Distributed server resource allocation system
By acquiring load and temperature data of server nodes, and combining this with hardware performance degradation rate and fault prediction, resource allocation is dynamically optimized, solving the problems of load balancing, high availability, and high energy efficiency in distributed server clusters, and realizing intelligent resource management of the system.
Patent Information
- Application Number
- CN202511922087.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-18
- Publication Date
- 2026-04-03
AI Technical Summary
Existing technologies cannot simultaneously meet the requirements of load balancing, high availability, and high energy efficiency in distributed server clusters. Traditional resource allocation systems fail to effectively prevent hardware degradation and failures, leading to decreased system performance and increased energy consumption.
By acquiring server node load, resource, and temperature data, and combining hardware performance degradation rate and fault prediction, resource allocation is dynamically optimized. Collaborative optimization is achieved by using node load resilience index, fault prediction confidence, and energy efficiency weighting coefficient.
It achieves accurate quantification of the actual load capacity of nodes, avoids resource waste and overload, improves system availability and energy efficiency, reduces total energy consumption, and solves the problems of load assessment bias and failure risk in traditional systems.
Smart Images
Figure CN121785774A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of resource allocation technology, specifically a distributed server resource allocation system. Background Technology
[0002] With the rapid development of technologies such as cloud computing, big data, and artificial intelligence, distributed server clusters have become the core infrastructure supporting various Internet businesses, such as e-commerce platforms, financial transaction systems, and AI training platforms. Such clusters typically consist of tens to thousands of servers and need to handle high concurrency and various types of business requests, such as computationally intensive AI model training, memory-intensive database queries, and network-intensive video streaming. In practical applications, the core objective of distributed server resource allocation is to balance load pressure, ensure system availability, and reduce energy consumption costs, but existing technologies still have shortcomings.
[0003] Firstly, after long-term server operation, hardware such as CPU, memory, and network will experience performance degradation due to issues such as thermal paste aging, electromigration, and interface oxidation. Traditional resource allocation systems only assess load based on real-time utilization, ignoring false idle situations caused by hardware degradation. Furthermore, in distributed clusters, single-node failures can trigger business interruptions. Most traditional allocation systems only perform resource migration after a failure occurs, making it difficult to achieve proactive prevention. Since existing technologies cannot simultaneously meet the multiple objectives of "load balancing, high availability, and high energy efficiency," there is an urgent need for an intelligent resource allocation system that integrates hardware status, failure risk, and energy efficiency characteristics to improve upon existing technologies. Summary of the Invention
[0004] The purpose of this invention is to provide a distributed server resource allocation system that solves the problems mentioned in the background section.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a distributed server resource allocation system, comprising:
[0006] Data acquisition component: used to acquire node load-related data, resource allocation-related data, and temperature-related data during server operation;
[0007] Storage generation component: Used to input node load-related data, resource allocation-related data, and temperature-related data acquired by the data acquisition component, and to clean the input data before inputting the cleaned data into the allocation calculation component;
[0008] Assignment calculation components:
[0009] The node load resilience index is output based on CPU utilization, CPU performance degradation rate, memory utilization, memory performance degradation rate, network utilization, network performance degradation rate and weighting coefficients. The node load resilience index accurately quantifies the actual load capacity of the node.
[0010] Based on the node load resilience index, fault prediction confidence, and fault prediction weight coefficient, resource allocation priority is output, and load conditions and fault prevention are collaboratively optimized through resource allocation priority.
[0011] Based on resource allocation priority, full load power, idle power, dynamic heat dissipation efficiency, and energy efficiency weighting coefficient, an energy efficiency optimization factor is output, which provides the final decision indicator for resource allocation.
[0012] Build allocation component: Dynamically allocate server resources based on energy efficiency optimization factors.
[0013] Optionally, the allocation calculation component includes a load calculation module, a priority allocation module, and an energy efficiency optimization module.
[0014] Optionally, the data acquisition process of the data acquisition component includes:
[0015] CPU utilization is collected through CPU performance monitoring sensors on the server motherboard, CPU core temperature is recorded through CPU temperature sensors, cumulative power-on time of the CPU is recorded based on CPU runtime sensors, memory utilization is directly collected through memory controller sensors on the server motherboard, the number of memory ECC errors is recorded through memory error rate sensors, network utilization is directly collected through network traffic sensors on the server network card, and the number of data packets lost by the network card is recorded based on the network card packet loss rate sensor.
[0016] Optionally, the data acquisition process of the data acquisition component further includes:
[0017] The server power module's power sensor directly collects full-load power and idle power, the server's air inlet temperature is collected by the temperature sensor at the front of the server, the server's air outlet temperature is collected by the temperature sensor at the rear of the server, and the data center's ambient temperature is collected by the ambient temperature sensor in the server room.
[0018] Optionally, the processing flow of the load calculation module is as follows:
[0019] The CPU performance degradation rate is calculated based on the cumulative power-on time of the CPU and the benchmark performance test data. The memory performance degradation rate is calculated based on the number of ECC errors of the memory and the benchmark bandwidth test data. The network performance degradation rate is calculated based on the number of packet loss of the network card and the benchmark throughput test data. The weight coefficients are trained based on the historical fault data of the system.
[0020] A1: Introduce quantitative decay of CPU utilization and CPU performance decay rate to make load assessment more in line with the actual hardware status.
[0021] A2: By incorporating memory usage and memory performance degradation into load assessment, the dynamic performance degradation of memory is quantified, making the load assessment more closely reflect the actual hardware condition.
[0022] A3: By incorporating network link performance degradation into load assessment through network utilization and network performance degradation rate, the degradation is quantified, making load assessment more accurate.
[0023] Optionally, the processing flow of the priority partitioning module is as follows:
[0024] The confidence level of fault prediction is obtained by training a Naive Bayes model based on CPU core temperature, memory ECC error count and network card packet loss count.
[0025] B1: By dividing the node load resilience by the sum of the load resilience of all nodes, resources are ensured to be tilted towards truly idle nodes, avoiding resource waste and overload risks.
[0026] B2: Introduce fault prediction confidence and fault prediction weight coefficient, and multiply them to bring fault prevention forward, ensuring that resources are tilted towards reliable nodes.
[0027] Optionally, the processing flow of the energy efficiency optimization module is as follows:
[0028] Dynamic heat dissipation efficiency is calculated by taking the server inlet temperature, server outlet temperature and data center ambient temperature, and energy efficiency weighting coefficient is determined based on the total energy consumption target of the system.
[0029] C1: By multiplying with resource allocation priority, a balance between performance and energy efficiency is achieved in resource allocation;
[0030] C2: The dynamic energy efficiency of a node is evaluated by the difference between full-load power and idle power, and then the ratio of the idle power to the full-load power is used to evaluate the dynamic energy efficiency, making energy efficiency optimization more in line with actual operating scenarios.
[0031] C3: By adjusting the energy efficiency ratio of 1 to dynamic heat dissipation efficiency, resource allocation is made more in line with the overall energy consumption characteristics of the data center.
[0032] Optionally, the processing flow for building the allocation component is as follows:
[0033] The energy efficiency optimization factors of all nodes are normalized to obtain the resource allocation ratio of each node, and CPU, memory and network bandwidth are allocated according to the allocation ratio.
[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0035] I. This invention achieves accurate quantification of the actual load capacity of a node by coupling the calculation of CPU utilization, memory utilization, network utilization and corresponding performance degradation rate, combined with dynamic weighting coefficients driven by business. Among them, CPU utilization, memory utilization and network utilization directly reflect the real-time load pressure of each hardware dimension of the node, while CPU performance degradation rate, memory performance degradation rate and network performance degradation rate correct the load assessment deviation caused by hardware aging, making the load index more in line with the actual hardware bearing state, and solving the load assessment mismatch problem caused by traditional systems ignoring hardware aging and business differences.
[0036] Second, this invention uses the node load resilience index as a basis, combined with fault prediction confidence and fault prediction weight coefficient, to achieve synergistic optimization of load balancing and fault prevention. The node load resilience index serves as the core basis for load balancing, ensuring that resources are tilted towards truly idle nodes. The fault prediction confidence quantifies fault risk and incorporates it into the resource allocation logic, giving higher priority to nodes with low fault risk. The fault prediction weight coefficient dynamically adjusts the priority of fault prevention according to business availability requirements, balancing the relationship between load balancing and fault prevention. This solves the problem of the disconnect between fault risk and resource allocation in traditional systems, and improves the overall availability of the system.
[0037] Third, this invention integrates dynamic power difference, dynamic heat dissipation efficiency, and energy efficiency weight coefficient driven by energy consumption targets, based on resource allocation priority, to achieve coordinated energy efficiency optimization of the system and infrastructure. Among them, resource allocation priority ensures that energy efficiency optimization does not sacrifice load balancing and fault prevention goals; dynamic power difference quantifies the dynamic energy consumption characteristics of nodes and prioritizes resource allocation to nodes with good dynamic energy efficiency; dynamic heat dissipation efficiency incorporates the data center heat dissipation system into the resource allocation logic to reduce heat dissipation energy consumption; and energy efficiency weight coefficient adjusts the priority of energy efficiency optimization according to the system energy consumption target, balancing the relationship between performance and energy efficiency. This solves the problem of the lack of coordination between dynamic energy efficiency and heat dissipation system in traditional systems and reduces the total energy consumption of the system.
[0038] Fourth, this invention achieves multi-objective optimization of "load balancing, high availability, and high energy efficiency," breaking through the limitations of "isolated optimization" in traditional distributed resource allocation systems. It integrates three cross-domain characteristics—"hardware performance degradation, fault prediction confidence, and dynamic heat dissipation efficiency"—into a unified resource allocation logic, realizing "dynamic, intelligent, and business-adaptive" resource allocation. It can dynamically adjust resource allocation strategies based on hardware status, fault risk, business type, and energy consumption targets, providing a quantifiable and scalable technical solution for resource management of distributed server clusters. This promotes the technological upgrade of distributed computing from "static allocation" to "intelligent collaborative allocation," improving system stability, availability, and energy efficiency. Attached Figure Description
[0039] Figure 1 This is a diagram illustrating the distribution process of the present invention;
[0040] Figure 2 A flowchart illustrating the allocation of computational components for this invention. Detailed Implementation
[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0042] Example: Please refer to Figure 1 and Figure 2 This invention provides a distributed server resource allocation system, comprising: a data acquisition component, which acquires node load-related data, resource allocation-related data, and temperature-related data during server operation. The data acquisition process of the data acquisition component includes acquiring CPU utilization data via a CPU performance monitoring sensor on the server motherboard. util,i The CPU core temperature is recorded by a CPU temperature sensor, the cumulative power-on time of the CPU is recorded by a CPU runtime sensor, and the memory usage (Mem) is directly collected by a memory controller sensor on the server motherboard. util,i The system records the number of ECC errors in the memory using a memory error rate sensor, and directly collects network usage data (Network usage) using a network traffic sensor on the server's network interface card. util,i The number of data packets lost by the network card is recorded based on the network card packet loss rate sensor;
[0043] This also includes directly acquiring the full-load power P through the power sensor of the server power module. active,i and idle power P idle,iThe server's air inlet temperature T is collected by a temperature sensor at the front of the server. in,i The server exhaust temperature T is collected by a temperature sensor at the rear of the server. out,i The ambient temperature T in the data center is collected by the ambient temperature sensor in the computer room. amb ;
[0044] The resource allocation system also includes a storage generation component: used to input node load-related data, resource allocation-related data, and temperature-related data acquired by the data acquisition component, and to clean the input data before inputting the cleaned data into the allocation calculation component; the allocation calculation component includes a load calculation module, which first calculates the CPU performance degradation rate λ based on the CPU's cumulative power-on time and benchmark performance test data. i The specific calculation method is as follows: λ i = 1 - Current benchmark score / Factory benchmark score, where the current benchmark score is obtained by periodically running computationally intensive testing programs, and the factory benchmark score is the standard performance value of the server at the time of manufacture. The memory performance degradation rate μ is calculated based on the number of ECC errors in the memory and the benchmark bandwidth test data. i Specific calculation method: μ i = 1 - Current memory bandwidth / Factory memory bandwidth, where the current memory bandwidth is obtained by periodically running a memory bandwidth test program, and the factory memory bandwidth is the standard bandwidth value when the server leaves the factory. The network performance degradation rate v is calculated based on the number of packet losses of the network card and the baseline throughput test data. i The specific calculation method is as follows: v i = 1 - Current network throughput / Factory network throughput, where the current network throughput is obtained by running network testing programs periodically, and the factory network throughput is the standard throughput value when the server leaves the factory. The weight coefficients are obtained by training based on historical system failure data, and the weights of each dimension are determined by minimizing the error through a linear regression model.
[0045] The load calculation module is based on CPU utilization. util,i CPU performance degradation rate λ i Memory usage (Mem) util,i Memory performance degradation rate μ i Network usage rate (Net) util,i Network performance degradation rate v i The node load resilience index (LRI) is output along with weighting coefficients, which accurately quantifies the actual load capacity of a node.
[0046] A1: Introduce quantitative decay of CPU utilization and CPU performance decay rate to make load assessment more in line with the actual hardware status.
[0047] A2: By incorporating memory usage and memory performance degradation into load assessment, the dynamic performance degradation of memory is quantified, making the load assessment more closely reflect the actual hardware condition.
[0048] A3: By incorporating network link performance degradation into load assessment through network utilization and network performance degradation rate, the degradation is quantified, making load assessment more accurate;
[0049] Specifically, the load accounting module calculates load as follows:
[0050]
[0051] CPU util,i CPU utilization, ranging from 0 to 1, represents the real-time percentage of CPU usage at node i. This data is collected directly from the CPU performance monitoring sensor on the server motherboard, located near the CPU cores. The sensor continuously monitors the CPU core's instruction execution duty cycle. util,i It reflects the real-time load pressure of the node's CPU. The higher the utilization rate, the lower the node's remaining capacity to process computing tasks, which directly affects the priority of resource allocation.
[0052] λ i LRI is the CPU performance degradation rate, ranging from 0 to 0.2. It represents the percentage of CPU performance degradation caused by long-term operation of node i. It is calculated based on the cumulative power-on time of the CPU recorded by the CPU runtime sensor and the benchmark performance test data. By correcting the defect of ignoring hardware aging in traditional load assessment, LRI can more accurately reflect the true load capacity of the node. The more severe the performance degradation, the greater the actual load pressure of the node under the same utilization rate, thus avoiding the risk of overload due to hardware aging.
[0053] Mem util,i This represents memory usage, ranging from 0 to 1. It is directly collected by a sensor on the server motherboard's memory controller, located near the memory slots. The sensor calculates the ratio of used memory capacity to total memory capacity in real time, and then transmits this data via Mem... util,i It reflects the real-time load pressure of node memory. The higher the utilization rate, the lower the remaining capacity of the node to handle memory-intensive tasks, which affects the balance of resource allocation.
[0054] μ i The memory performance degradation rate, ranging from 0 to 0.2, is calculated based on the number of ECC errors recorded by a memory error rate sensor and benchmark bandwidth test data, expressed in μ. i This corrects the flaw in traditional memory load assessment that ignores hardware aging, making LRI more accurately reflect the node's true memory load capacity. The more severe the performance degradation, the greater the actual memory pressure on the node under the same utilization rate, thus avoiding service lag caused by memory aging.
[0055] Net util,i This represents network utilization, ranging from 0 to 1. It is directly collected via the network traffic sensor on the server's network interface card (NIC). The sensor is located inside the NIC chip and calculates the ratio of the NIC's transmit / receive traffic to the total bandwidth in real time. util,i It reflects the real-time load pressure of the node network. The higher the utilization rate, the lower the remaining capacity of the nodes to handle network-intensive tasks, thus affecting the communication efficiency of the distributed system.
[0056] v i The network performance degradation rate, ranging from 0 to 0.2, is calculated based on the number of packet losses recorded by the network card packet loss rate sensor and benchmark throughput test data. This is achieved through v... i It corrects the shortcomings of traditional network load assessment that ignores hardware aging, making LRI more accurately reflect the actual network load capacity of a node. The more severe the performance degradation, the greater the actual network pressure on the node under the same utilization rate, thus avoiding communication delays caused by network card aging.
[0057] α, β, and γ are weighting coefficients, representing the contribution weights of CPU, memory, and network utilization in the Node Load Resilience Index (LRI), respectively, satisfying α + β + γ = 1. These coefficients are derived from training based on historical system failure data. The error between LRI and actual system performance is minimized through a linear regression model to determine the weights of each dimension. The priority of each resource dimension is dynamically adjusted according to the business type, making LRI more closely match the load characteristics of actual business scenarios, improving the accuracy of resource allocation, and avoiding mismatch between resource allocation and business needs. In this embodiment, α is set to 0.35-0.45, β to 0.25-0.35, and γ to 0.25-0.35. Considering that CPU is the hardware dimension most relied upon by the business, the impact of failures in this item is relatively higher.
[0058] It is worth noting that α×CPU util,i / (1-λ i The item adopts a fractional structure of "load rate ÷ (1 - attenuation rate)" to couple hardware performance degradation with load rate, and restores the degraded hardware capability to the load rate under "equivalent healthy hardware capability". This ensures that the calculation results reflect the actual load pressure of the node. Considering that hardware performance will degrade due to thermal grease aging, electromigration and other factors during long-term operation, if the attenuation is ignored and the utilization rate is used directly to evaluate the load, the pressure on the aging node will be underestimated and overload will be caused. This item quantifies the attenuation through physical layer sensor data, so that the load evaluation is more in line with the actual hardware state.
[0059] Considering that long-term memory operation can lead to increased ECC error rates and read / write latency, evaluating only capacity utilization could result in a misjudgment of "idle memory capacity but performance bottleneck." This item uses β×Mem... util,i / (1-μ)i The dynamic performance degradation of memory is incorporated into the load assessment. The degraded memory performance is restored to the equivalent healthy memory load rate using "capacity utilization ÷ (1 - degradation rate)". This ensures that the load assessment covers both capacity and performance dimensions, making the assessment more comprehensive. However, long-term operation of network cards can lead to increased packet loss and transmission latency. Assessing only bandwidth utilization would result in wasted resources such as "idle bandwidth but failed transmissions". Therefore, γ×Net addresses this issue. util,i / (1-v) i The load assessment incorporates network link performance degradation and uses "bandwidth utilization ÷ (1 - degradation rate)" to restore the degraded network link performance to the load rate of an equivalent healthy link, making the load assessment more accurate.
[0060] Thus, by CPU utilization CPU util,i Memory usage (Mem) util,i Network usage rate (Net) util,i With network performance degradation rate v i By combining coupled calculations with business-driven weighting coefficients, a precise quantification of the actual load capacity of nodes is achieved, including CPU utilization. util,i Memory usage (Mem) util,i Network usage rate (Net) util,i Directly reflects the real-time load pressure of each hardware dimension of the node, CPU performance degradation rate λ i Memory performance degradation rate μ i Network performance degradation rate v i This corrects the load assessment bias caused by hardware aging, making the load index more closely match the actual hardware load status. The dynamic weight coefficients α, β, and γ adjust the priority of each resource dimension according to the business type, ensuring that the load assessment is highly adapted to business needs. This provides an accurate quantitative basis for load resilience for subsequent resource allocation and solves the problem of load assessment mismatch caused by traditional systems ignoring hardware aging and business differences.
[0061] The allocation computing component also includes a priority partitioning module. This module first trains a Naive Bayes model based on CPU core temperature, memory ECC error count, and network card packet loss count to predict fault prediction confidence. Inputs include CPU core temperature, recorded ECC error count, and packet loss count. The model output is the fault probability P(Failure). i,t ), Conf i,t =1 - P(Failure) i,t Based on the node load resilience index LRI i Fault prediction confidence Conf i,tThe fault prediction weight coefficient k outputs the resource allocation priority RAP, and the resource allocation priority RAP is used to perform collaborative optimization of load conditions and fault prevention.
[0062] B1: By dividing the node load resilience by the sum of the load resilience of all nodes, resources are ensured to be tilted towards truly idle nodes, avoiding resource waste and overload risks.
[0063] B2: Introduce fault prediction confidence and fault prediction weight coefficient, and multiply them to advance fault prevention and ensure that resources are tilted towards reliable nodes;
[0064] Specifically, the calculation method for prioritizing module division is as follows:
[0065]
[0066] LRI i The node load resilience index, ranging from 0 to 1, serves as the basis for resource allocation. Nodes with a lower LRI can be allocated more resources, ensuring that resources are tilted towards idle nodes to achieve load balancing.
[0067] Conf i,t The fault prediction confidence level, ranging from 0 to 1, represents the complement of the fault probability of node i in the future time t, ranging from 0 to 1. The higher the value, the lower the fault risk. Fault prevention is integrated into resource allocation, and resources are allocated to nodes with low fault risk first, reducing service interruptions caused by node failures and improving system availability.
[0068] k is the fault prediction weight coefficient, ranging from 0 to 0.5, which represents the contribution weight of fault prediction confidence in resource allocation priority (RAP). It is determined according to the system service level agreement. By balancing the relationship between load balancing and fault prevention, when the system availability requirements are high, the weight of fault prediction is increased to ensure that resource allocation is tilted towards reliable nodes.
[0069] in As a fundamental component of load balancing, considering that traditional load balancing only focuses on "current utilization" and ignores "false idle" caused by hardware aging, this component is based on the Node Load Resilience Index (LRI). It breaks through the traditional static balancing thinking of "equal distribution of resources according to utilization" and adopts dynamic balancing driven by load resilience to achieve the balance of "true idle capacity". Through the fractional structure of "node load resilience ÷ sum of the load resilience of all nodes", it achieves the normalization of resource allocation and ensures the fairness and scalability of resource allocation.
[0070] 1 + k × Conf i,tAs a fault prevention correction item, considering that the availability of a distributed system depends on the reliability of nodes, traditional resource allocation does not take into account the risk of node failure, leading to the hidden danger of "resources being concentrated on high-failure nodes". This item uses a linear correction structure of "1 + weight × confidence" to ensure that resources are tilted towards reliable nodes, breaking through the traditional passive operation and maintenance thinking of "failure first and then recovery". It integrates the confidence of fault prediction into resource allocation to achieve proactive prevention. Considering that different businesses have significantly different availability requirements, fixed weights will lead to excessive redundancy or insufficient availability. By multiplying with the fault prediction weight coefficient k, the priority of fault prevention in resource allocation is ensured, making resource allocation more in line with the availability requirements of the business.
[0071] Thus, the node load resilience index LRI is output. i Based on this, combined with fault prediction confidence Conf i,t By combining the fault prediction weighting coefficient k, load balancing and fault prevention are synergistically optimized, where the node load resilience index LRI is used. i As the core basis of load balancing, ensuring that resources are tilted towards truly idle nodes, the failure prediction confidence level (Conf) is crucial. i,t By quantifying and incorporating fault risk into the resource allocation logic, nodes with low fault risk are given higher priority. The fault prediction weight coefficient k is dynamically adjusted based on business availability requirements to prioritize fault prevention, thus balancing load balancing and fault prevention. This solves the problem of the disconnect between fault risk and resource allocation in traditional systems and improves the overall availability of the system.
[0072] The distribution computing component also includes an energy efficiency optimization module, which uses the server's air inlet temperature T... in,i Server exhaust temperature T out,i and data center ambient temperature T amb Calculate the dynamic heat dissipation efficiency Eff cool,i The energy efficiency weighting coefficient δ is determined based on the total energy consumption target of the system, and the dynamic heat dissipation efficiency is calculated as Eff. cool,i = (T) in,i -T amb ) / (T out,i -T in,i ), and based on resource allocation priority RAP i Full load power P active,i Idle power P idle,i Dynamic heat dissipation efficiency (Eff) cool,i The energy efficiency weighting coefficient δ outputs the energy efficiency optimization factor EOF, which provides the final decision indicator for resource allocation.
[0073] C1: By multiplying with resource allocation priority, a balance between performance and energy efficiency is achieved in resource allocation;
[0074] C2: The dynamic energy efficiency of a node is evaluated by the difference between full-load power and idle power, and then the ratio of the idle power to the full-load power is used to evaluate the dynamic energy efficiency, making energy efficiency optimization more in line with actual operating scenarios.
[0075] C3: By adjusting the energy efficiency ratio of 1 to dynamic heat dissipation efficiency, resource allocation is made more in line with the overall energy consumption characteristics of the data center.
[0076] Specifically, the calculation method for the energy efficiency optimization module is as follows:
[0077]
[0078] RAP i Prioritize resource allocation, ranging from 0 to 1, as the basis for energy efficiency optimization. Nodes with higher resource allocation priority (RAP) can be allocated more resources. At the same time, adjustments are made in conjunction with energy efficiency indicators to achieve synergistic optimization of performance and energy efficiency.
[0079] P active,i The full-load power reflects the real-time power of node i when it is running at full load. It is directly collected by the power sensor of the server power module. The sensor is located at the power input interface and counts the total power consumption of the server in real time. It reflects the energy consumption level of the node when it is running at full load. The higher the power, the higher the energy consumption cost of the node, which affects the total energy consumption of the system.
[0080] P idle,i Idle power represents the real-time power of node i when it is idle. It is directly collected by the power sensor of the server power module. The sensor is located at the power input interface and counts the total power consumption of the server in real time. It mainly reflects the energy consumption level of the node when it is idle. The higher the power, the higher the standby power consumption of the node, which affects the total energy consumption of the system.
[0081] Eff cool,i Dynamic heat dissipation efficiency represents the efficiency of node i's heat dissipation system in transferring heat from the inside of the server to the external environment, ranging from 1 to 5. Based on calculations, by co-optimizing the data center heat dissipation system with server resource allocation, nodes with higher heat dissipation efficiency have better energy efficiency, can allocate more resources, and reduce additional energy consumption caused by insufficient heat dissipation.
[0082] δ is the energy efficiency weight coefficient, which represents the contribution weight of energy efficiency indicators in the energy efficiency optimization factor EOF. It ranges from 0 to 0.4 and is determined according to the total energy consumption target of the system. This balances the relationship between performance and energy efficiency. When the system energy consumption target is high, the weight of energy efficiency indicators is increased to ensure that resource allocation is tilted towards high energy efficiency nodes and reduce the total energy consumption cost of the system.
[0083] EOF i Energy efficiency optimization factor;
[0084] It is worth noting that, considering that the energy consumption of distributed systems mainly comes from dynamic loads, such as full-load operation during peak business periods, traditional static energy efficiency assessments are difficult to reflect the actual energy consumption characteristics of nodes. Therefore, an energy efficiency ratio correction term is used: δ×(P) active,i -P idle,i ) / P active,i Breaking away from the traditional static energy efficiency thinking that "only focuses on full load power", it adopts dynamic power difference driven energy efficiency assessment. Through the fractional structure of "(full load power - idle power) ÷ full load power", it quantifies dynamic energy efficiency, making energy efficiency optimization more in line with actual operating scenarios. It also multiplies with the energy efficiency weight coefficient δ to optimize the traditional static optimization thinking with fixed energy efficiency weight, making energy efficiency optimization more in line with the energy consumption needs of data centers.
[0085] Furthermore, considering that cooling systems account for a significant portion of data center energy consumption, and that traditional energy efficiency optimization often overlooks cooling costs, leading to resource waste such as "low server energy consumption but high cooling energy consumption," a cooling efficiency correction term 1 / Eff is used. cool,i This approach breaks away from the traditional isolated optimization mindset that separates IT systems from infrastructure. It uses a fractional structure of "1 ÷ heat dissipation efficiency" to mathematically map heat dissipation costs, making resource allocation more aligned with the overall energy consumption characteristics of the data center. Furthermore, since energy efficiency optimization in distributed systems must be based on performance and availability, simply pursuing energy efficiency can lead to performance degradation or increased failures. Therefore, it is closely linked to resource allocation priority (RAP). i The product of these factors ensures that energy efficiency optimization is carried out within the range of "performance meeting standards and fault controllable".
[0086] Based on the resource allocation priority of the output RAP i Based on this, it integrates dynamic power difference and dynamic heat dissipation efficiency Eff cool,i With the energy efficiency weighting coefficient δ driven by energy consumption targets, collaborative energy efficiency optimization of IT systems and infrastructure is achieved, among which resource allocation priority RAP i Ensure that energy efficiency optimization does not sacrifice load balancing and fault prevention goals, dynamic power difference (full load power P) active,i - Idle power P idle,i Quantify the dynamic energy consumption characteristics of nodes and prioritize resource allocation to nodes with good dynamic energy efficiency (Eff). cool,i By incorporating the data center cooling system into the resource allocation logic, heat dissipation energy consumption is reduced. The energy efficiency weight coefficient δ adjusts the priority of energy efficiency optimization according to the system energy consumption target, balancing the relationship between performance and energy efficiency. This solves the problem of the lack of coordination between the dynamic energy efficiency and the cooling system in traditional systems, and reduces the total energy consumption of the system.
[0087] The resource allocation system also includes a configuration allocation component, which dynamically allocates resources to servers based on the Energy Efficiency Optimization Factor (EOF). Specifically, it allocates the EOF of all nodes. i Normalization is performed to obtain the resource allocation ratio (Alloc) for each node. i , that is Alloc i Allocate CPU, memory, and network bandwidth.
[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A distributed server resource allocation system, characterized in that: include: Data acquisition component: used to acquire node load-related data, resource allocation-related data, and temperature-related data during server operation; Storage generation component: Used to input node load-related data, resource allocation-related data, and temperature-related data acquired by the data acquisition component, and to clean the input data before inputting the cleaned data into the allocation calculation component; Assignment calculation components: The node load resilience index is output based on CPU utilization, CPU performance degradation rate, memory utilization, memory performance degradation rate, network utilization, network performance degradation rate and weighting coefficients. The node load resilience index accurately quantifies the actual load capacity of the node. Based on the node load resilience index, fault prediction confidence, and fault prediction weight coefficient, resource allocation priority is output, and load conditions and fault prevention are collaboratively optimized through resource allocation priority. Based on resource allocation priority, full load power, idle power, dynamic heat dissipation efficiency, and energy efficiency weighting coefficient, an energy efficiency optimization factor is output, which provides the final decision indicator for resource allocation. Build allocation component: Dynamically allocate server resources based on energy efficiency optimization factors.
2. The distributed server resource allocation system according to claim 1, characterized in that: The allocation calculation component includes a load calculation module, a priority allocation module, and an energy efficiency optimization module.
3. The distributed server resource allocation system according to claim 2, characterized in that: The data acquisition process of the data acquisition component includes: CPU utilization is collected through CPU performance monitoring sensors on the server motherboard, CPU core temperature is recorded through CPU temperature sensors, cumulative power-on time of the CPU is recorded based on CPU runtime sensors, memory utilization is directly collected through memory controller sensors on the server motherboard, the number of memory ECC errors is recorded through memory error rate sensors, network utilization is directly collected through network traffic sensors on the server network card, and the number of data packets lost by the network card is recorded based on the network card packet loss rate sensor.
4. The distributed server resource allocation system according to claim 3, characterized in that: The data acquisition process of the data acquisition component also includes: The server power module's power sensor directly collects full-load power and idle power, the server's air inlet temperature is collected by the temperature sensor at the front of the server, the server's air outlet temperature is collected by the temperature sensor at the rear of the server, and the data center's ambient temperature is collected by the ambient temperature sensor in the server room.
5. The distributed server resource allocation system according to claim 4, characterized in that: The processing flow of the load calculation module is as follows: The CPU performance degradation rate is calculated based on the cumulative power-on time of the CPU and the benchmark performance test data. The memory performance degradation rate is calculated based on the number of ECC errors of the memory and the benchmark bandwidth test data. The network performance degradation rate is calculated based on the number of packet loss of the network card and the benchmark throughput test data. The weight coefficients are trained based on the historical fault data of the system. A1: Introduce quantitative decay of CPU utilization and CPU performance decay rate to make load assessment more in line with the actual hardware status. A2: By incorporating memory usage and memory performance degradation into load assessment, the dynamic performance degradation of memory is quantified, making the load assessment more closely reflect the actual hardware condition. A3: By incorporating network link performance degradation into load assessment through network utilization and network performance degradation rate, the degradation is quantified, making load assessment more accurate.
6. The distributed server resource allocation system according to claim 5, characterized in that: The processing flow of the priority partitioning module is as follows: The confidence level of fault prediction is obtained by training a Naive Bayes model based on CPU core temperature, memory ECC error count and network card packet loss count. B1: By dividing the node load resilience by the sum of the load resilience of all nodes, resources are ensured to be tilted towards truly idle nodes, avoiding resource waste and overload risks. B2: Introduce fault prediction confidence and fault prediction weight coefficient, and multiply them to bring fault prevention forward, ensuring that resources are tilted towards reliable nodes.
7. The distributed server resource allocation system according to claim 6, characterized in that: The processing flow of the energy efficiency optimization module is as follows: Dynamic heat dissipation efficiency is calculated by taking the server inlet temperature, server outlet temperature and data center ambient temperature, and energy efficiency weighting coefficient is determined based on the total energy consumption target of the system. C1: By multiplying with resource allocation priority, a balance between performance and energy efficiency is achieved in resource allocation; C2: The dynamic energy efficiency of a node is evaluated by the difference between full-load power and idle power, and then the ratio of the idle power to the full-load power is used to evaluate the dynamic energy efficiency, making energy efficiency optimization more in line with actual operating scenarios. C3: By adjusting the energy efficiency ratio of 1 to dynamic heat dissipation efficiency, resource allocation is made more in line with the overall energy consumption characteristics of the data center.
8. The distributed server resource allocation system according to claim 7, characterized in that: The processing flow for constructing and allocating components is as follows: The energy efficiency optimization factors of all nodes are normalized to obtain the resource allocation ratio of each node, and CPU, memory and network bandwidth are allocated according to the allocation ratio.