Method for distributing computational loads to improve energy efficiency of data center it equipment operation

Abstract

A computing capacity allocation method for improving the energy efficiency of IT equipment in a data center, having n operating racks with similar energy efficiency, the optimal rack operating number of IT equipment and the optimal computing capacity allocation method of each operating rack: keep the operating energy efficiency of each rack of n operating racks equal, i.e. η1=η2=…=η n The switching point of the rack operating number from n to m is the point at which the overall energy efficiency of n operating racks η tn And the overall energy efficiency of m operating racks η tm Is equal, that is, η tn =η tm ; when all n operating racks are racks of the same model, the computing capacity is evenly distributed: Q1=Q2=…=Q n =Q t / n, m is one of (n-1) and (n+1); when all operating racks are not the same model, m value includes m=n and m≠n two cases.

Description

Technical Field

[0001] This invention relates to a method for allocating computing power to data center IT equipment, and more particularly to a method for allocating computing power to improve the energy efficiency of data center IT equipment operation. Background Technology

[0002] With the rapid development of artificial intelligence, the number of big data centers and their electricity consumption have also grown explosively, which has added new pressure to the achievement of the "dual carbon" goal. Improving the energy efficiency of data centers and reducing carbon emissions has become an inevitable trend.

[0003] Power Usage Effectiveness (PUE) is a standard for measuring the energy efficiency of data centers. PUE = Total power consumption of data center / Power consumption of IT equipment = 1 + Power consumption of non-IT equipment / Power consumption of IT equipment. The closer the PUE value is to 1, the less power the non-IT equipment (such as lighting, cooling, etc.) consumes. Measured by this standard, the data center has higher energy efficiency.

[0004] While this metric sets requirements for reducing power consumption of non-IT equipment, it doesn't put much pressure on reducing power consumption of IT equipment. On the contrary, the higher the power consumption of IT equipment, the lower the PUE. Under the same computing power, racks with high power consumption have lower PUE, and measured by the PUE standard, they are actually more energy efficient. This is a shortcoming of using the PUE metric to measure the energy efficiency of data centers.

[0005] The industry commonly uses "power consumption per rack (kW)" to represent the "power density" of a data center. A data center's power density is related to the performance and power of its IT equipment, as well as the types of services and functions it supports. "Computational power" represents a data center's "ability to process data and perform computational tasks," encompassing information processing, network capacity, and data storage, measured in floating-point operations per second (FLOPS). Once the data center's hardware is determined, metrics such as "power density" and "ability to process data and perform computational tasks" remain fixed. However, the actual computational load and power consumption of the data center are constantly changing; in other words, the data center's energy efficiency is constantly evolving.

[0006] The single-rack operating energy efficiency η is expressed as the single-rack computing power divided by the single-rack power. i The total energy efficiency (η) of data center IT equipment is expressed as the total computing power of the data center divided by the total power of all operating racks. t .

[0007] Under the condition of the same total computing load in the data center, by reasonably allocating the computing load of each rack, the overall operating energy efficiency of IT equipment can be optimized to reduce the total energy consumption of data center IT equipment.

[0008] There are many studies on reducing the energy consumption of data center IT equipment, but the total operating energy efficiency η of IT equipment is still the most important factor. t There are not many solutions to optimization problems that provide a clear optimal operating method. Optimization problems include: determining the number of racks to use and determining the computational cost of each operating rack.

[0009] Patent (US20250217160A1) describes a method that allows multiple smaller CPU cores to dynamically combine and jointly execute a single-threaded task, simulating a more powerful core. This approach can improve performance without increasing voltage and frequency (traditional methods for improving single-core performance), thus significantly improving performance per watt. Patent (US10133323B2) proposes DVFS technology based on Performance Monitoring Unit (PMU) events, dynamically adjusting voltage and frequency according to real-time processor activity (such as memory-intensive tasks) to achieve higher energy efficiency within power consumption constraints. Patent (CN102213475A) describes a method that predicts server hotspot distribution using neural networks and adjusts the airflow at air conditioning terminals in real time. Patent (CN 118260066B) dynamically migrates computing tasks (bits) based on different data centers' power costs, renewable energy ratios, and real-time Power Usage Effectiveness (PUE) status to achieve globally optimal energy utilization.

[0010] Patent (CN102213475A) provides a partitioned control method that dynamically adjusts server power consumption based on real-time business load and combines it with fuzzy neural network control of air conditioning to achieve coordinated energy saving in IT and cooling. Patent (CN109542206A) describes a low-energy system combining hibernation, frequency reduction and voltage reduction (DVFS), and liquid cooling technology. With the surge in AI computing power demand in 2026, liquid cooling penetration is projected to reach 35%. Patent (CN103345298A) utilizes a virtual machine placement algorithm to dynamically integrate virtual resources based on business volume and shut down idle physical servers. Patent (US11853936B2) proposes a method for redistributing workloads across multiple data centers. The system monitors the proportion of renewable energy in each location and migrates computing tasks to data centers with more abundant clean energy to reduce overall environmental impact. Patent (ZL201210331205.9) employs IoT technology to monitor the data center environment in real time, reducing local hotspots through precise data collection and avoiding energy waste caused by over-cooling. The patent (CN106152394A) provides a comprehensive information management method that can issue control commands to terminal devices for energy-saving adjustments based on server workload.

[0011] Regarding the optimal study of energy efficiency in multi-unit systems, the literature "Efficient Energy-Saving Control and Optimization for Multi-Unit Systems - A Guide for Electrical Engineers" (published by Springer) provides the optimal solution for operating energy efficiency, proves Yao's Theorem 1 for achieving optimal load allocation, and proves Yao's Theorem 2 for achieving optimal unit switching. However, the book does not provide an engineering implementation method for achieving optimal energy efficiency in multi-rack operation of data centers. Summary of the Invention

[0012] To address the problem of optimizing the overall operating energy efficiency of IT equipment in data centers, this invention, based on Yao's Theorem 1 and Yao's Theorem 2, provides a specific method for optimizing the overall operating energy efficiency of IT equipment in data centers. For each total computing load in a data center, it provides a method for determining the number of racks in operation and a method for allocating computing load to each operating rack.

[0013] The technical solution adopted by this invention to solve its technical problem is: in a data center composed of multiple racks, the total computing power of IT equipment is Q. t Q t The total number of floating-point operations performed per second in the data center, and the total power consumed by the IT equipment, are given in W.t The energy efficiency curve η1 of the first rack is defined as the curve of the rack's "computational load / watts" changing with the rack's "computational load," where "computational load" is the number of floating-point operations per second (Q1), and "watts" is the rack's power consumption (W1) corresponding to Q1. Q1 is the horizontal axis, and energy efficiency η1 = Q1 / W1 is the vertical axis. The second derivative of the energy efficiency curve η1 is less than zero. The energy efficiency curve of the i-th rack is η1. i Satisfying η i (Q) i ) =η1(β i *Q1), β i Let η be a constant. The i-th rack and the 1st rack are called energy-efficiency similar racks. When all n operating racks are energy-efficiency similar racks, the optimal computational allocation method is to keep the operating energy efficiency of each of the n operating racks equal, that is, η1=η2=…=η n The switching point for the number of racks operating from n to m is at the overall energy efficiency η of the n operating racks. tn The overall energy efficiency η of m operating racks tm The point of equality, i.e., η tn =η tm Point; when all n running racks are of the same model, the computational load is evenly distributed: Q1=Q2=…=Q n =Q t / n, where m is one of (n-1) and (n+1); when all running racks are not the same model, the value of m includes two cases: m=n and m≠n.

[0014] The beneficial effect of this invention is that this method provides a high-efficiency operation control method for IT equipment in multi-rack data centers. Since the technical measures adopted in this invention are all mature and feasible, they can be implemented. Attached Figure Description

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0016] Figure 1 This is the first preferred embodiment of the present invention.

[0017] Figure 2 This is the second preferred embodiment of the present invention. Detailed Implementation

[0018] exist Figure 1 In the diagram, the energy efficiency curve of the first rack is η1(Q), and the energy efficiency curve of the i-th rack is η. i (Q), where Q is the computational quantity, in Q=Q 1e At point Q=Q, the energy efficiency of the first rack reaches its maximum value. ieAt point i, the energy efficiency of the i-th rack reaches its maximum value. The two racks are different models. When Q=Q 1-i Point, η1(Q) = η i (Q), the total computing power of data center IT equipment is Q. t Q t 1-i At that time, only the first rack is running, and the IT equipment operates with the highest energy efficiency; Q t =Q 1-i At this time, the energy efficiency of running the first rack and running the i-th rack are equal, and the number of racks on which IT equipment is running does not need to be switched; Q t Q 1-i When only the i-th rack is running, the IT equipment has the highest energy efficiency. Switching from the 1st rack (n=1) to the i-th rack (m=1, m=n), the switching point is when the overall energy efficiency of the n-rack system equals that of the m-rack system, i.e., η1(Q) = η. i (Q) point.

[0019] exist Figure 2 In this system, all racks in the data center IT equipment are of the same model, and the total computing power of the data center IT equipment is Q. t Currently, n racks are running simultaneously. The optimal method for allocating computational load is to maintain a computational load of Q per rack. t / n, Q1=Q2=…=Q n =Q t / n, n operating racks, each rack has the same operating energy efficiency, η1=η2=…=η n Overall energy efficiency η t =η1, where n is the optimal number of operating racks, and the overall energy efficiency η1 (Q) of operating n racks is... t / n) compared to the overall energy efficiency η1 (Q) of (n-1) racks t The overall energy efficiency η1 (Q) of the operation of (n-1) and (n+1) racks t / (n+1)) must be higher, when the total computing power Q of the data center IT equipment is Q t When it increases, η1(Q) t / (n+1)) keeps increasing, η1(Q) t / n) first increases and then decreases, when η1(Q) t / n)=η1(Q t When / (n+1)), if Q t Continue to increase, η1(Q) t / n)<η1(Q t If the total computing power Q of the data center IT equipment is / (n+1), then it should be switched from running on n racks to running on (n+1) racks. Similarly, when the total computing power Q of the data center IT equipment is Q, t ​When it decreases, η1(Q) t / (n-1)) keeps increasing, η1(Q) t / n) is decreasing, when η1(Q) t / n)=η1(Q t When / (n-1)), if Q t Continue to decrease, η1(Q) t / n)<η1(Q t If the number of racks is (n-1), then the operation should be switched from running on n racks to running on (n-1) racks.

[0020] Those skilled in the art will recognize that although only two embodiments have been described above, they are not all forms of the present invention. It should be understood that many modifications can be made without departing from the spirit and scope of the invention. In the description of the present invention, it should be understood that terms such as "first unit," "i-th unit," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first unit," "i-th unit," etc., may explicitly or implicitly include one or more of that feature. In the description of the present invention, unless otherwise stated, "a plurality of" means two or more. The definitions of performance indicators such as "computing power," "watt," etc., can also be replaced with "computing power," "kilowatt," etc., only for the convenience of describing the present invention and simplifying the description, and not to indicate or imply necessary descriptions and definitions, and therefore should not be construed as limiting the present invention. Those skilled in the art can implement the present invention in other forms without departing from the concept of the present invention; therefore, other embodiments are also within the scope of the claims of the present invention.

Claims

1. A method for allocating computational load to improve the energy efficiency of IT equipment in a data center. In a multi-rack data center, the total computational load of IT equipment is Q. t Q t The total number of floating-point operations performed per second in the data center, and the total power consumed by the IT equipment, are given in W. t The energy efficiency curve η1 of the first rack is defined as the curve of the rack's "computation load / watt" changing with the rack's "computation load," where "computation load" is the number of floating-point operations per second (Q1), and "watts" is the rack's power consumption (W1) corresponding to that "computation load" Q1. Q1 is the horizontal axis, and energy efficiency η1 = Q1 / W1 is the vertical axis. The second derivative of the energy efficiency curve η1 is less than zero. The energy efficiency curve of the i-th rack is η1. i Satisfying η i (Q) i ) =η1(β i *Q1), β i Let be a constant. The i-th rack and the 1-th rack are called energy-efficient racks, characterized by: When n operating racks are all energy-efficient racks, the optimal method for allocating computational resources is to ensure that the operating energy efficiency of each of the n operating racks is equal, i.e., η1=η2=…=η n The switching point for the number of racks operating from n to m is at the overall energy efficiency η of the n operating racks. tn The overall energy efficiency η of m operating racks tm The point of equality, i.e., η tn =η tm point.

2. The computational load allocation method for improving the energy efficiency of data center IT equipment according to claim 1, characterized in that: When all n running racks are of the same model, the computational load is evenly distributed: Q1=Q2=…=Q n =Q t / n, where m is a value between (n-1) and (n+1).

3. The computational load allocation method for improving the energy efficiency of data center IT equipment according to claim 1, characterized in that: When all running racks are not the same model, the value of m includes two cases: m=n and m≠n.

Images