizing the total effective power P eff The linear programming problem yields the switching state combination s and power allocation vector P that maximize the total effective power; where: the total effective power P eff Remove the idle rack power P from the power of all open racks. idle The sum after; S4, if the total effective power P is obtained eff Total effective power P close to actual demand demand If the result is positive, the iteration ends; otherwise, the temperature limit T is updated based on the calculation result of the previous step. limit Return to step S2 and continue to the next iteration; where: In step S3, if the obtained total effective power P eff Total effective power P required by actual needs demand The difference is less than P demand If the value is less than one ten-thousandth, the iteration ends; the formula for updating the temperature limit during iteration is: ； in: This indicates an updated temperature limit.

2. The power distribution method according to claim 1, characterized in that, In step S2, the linear influence coefficient matrix is ​​obtained by linearizing the computational fluid dynamics model and energy equation, or by fitting actual measured data from the computer room temperature sensor.

3. The power distribution method according to claim 1, characterized in that: In step S3, the linear programming problem to be solved includes the idle rack power P of each rack. idle With maximum power constraints.

4. The power distribution method according to claim 1, characterized in that: In step S3, the HiGHS, Interior-Point, or Simplex algorithm is used to solve the linear programming problem.

Images