A multi-node server network topology and energy consumption-aware task scheduling method

Abstract

The application discloses a kind of multi-node server network topology and energy consumption perception's task scheduling method, belong to data center and high performance computing technical field.It includes: constructing the weighted directed graph topology model including server node and switch node, the dynamic energy consumption weight function of link is defined;Establish the three-dimensional energy consumption and thermal interference evaluation model of node, introduce space thermal interference factor;Analysis the characteristics of task to be scheduled;Based on energy consumption potential field screening candidate node;Build the global energy consumption cost function including computing energy consumption, network transmission energy consumption and thermal coupling penalty energy consumption;According to the principle of minimum total energy consumption, select target node to execute scheduling and update topology state.The application solves the problem of splitting computing and network energy consumption, ignoring the space coupling effect of heat dissipation in the prior art by the fusion of network topology perception and multi-dimensional energy consumption model, and realizes the optimization of global energy consumption of data center.

Description

Technical Field

[0001] This invention relates to the field of data center and high-performance computing technology, and in particular to a task scheduling method for multi-node server network topology and energy consumption awareness. Background Technology

[0002] With the rapid development of cloud computing and big data technologies, the scale of servers in data centers is growing rapidly, and energy consumption has become a key bottleneck restricting development. Electricity costs account for a large proportion of data center operating costs, including not only the energy consumption of server computing components (CPU, memory, etc.), but also the energy consumption of network equipment (switches, routers) and cooling systems.

[0003] Existing task scheduling algorithms, such as the Kubernetes default scheduler and Hadoop YARN, primarily focus on resource utilization (e.g., CPU and memory utilization) or simple load balancing. However, these existing technologies have the following significant drawbacks and shortcomings in practical applications: The one-sidedness of network topology awareness: Existing technologies typically only consider the "distance" of network topology (e.g., determining whether they are located in the same rack or connected to the same switch) to reduce network latency. This approach lacks a quantitative consideration of the dynamic power consumption of switches caused by network link congestion. In reality, the power consumption of network devices has a non-linear relationship with traffic load. Ignoring this leads to a surge in implicit power consumption at the network layer; that is, although tasks are assigned to nodes with sufficient resources, congestion on the network transmission path results in extremely high additional power consumption.

[0004] The disconnect between computing and network energy consumption: Existing technologies often optimize "node computing energy consumption" and "network transmission energy consumption" separately. For example, some strategies tend to schedule tasks to low-power nodes to save computing energy. However, this practice may lead to a large amount of remote data transmission, and the resulting network energy consumption may far exceed the computing energy saved by the nodes, resulting in an increase in global energy consumption and a phenomenon of "saving computing power but wasting network power."

[0005] Ignoring the spatial coupling effect of heat dissipation: Existing technologies primarily focus on temperature control of single nodes to prevent overheating. However, in multi-node servers (such as blade servers or high-density racks), the physical layout of nodes exhibits strong spatial correlation. For example, in rack-mount servers, the high power consumption and heat generated by upstream nodes (inlet side) can cause exhaust hot air to flow to downstream nodes (outlet side), leading to an increase in the ambient temperature of downstream nodes. This temperature increase forces downstream node fan speeds to increase, and may even increase the cooling load on the data center's air conditioning, resulting in a non-linear increase in cooling energy consumption. Existing scheduling algorithms do not incorporate this "thermal interference conduction loss" into the scheduling decision model, resulting in low overall cooling efficiency.

[0006] In summary, there is an urgent need for a task scheduling method that can comprehensively consider the dynamic energy consumption of network topology, the computing energy consumption of nodes, and the spatial thermal interference effect in order to minimize the global energy consumption of the data center. Summary of the Invention

[0007] The purpose of this invention is to provide a task scheduling method based on multi-node server network topology and energy consumption awareness. It aims to solve the problems of single energy consumption optimization dimension and neglect of thermal spatial coupling effect in the prior art by constructing a joint optimization model that integrates network path energy consumption, node computing energy consumption and spatial thermal interference cost, thereby minimizing the global energy consumption of the data center.

[0008] To achieve the above objectives, this invention provides a task scheduling method for multi-node server network topology and energy consumption awareness, comprising the following steps: S1. Construct a multi-dimensional energy consumption-aware network topology model, abstracting the physical cluster into a weighted directed graph. , where the set of nodes Includes server nodes and switch nodes, edge set Represents the physical link; for each link Define dynamic energy consumption weight function The dynamic energy consumption weighting function includes the static transmission power consumption of the link, the dynamic power consumption increment caused by the current load rate of the link, and the activation energy consumption coefficient of the switch ports at both ends of the link. This allows the weights in the topology graph to reflect network congestion and energy consumption status in real time.

[0009] S2. Establish a three-dimensional energy consumption and thermal interference evaluation model for each server node. Establish energy consumption state vector ;in, To estimate energy consumption, This refers to static power consumption. The spatial thermal interference factor is defined as the heat transfer coefficient to downstream nodes based on the physical location of the node in the rack. The heat transfer coefficient to downstream nodes is defined based on the node's physical location (U position) within the rack. If the node... Located on the air inlet side, its high power consumption operation will cause the exhaust hot air to flow to downstream nodes, thereby increasing the fan energy consumption and cooling energy consumption of downstream nodes.

[0010] S3. Task Feature Profiling and Requirement Analysis: Analyzing the Tasks to be Scheduled Extracting computational load Data transmission volume and the set of data block locations it depends on This step clarifies the resource requirements and data dependencies of the task.

[0011] S4. Candidate node selection based on energy consumption potential field: In the topology graph, with the data source location as the center, calculate the energy consumption potential of each node. Nodes with potential energy values ​​below a preset threshold are selected as the candidate node set. ; S5. Construct a global energy consumption cost function and make scheduling decisions for the set. For each node in the process, calculate the total energy cost. Total energy consumption cost This includes the direct computational energy consumption generated during task execution. Total energy consumption generated by data transmission via network path and thermal coupling penalty energy consumption ; S6. Perform scheduling and dynamic topology state update, and select the total energy cost. The smallest node is designated as the target node for issuing tasks, and the link load status and node energy consumption status vectors in the topology graph are updated in real time.

[0012] Preferably, in step S2, the space thermal interference factor The quantification process is as follows: Assuming the rack uses vertical air cooling with cool air flowing from bottom to top, define the thermal interference radius of node $n_i$. thermal interference radius For the node The set of downstream nodes affected by exhaust emissions; Thermal coupling penalty energy consumption The calculation formula is: ; in, Represents a node The increased air outlet temperature due to workload caused downstream nodes to... The increase in fan power consumption to maintain its own temperature The estimated execution time for the task.

[0013] Preferably, downstream nodes Increased fan power consumption Estimated using a heat conduction model: node The air outlet temperature rises due to workload. According to the heat conduction formula Estimate the heat transferred to downstream nodes, and then calculate the downstream nodes. The increased power consumption required to offset this heat.

[0014] Preferably, in step S4, the energy consumption potential energy The calculation formula is: ; in, Represents a node hop count to the data source location This represents the cumulative dynamic energy consumption weight along the path. and These are the weighting coefficients.

[0015] Preferably, in step S5, network transmission energy consumption The calculation formula is: ; in, For data transmission volume, For the basic power consumption of the link, The dynamic power consumption per unit flow rate. The current link load rate, This represents the internal processing power consumption of the switches along the path.

[0016] Preferably, in step S5, energy consumption is directly calculated. The CPU frequency-voltage-power curve was calculated based on historical load data.

[0017] Preferably, in step S5, the thermal coupling penalty energy consumption This is used to achieve physical-level thermal avoidance, so that computationally intensive tasks are prioritized and scheduled to the rack's exhaust side to prevent hot air from blowing to downstream nodes.

[0018] Preferably, the method further includes the step of using an improved ant colony algorithm to solve for the optimal target node: Total energy consumption cost The reciprocal of the value is used as heuristic information to guide the ant colony's search; Once the task is completed, the energy pheromone on the path is updated based on the actual measured total energy consumption value. The lower the energy consumption of a path or node, the higher the pheromone concentration.

[0019] Preferably, before step S1, a step of collecting topology and energy consumption data is also included: Deploy a lightweight agent on each server node in the cluster to collect CPU utilization, fan speed, inlet temperature, and outlet temperature. On the switch side, port traffic counters and total switch power consumption are collected via SNMP or OpenFlow protocols. The dispatch center summarizes the above information and constructs a real-time topology status table.

[0020] Preferably, in step S6, the scheduling process includes: The scheduler encapsulates tasks into containers or virtual machine images and sends them to the selected target nodes; If a task involves a large amount of data transmission, the scheduler sends flow tables to the switches through the SDN controller to reserve bandwidth resources.

[0021] Therefore, the multi-node server network topology and energy-aware task scheduling method of the present invention, which adopts the above structure, has the following beneficial effects: (1) This invention breaks through the limitations of traditional optimization of only a single dimension, incorporates network device energy consumption into task scheduling considerations, avoids the phenomenon of "saving computing electricity costs but wasting network electricity costs", and realizes joint optimization of computing and network energy consumption.

[0022] (2) This invention introduces a “spatial thermal interference factor” at the scheduling level, which effectively reduces the cooling energy consumption of the entire rack by prioritizing the scheduling of low-power tasks on the air inlet side (upstream of the cold air) or avoiding overheating of upstream nodes. This is a micro-innovation that is not easily thought of by those skilled in the art, and it has significant technical effects.

[0023] (3) The present invention can sense the network congestion status, dynamically adjust the link weight, avoid the surge in network device energy consumption caused by traffic congestion, and realize peak shaving and valley filling of network energy consumption.

[0024] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a task scheduling method for multi-node server network topology and energy consumption awareness according to the present invention. Figure 2 This is a schematic diagram of thermal interference in a multi-node server rack, illustrating a task scheduling method for multi-node server network topology and energy consumption awareness according to the present invention. Detailed Implementation

[0026] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0028] Example 1 like Figure 1 As shown, this invention provides a task scheduling method for multi-node server network topology and energy consumption awareness, comprising the following steps: Step 1: Topology and Energy Consumption Data Acquisition To obtain accurate energy consumption model parameters, data collection is required within the cluster. A lightweight agent is deployed on each server node to collect data such as CPU utilization, fan speed, inlet temperature, and outlet temperature. On the switch side, if the switch supports SNMP or OpenFlow protocols, port traffic counters and total switch power consumption are collected. The scheduling center aggregates this information to construct a real-time topology status table, which serves as the basis for subsequent calculations.

[0029] Step 2: Construct a multi-dimensional energy consumption-aware network topology model: The system abstracts the physical cluster as a weighted directed graph. The set of nodes Includes server nodes and switch nodes, edge set Represents the physical link.

[0030] For each link Define the dynamic energy consumption weight function. The design of this function not only considers static transmission power consumption, but more importantly, it introduces a dynamic factor: ; in, For the basic power consumption of the link, This addresses the dynamic power consumption increment caused by link load rate. It solves the problem of existing technologies neglecting the dynamic energy consumption caused by network congestion.

[0031] Step 3: Quantitative calculation of thermal interference factor: like Figure 2 As shown, assuming the data center rack uses vertical air cooling, the cool air flows from bottom to top (i.e., the lower part of the rack is the air inlet side and the upper part is the air outlet side, or as defined by the specific airflow layout of the data center).

[0032] Define nodes thermal interference radius That is, the set of downstream nodes affected by its exhaust.

[0033] For nodes Its thermal coupling energy consumption penalty term The calculation of this value is the key innovation of this invention. The formula is as follows: ; in, Represents a node The air outlet temperature rises due to workload. Subsequently, this led to downstream nodes The increase in fan power consumption to maintain its own temperature. The estimation can be achieved through a heat conduction model. In progress, among which... For heat, The specific heat capacity of air, Air mass flow rate. The estimated execution time for the task.

[0034] Step S3 causes the scheduler to avoid hot air blowing onto a large number of downstream nodes if it schedules high-computation-intensive tasks to the bottom of the rack (air inlet side, upstream), resulting in huge energy consumption due to thermal coupling penalties. Therefore, the scheduler tends to schedule high-power tasks to the top of the rack (air outlet side, downstream) or restrict high-power tasks on upstream nodes, thereby avoiding interference from hot air to a large number of downstream nodes and achieving "thermal avoidance" at the physical level.

[0035] Step 4: Refined calculation of network path energy consumption: Traditional methods only calculate hop count, ignoring path load differences. This invention defines path energy consumption. for: ; in, This refers to the amount of data transmitted. This represents the basic power consumption of the link. Dynamic power consumption per unit flow rate; This represents the current link load rate. Energy consumption for a switch to process data packets.

[0036] Even if two paths have the same number of hops, the scheduler will choose the path with lower load (lower energy consumption) to avoid network hotspots and achieve peak shaving and valley filling of network energy consumption.

[0037] Step 5: Solving based on the improved ant colony algorithm: Since finding the optimal node is an NP-hard problem, this embodiment uses an improved ant colony algorithm to solve it.

[0038] Traditional ant colony algorithms only consider distance; this invention addresses the aforementioned total energy consumption cost. The reciprocal as heuristic information ,Right now This means that nodes with lower energy consumption are more attractive to "ants".

[0039] Once the task is completed, the "energy pheromone" on the path is updated based on the actual measured total energy consumption. The lower the energy consumption of a path / node, the higher the pheromone concentration, and the greater the probability that subsequent tasks will be scheduled to that location.

[0040] Step Six: Task Feature Profiling and Candidate Node Selection: Analyze the tasks to be scheduled computational load Data transmission volume and data block location .

[0041] Calculate the energy consumption potential of each node. Set a threshold Filter out The nodes are selected as the candidate set. This step quickly eliminates nodes with excessively high network transmission costs, thus narrowing the search space.

[0042] Step 7: Scheduling Decision and Execution Comprehensive calculation of candidate nodes Select the node corresponding to the minimum value as the target node.

[0043] The scheduler encapsulates tasks into containers or virtual machine images and sends them to the selected target nodes. If a task involves a large amount of data transmission, the scheduler simultaneously sends flow tables to the switches through the SDN controller, reserving bandwidth resources to ensure that network energy consumption is under control.

[0044] Finally, update the global topology graph state and wait for the next task's scheduling request.

[0045] Through the above embodiments, the present invention effectively solves the problem of the separation of computing energy consumption, network energy consumption and thermal energy consumption in data center task scheduling, and achieves in-depth optimization of global energy consumption.

[0046] Example 2 Taking a data center server room that uses a standard 42U high-density rack and has multiple blade servers deployed inside as an example, the cooling method adopts the traditional "bottom air intake, top air exhaust" vertical air cooling mode, that is, cold air enters from the bottom of the rack, is heated by the server, and is exhausted from the top of the rack.

[0047] Physical resources: Node A ( Located at the bottom of the cabinet (position 5U), on the air inlet side, with a low ambient temperature (25℃), it is an upstream node.

[0048] Node B ( Located at the top of the cabinet (position 40U), on the air outlet side, the ambient temperature is relatively high (30℃). It is a downstream node, and its air intake is affected by the exhaust of node A.

[0049] Data storage node: Located in another rack, connected to node A and node B via a core switch, with the same basic hop count for the network link.

[0050] Task to be scheduled: A deep learning model training task .

[0051] Task characteristics: Computationally intensive task, estimated execution time Hours; requires reading a large amount of training data from data storage nodes, data transfer volume Significantly high; the CPU is expected to be fully loaded during task execution, resulting in an increase in node power consumption of approximately 200W.

[0052] The scheduling decision-making process adopts a multi-node server network topology and energy consumption-aware task scheduling method provided by this invention, as detailed below: Step 1: Energy Consumption Data Acquisition and Topology Construction The dispatch center, through agents deployed on the server, detected that the current CPU utilization of nodes A and B is low, indicating sufficient resources. Simultaneously, it detected that the current network link load is low, and the network topology model has been completed.

[0053] Step 2: Calculate energy consumption directly Based on the power consumption curve fitted from historical load data, regardless of whether the task is scheduled to node A or node B, the direct computational energy consumption increment is basically the same because the task's computational load is fixed. ; Step 3: Calculate network transmission power consumption Since the number of network path hops from node A and node B to the data storage node is the same, and the current link load is low with no congestion, the difference in network transmission energy consumption between the two is minimal and can be approximated as equal. ; Step 4: Calculate the energy consumption of thermal coupling penalty If node A (upstream node) is selected: Node A is located at the bottom of the rack, and the high power consumption generated by its operations causes a significant increase in the outlet temperature. The hot air flows upward with the airflow and blows directly onto Node B, located at the top of the rack, and other downstream nodes.

[0054] Assuming this causes an increase in the inlet temperature of node B, in order to maintain the normal operating temperature of node B, the fan speed of node B will be forced to increase. Estimate the increased fan power consumption. It is 50W.

[0055] According to the formula: ; If node B (downstream node) is selected: Node B is located at the top of the rack. The hot air generated by its operation is directly discharged into the return air duct of the computer room and does not pass through other server nodes, so it will not cause thermal interference to other nodes.

[0056] According to the formula: ; Step 5: Comparison and Decision of Total Energy Cost $Cost_{total}$ Total energy consumption cost of node A: ; Total energy consumption cost of node B: ; Decision result: After calculation, Although nodes A and B are nearly identical in terms of computational and network energy consumption, the scheduler recognizes that placing high-power tasks on upstream node A would incur additional cooling energy consumption due to the thermal coupling penalty mechanism introduced in this invention. Therefore, the scheduler ultimately selects node B as the target node for task distribution.

[0057] Technical effect verification: As can be seen from the above application examples, the present invention successfully achieves "thermal avoidance" at the physical level by introducing a spatial thermal interference factor, scheduling high-heat tasks to the downstream of the airflow, avoiding the thermal cascading effect, and implicitly reducing the overall cooling cost of the data center without sacrificing computing performance, thus verifying the effectiveness of the global energy consumption optimization method of the present invention.

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A task scheduling method based on multi-node server network topology and energy consumption awareness, characterized in that, Includes the following steps: S1. Construct a multi-dimensional energy consumption-aware network topology graph model, abstracting the physical cluster into a weighted directed graph. , where the set of nodes Includes server nodes and switch nodes, edge set Represents the physical link; for each link Define dynamic energy consumption weight function The dynamic energy consumption weighting function includes the static transmission power consumption of the link, the dynamic power consumption increment caused by the current load rate of the link, and the activation energy consumption coefficient of the switch ports at both ends of the link. S2. Establish a three-dimensional energy consumption and thermal interference evaluation model for each server node. Establish energy consumption state vector ;in, To estimate energy consumption, This refers to static power consumption. The spatial thermal interference factor is defined as the heat transfer coefficient from the node to the downstream node based on the node's physical location in the rack. S3. Task Feature Profiling and Requirement Analysis: Analyzing the Tasks to be Scheduled Extracting computational load Data transmission volume and the set of data block locations it depends on ; S4. Candidate node selection based on energy consumption potential field: In the topology graph, with the data source location as the center, calculate the energy consumption potential of each node. Nodes with potential energy values ​​below a preset threshold are selected as the candidate node set. ; S5. Construct a global energy consumption cost function and make scheduling decisions for the set. For each node in the process, calculate the total energy cost. Total energy consumption cost This includes the direct computational energy consumption generated during task execution. Total energy consumption generated by data transmission via network path and thermal coupling penalty energy consumption ; S6. Perform scheduling and dynamic topology state update, and select the total energy cost. The smallest node is designated as the target node for issuing tasks, and the link load status and node energy consumption status vectors in the topology graph are updated in real time.

2. The task scheduling method for multi-node server network topology and energy consumption awareness according to claim 1, characterized in that, In step S2, the space thermal interference factor The quantification process is as follows: Assuming the rack uses vertical air cooling with cool air flowing from bottom to top, define the thermal interference radius of node $n_i$. thermal interference radius For the node The set of downstream nodes affected by exhaust emissions; Thermal coupling penalty energy consumption The calculation formula is: ； in, Represents a node The increased air outlet temperature due to workload caused downstream nodes to... The increase in fan power consumption to maintain its own temperature The estimated execution time for the task.

3. The task scheduling method for multi-node server network topology and energy consumption awareness according to claim 2, characterized in that, Downstream nodes Increased fan power consumption Estimated using a heat conduction model: node The air outlet temperature rises due to workload. According to the heat conduction formula Estimate the heat transferred to downstream nodes, and then calculate the downstream nodes. The increased power consumption required to offset this heat.

4. The task scheduling method for multi-node server network topology and energy consumption awareness according to claim 1, characterized in that, In step S4, energy is consumed. The calculation formula is: ； in, Represents a node hop count to the data source location This represents the cumulative dynamic energy consumption weight along the path. and These are the weighting coefficients.

5. The task scheduling method for multi-node server network topology and energy consumption awareness according to claim 1, characterized in that, In step S5, network transmission energy consumption The calculation formula is: ； in, For data transmission volume, For the basic power consumption of the link, The dynamic power consumption per unit flow rate. The current link load rate, This represents the internal processing power consumption of the switches along the path.

6. The task scheduling method for multi-node server network topology and energy consumption awareness according to claim 1, characterized in that, In step S5, energy consumption is calculated directly. The CPU frequency-voltage-power curve was calculated based on historical load data.

7. The task scheduling method for multi-node server network topology and energy consumption awareness according to claim 1, characterized in that, In step S5, thermal coupling penalty energy consumption This is used to achieve physical-level thermal avoidance, so that computationally intensive tasks are prioritized and scheduled to the rack's exhaust side to prevent hot air from blowing to downstream nodes.

8. The task scheduling method for multi-node server network topology and energy consumption awareness according to claim 1, characterized in that, The method also includes the step of using an improved ant colony algorithm to find the optimal target node: Total energy consumption cost The reciprocal of the value is used as a heuristic to guide the ant colony's search. Once the task is completed, the energy pheromone on the path is updated based on the actual measured total energy consumption value. The lower the energy consumption of a path or node, the higher the pheromone concentration.

9. The task scheduling method for multi-node server network topology and energy consumption awareness according to claim 1, characterized in that, Before step S1, there is also a step of collecting topology and energy consumption data: Deploy a lightweight agent on each server node in the cluster to collect CPU utilization, fan speed, inlet temperature, and outlet temperature. On the switch side, port traffic counters and total switch power consumption are collected via SNMP or OpenFlow protocols. The dispatch center summarizes the above information and constructs a real-time topology status table.

10. A task scheduling method for multi-node server network topology and energy consumption awareness according to claim 1, characterized in that, In step S6, the scheduling process includes: The scheduler encapsulates tasks into containers or virtual machine images and sends them to the selected target nodes; If a task involves a large amount of data transmission, the scheduler sends flow tables to the switches through the SDN controller to reserve bandwidth resources.

Images