AI resource scheduling method and device applied to distributed architecture and electronic equipment
By comprehensively extracting basic node health features and hardware thermal stress features, and combining them with task queue statistical features, the scheduling weights are dynamically adjusted and a probabilistic random allocation strategy is adopted. This solves the problems of node overheating, task congestion, and hardware aging in the distributed AI resource scheduling system, thereby improving the system's real-time performance and reliability.
Patent Information
- Application Number
- CN202610807431.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-05
- Publication Date
- 2026-08-25
AI Technical Summary
In the distributed AI resource scheduling system of smart factories, edge computing nodes often face problems such as overheating, task congestion and hardware aging, resulting in decreased computing power and processing delays. Existing monitoring methods are difficult to capture slow performance degradation or transient risks of thermal shock, and there is a lack of a joint scheduling weight mechanism between task waiting time and node health.
By extracting basic node health characteristics (temperature deviation, performance baseline, scheduling congestion) and hardware thermal stress characteristics, and combining them with task queue statistical characteristics, the node health and scheduling urgency are calculated. The scheduling weights are dynamically adjusted and a probabilistic random allocation strategy is adopted to generate a scheduling strategy.
It enables accurate health assessment and dynamic task allocation for edge computing nodes, reduces node overheating and frequency reduction events and task timeout rates, and improves the real-time performance of AI inference on the production line and the long-term reliability of the system.
Smart Images
Figure CN122633346A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer technology, and in particular to an AI resource scheduling method, apparatus, and electronic device applied to distributed architectures. Background Technology
[0002] In distributed AI resource scheduling systems in smart factories, edge computing nodes often face problems such as overheating, task congestion, and hardware aging, leading to decreased computing power and processing delays. Existing monitoring methods mostly use fixed threshold alarms, which are difficult to detect slow performance degradation or transient risks from thermal shocks. Traditional scheduling algorithms, such as round-robin, least connections, or those based solely on the current queue length, ignore the historical trends of node health and the cumulative effects of hardware stress.
[0003] Furthermore, task waiting time and node health are typically handled separately, lacking a joint scheduling weighting mechanism. Therefore, there is an urgent need for a real-time monitoring and resource scheduling method that can comprehensively consider temperature deviation, performance baseline, scheduling congestion, and thermal stress characteristics, and dynamically adjust scheduling priorities. Summary of the Invention
[0004] The purpose of this invention is to provide an AI resource scheduling method, apparatus, and electronic device applied to a distributed architecture, so as to solve at least one of the problems existing in the prior art.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] An AI resource scheduling method applied to a distributed architecture includes:
[0007] Based on CPU temperature, utilization, and task queue, basic node health characteristics are extracted, including temperature deviation characteristics, performance baseline deviation characteristics, and scheduling congestion characteristics.
[0008] Hardware thermal stress characteristics are extracted based on node operating current and core voltage.
[0009] The node health is determined based on the node's basic health characteristics and hardware thermal stress characteristics. The node health is negatively correlated with the average value of hardware thermal stress characteristics, temperature deviation characteristics, performance baseline deviation characteristics, and scheduling congestion characteristics in each real-time cycle.
[0010] Calculate the node scheduling urgency based on the statistical characteristics of the task queue;
[0011] Dynamic scheduling weights are determined based on node health and scheduling urgency, and then a scheduling strategy is generated.
[0012] Furthermore, the extraction process of the temperature deviation feature is as follows:
[0013] The average and maximum values T of the CPU temperature sequence within the evaluation window are calculated. Then, the temperature ratio factor and temperature peak penalty are calculated. Based on the temperature ratio factor and temperature peak penalty, the temperature deviation characteristics are determined.
[0014] The temperature ratio factor is calculated using the average value of the CPU temperature sequence, the upper limit of the normal operating temperature of the node, and the stable idle temperature of the node. The temperature peak penalty is calculated using the maximum value of the CPU temperature sequence, the upper limit of the normal operating temperature of the node, and the critical temperature of the node. The temperature deviation feature is a weighted fusion value of the temperature ratio factor and the temperature peak penalty.
[0015] Furthermore, the extraction process for the performance baseline deviation feature is as follows:
[0016] Calculate the average CPU utilization and average memory usage within the evaluation window, and calculate the CPU deviation and memory stress factor to determine the performance baseline deviation characteristics.
[0017] The CPU deviation is calculated by the average CPU utilization rate and the idle CPU utilization rate measured during node initialization. The memory pressure factor is positively correlated with the average memory occupancy rate. The performance baseline deviation feature is a weighted fusion value of the CPU deviation and the memory pressure factor.
[0018] Furthermore, the analysis process for the scheduling congestion characteristics is as follows:
[0019] Calculate the standard deviation and average queue length of the task queue within the evaluation window, and use the ratio of the standard deviation and average queue length of the task queue within the evaluation window as the queue fluctuation factor;
[0020] Calculate the average waiting time of the oldest task, and then determine the waiting time factor, which is positively correlated with the average waiting time of the oldest task;
[0021] The number of overload events within the statistical evaluation window is calculated, and the determination of the overload penalty factor is positively correlated with the number of overload events.
[0022] Furthermore, an overload event is defined as follows: when the queue length rises from below the queue length overload threshold to above the queue length overload threshold, and remains above the queue length overload threshold for five consecutive sampling points, it is counted as an overload event; only when the queue length drops below the queue length overload threshold and meets the above condition again is it counted as a new overload event.
[0023] The queue fluctuation factor, waiting time factor, and overload penalty factor are weighted and fused to determine the scheduling congestion characteristics.
[0024] Furthermore, within each real-time cycle, the node operating current In and core voltage Vc are processed:
[0025] The instantaneous power consumption of a computing node is Pn = In × Vc;
[0026] Calculate the power change ΔP = |Pn(t) - Pn(t-1)|, where t-1 represents the previous sampling time.
[0027] Calculate the hardware thermal stress characteristic Fst, Fst=min(1,ΔP / ΔPmax), where ΔPmax is the preset power change benchmark.
[0028] Furthermore, within an evaluation window, the average waiting time Wa of all tasks passing through the node within the evaluation window is extracted to determine the node scheduling urgency, which is positively correlated with Wa.
[0029] Furthermore, at the end of each evaluation window, the dynamic scheduling weight Wdn is calculated for all available nodes in the cluster:
[0030] The dynamic scheduling weight Wdn is determined based on the node health Hn and the scheduling urgency Un, and Wdn = Hn × (1 - Un);
[0031] During the next evaluation window, when new tasks arrive, the scheduler allocates tasks based on the dynamically calculated scheduling weights of each node:
[0032] The dynamic scheduling weight of each node is divided by the sum of the weights of all nodes in the cluster to obtain the probability value of that node being selected. Then, the scheduler randomly selects a node according to this probability distribution and assigns the new task to that node.
[0033] According to another aspect of this application, an AI resource scheduling device for a distributed architecture is provided, comprising:
[0034] Extract the health base unit, which is used to extract the basic health characteristics of the node based on CPU temperature, utilization, and task queue;
[0035] The thermal stress extraction unit is used to extract hardware thermal stress characteristics based on node operating current and core voltage;
[0036] Determine the health unit, which is used to determine the health of a node based on its basic health characteristics and hardware thermal stress characteristics;
[0037] The urgency calculation unit is used to calculate the scheduling urgency of nodes based on the statistical characteristics of the task queue.
[0038] A scheduling unit is generated to determine dynamic scheduling weights based on node health and scheduling urgency, and then generate a scheduling strategy.
[0039] According to another aspect of this application, an electronic device is provided, the electronic device comprising:
[0040] One or more processors;
[0041] Storage device for storing one or more programs;
[0042] When the one or more programs are executed by the one or more processors, the one or more processors implement the AI resource scheduling method applied to a distributed architecture.
[0043] The beneficial effects of this invention are as follows: This invention achieves accurate health assessment and dynamic task allocation for edge computing nodes in smart factories through multi-dimensional feature fusion and two-stage weighted scheduling. At the health assessment level, it innovatively introduces performance deviation features based on the idle baseline, congestion features based on queue fluctuations and overload events, and hardware thermal stress features based on power changes, compensating for the shortcomings of single temperature monitoring in responding to thermal shocks and gradual performance changes. At the scheduling decision level, node health and task scheduling urgency are fused into dynamic scheduling weights using a multiplicative formula, and a probabilistic random allocation strategy is adopted to avoid assigning tasks to nodes about to overheat or to nodes already heavily backed up. Practice shows that this method can significantly reduce node overheating and frequency reduction events and task timeout rates, greatly improving the real-time performance of AI inference on the production line and the long-term reliability of the system. Attached Figure Description
[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0045] Figure 1 This is a flowchart illustrating the AI resource scheduling method applied to a distributed architecture in this embodiment.
[0046] Figure 2 This is a flowchart illustrating the scheduling strategy generation method in this embodiment.
[0047] Figure 3 This is a schematic diagram of the AI resource scheduling device applied to a distributed architecture in this embodiment.
[0048] Figure 4 This is a schematic diagram of the electronic device in this embodiment. Detailed Implementation
[0049] To more clearly illustrate the present invention, the following description, in conjunction with preferred embodiments and accompanying drawings, further explains the invention. Similar components in the drawings are indicated by the same reference numerals. Those skilled in the art should understand that the specific description below is illustrative rather than restrictive and should not be construed as limiting the scope of protection of the present invention.
[0050] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0051] Specifically, this embodiment is applied to a distributed AI resource scheduling system. This system is deployed in a smart factory to monitor and dynamically schedule edge computing nodes distributed throughout the production line in real time, aiming to prevent the decline in computing power and processing delays caused by node overheating, task congestion, or hardware aging.
[0052] Please see Figure 1 As shown, this is a flowchart illustrating the AI resource scheduling method applied to a distributed architecture in this embodiment. Before the method is executed, the system synchronously collects data through monitoring modules deployed on each edge computing node and task queue, including:
[0053] The data includes:
[0054] Node performance data: CPU temperature, CPU utilization, and memory usage are collected in real time through the hardware monitoring interface built into each edge computing node, forming a time series.
[0055] Task load data: The number of tasks waiting to be processed in the current node's waiting queue, the waiting time of the oldest task in the queue, and the waiting time of all tasks in the queue are read through the internal counter of the task scheduler.
[0056] Power consumption and current data: Read the real-time operating current and core voltage of the node through the current sampling resistor on the node power supply line;
[0057] This embodiment does not specifically limit the communication protocol and transmission method of the above data. Those skilled in the art can freely set it according to the on-site network conditions, for example, using the MQTT protocol to upload to the central dispatch server via an industrial wireless network. The method uses a 30-second real-time cycle and a 10-minute evaluation window.
[0058] The method includes:
[0059] Step S1: Extract basic health characteristics of nodes based on CPU temperature, usage, and task queue.
[0060] Specifically, at the end of each evaluation window, based on CPU temperature, utilization, and task queue data collected within the evaluation window, temperature deviation features, performance baseline deviation features, and scheduling congestion features are extracted.
[0061] The analysis process for the temperature deviation characteristics is as follows:
[0062] Calculate the average value Ta and maximum value Tmax of the CPU temperature sequence within the evaluation window, then calculate the temperature ratio factor Rt and the temperature peak penalty Pp, and determine the temperature deviation characteristic Ft based on the temperature ratio factor Rt and the temperature peak penalty, where:
[0063] Rt=max(0,(Ta-Ti) / (Tn-Ti));
[0064] Pp=min(1,max(0,(Tmax-Tn) / (Tc-Tn)));
[0065] Ft = min(1, 0.7 × Rt + 0.3 × Pp);
[0066] Where Tn is the upper limit of the normal operating temperature of the node, Tc is the critical temperature of the node, and Ti is the stable temperature of the node under no-load conditions.
[0067] The analysis process for the performance baseline deviation characteristics is as follows:
[0068] Calculate the average CPU utilization (Uavg) and average memory usage (Mavg) within the evaluation window, and calculate the CPU deviation (Dc) and memory stress factor (Dm) to determine the performance baseline deviation characteristic (Fp), where:
[0069] Dc=min(1,max(0,(Uavg-Ub) / (max(0.3×Ub),0.05)));
[0070] Dm=min(1,max(0,(Mavg-0.5) / 0.4));
[0071] Fp = 0.8 × Dc + 0.2 × Dm;
[0072] Where Ub is the idle CPU utilization rate measured when the node is initialized;
[0073] The analysis process for the scheduling congestion characteristics is as follows:
[0074] Calculate the standard deviation of the task queue length σQ and the average queue length Qavg within the evaluation window, and then determine the queue fluctuation factor VP, VP=σQ / Qavg;
[0075] Calculate the average waiting time Wavg of the oldest task, and then determine the waiting time factor Wf, Wf=min(1,Wavg / Wth1), where Wth1 is the first preset waiting time threshold;
[0076] The number of overload events Nv within the statistical evaluation window is used to determine the overload penalty factor Op, where Op = min(1, Nv / N0), and N0 is the preset number of events. An overload event is defined as follows: when the queue length rises from below Qth to ≥Qth, and remains ≥Q_th for the next 5 consecutive sampling points, it is counted as an overload event. A new overload event is only counted when the queue length drops to <Qth and meets the above condition again. Qth is the queue length overload threshold.
[0077] The scheduling congestion characteristic Fc is determined based on the queue fluctuation factor VP, the waiting time factor Wf, and the overload penalty factor Op, where Fc = min(1, 0.4 × Vq + 0.4 × Wf + 0.2 × Op).
[0078] Preferably, in this embodiment, the upper limit of the normal operating temperature of the node is 75℃, the critical temperature of the node is 95℃, the stable idle temperature of the node is ambient temperature +15℃, the first preset waiting time threshold is 500ms, the preset event is 5 items, and the queue length overload threshold is 10 tasks.
[0079] Specifically, by performing multi-dimensional statistics on CPU temperature, utilization, memory usage, and task queues within an evaluation window, temperature deviation features, performance baseline deviation features, and scheduling congestion features are extracted. Through weighted fusion of temperature ratio factor and peak penalty, performance deviation calculation based on idle baseline, and joint analysis of queue fluctuations and overload events, the sensitivity of identifying early performance degradation and congestion precursors of nodes is significantly improved, providing structured and comparable input for subsequent health assessments.
[0080] Please continue reading. Figure 1 As shown, the AI resource scheduling method applied to a distributed architecture further includes:
[0081] Step S2: Extract hardware thermal stress characteristics based on node operating current and core voltage.
[0082] Specifically, within each real-time cycle, the node operating current In and core voltage Vc are processed:
[0083] The instantaneous power consumption of a computing node is Pn = In × Vc;
[0084] Calculate the power change ΔP = |Pn(t) - Pn(t-1)|, where t-1 represents the previous sampling time.
[0085] Calculate the hardware thermal stress characteristic Fst, Fst=min(1,ΔP / ΔPmax), where ΔPmax is the preset power change benchmark.
[0086] Preferably, in this embodiment, the preset power variation benchmark is 0.2×Prated, where Prated is the rated thermal design power of the node.
[0087] Specifically, within each real-time cycle, instantaneous power consumption and power change are calculated using the node's real-time operating current and core voltage, thereby obtaining hardware thermal stress characteristics. These characteristics directly reflect the degree of thermal cycling impact of power fluctuations on chip packaging and solder joints. Compared to monitoring only the absolute temperature value, the power change rate can capture thermal transient stress caused by sudden task changes earlier, preventing accelerated aging due to thermal fatigue. By comparing and normalizing the power change with a preset power change benchmark, this characteristic can be integrated with basic health characteristics, filling the gap in traditional temperature monitoring's lag in responding to thermal shocks.
[0088] Please continue reading. Figure 1 As shown, the AI resource scheduling method applied to a distributed architecture further includes:
[0089] Step S3: Determine the node health based on the node's basic health characteristics and hardware thermal stress characteristics.
[0090] Specifically, at the end of each evaluation window, the average value Fsta of the hardware thermal stress characteristics for each real-time cycle is calculated, and the node health Hn is calculated in combination with the node health basic characteristics:
[0091] Hn=1-(w1×Ft+w2×Fp+w3×Fc+w4×Fsta);
[0092] Where w1 is the temperature deviation weight, w2 is the performance deviation weight, w3 is the scheduling congestion weight, w4 is the thermal stress weight, and w1+w2+w3+w4=1.
[0093] Preferably, in this embodiment, the weighting for temperature deviation is 0.3, the weighting for performance deviation is 0.2, the weighting for scheduling congestion is 0.3, and the weighting for thermal stress is 0.2.
[0094] Specifically, at the end of the evaluation window, the average value of the hardware thermal stress characteristics of each real-time cycle is calculated and weighted and fused with the obtained temperature deviation characteristics, performance deviation characteristics, and scheduling congestion characteristics to obtain the node health score. This health score can comprehensively reflect the coupled effects of multiple factors such as hardware aging, thermal stress, and task congestion, providing a single, intuitive, and comparable node status score for dynamic scheduling.
[0095] Please continue reading. Figure 1 As shown, the AI resource scheduling method applied to a distributed architecture further includes:
[0096] Step S4: Calculate the node scheduling urgency based on the statistical characteristics of the task queue.
[0097] Specifically, within an evaluation window, the average waiting time Wa of all tasks passing through the node within the evaluation window is extracted, and then the node scheduling urgency Un is determined, Un=min(1,Wa / Wth2), where Wth2 is the second preset waiting time threshold.
[0098] Preferably, in this embodiment, the second preset waiting time threshold is 300ms.
[0099] Specifically, the average waiting time of all tasks passing through the node within the evaluation window is extracted and normalized to a second preset waiting time threshold to obtain the scheduling urgency. This indicator is independent of the node's own health and measures the node's sensitivity to task delays purely from the perspective of task service quality. When the average waiting time is close to or exceeds the threshold, the urgency approaches the highest value, indicating that the node has violated or is about to violate the real-time requirements of the production line tasks. This step quantifies the user experience into parameters that can be used for scheduling, enabling the scheduler not only to passively avoid "bad nodes" but also to proactively prioritize serving "urgent nodes".
[0100] Please continue reading. Figure 1 As shown, the AI resource scheduling method applied to a distributed architecture further includes:
[0101] Step S5: Determine the dynamic scheduling weight based on node health and scheduling urgency, and then generate a scheduling strategy.
[0102] Please see Figure 2 As shown, the scheduling policy generation method includes:
[0103] Step S51: Determine the dynamic scheduling weight of each node based on node health and scheduling urgency.
[0104] Specifically, at the end of each evaluation window, the dynamic scheduling weight Wdn is calculated for all available nodes in the cluster:
[0105] The dynamic scheduling weight Wdn is determined based on the node health Hn and the scheduling urgency Un, where Wdn = Hn × (1 - Un).
[0106] Please continue reading. Figure 2 As shown, the scheduling policy generation method further includes:
[0107] Step S52: Generate a scheduling strategy based on the dynamic scheduling weights of each node.
[0108] Specifically, during the next evaluation window, when new tasks arrive, the scheduler allocates tasks based on the dynamically calculated scheduling weights of each node:
[0109] The dynamic scheduling weight of each node is divided by the sum of the weights of all nodes in the cluster to obtain the probability value of that node being selected. Then, the scheduler randomly selects a node according to this probability distribution and assigns the new task to that node.
[0110] Specifically, within each real-time period, if the Hn of a node drops below 0.3, a weight recalculation is immediately triggered without waiting for the window to close.
[0111] Specifically, by combining node health with scheduling urgency, dynamic scheduling weights are generated, and new tasks are allocated according to the weighted probability distribution in the next evaluation window. This strategy achieves the intelligent scheduling logic of "assigning more tasks to healthy and non-urgent nodes, and fewer tasks to unhealthy or very urgent nodes." Compared with round-robin or least-connection algorithms, this step simultaneously prevents node overheating and crashes as well as task timeouts. Furthermore, it uses multiplication instead of addition in weight calculation, ensuring that any deterioration in any metric significantly reduces the node's selection probability, reflecting the "weakest link" safety design principle.
[0112] Please see Figure 3 As shown, the AI resource scheduling device applied to a distributed architecture includes:
[0113] Extract the health base unit, which is used to extract the basic health characteristics of the node based on CPU temperature, utilization, and task queue;
[0114] The thermal stress extraction unit is used to extract hardware thermal stress characteristics based on node operating current and core voltage;
[0115] Determine the health unit, which is used to determine the health of a node based on its basic health characteristics and hardware thermal stress characteristics;
[0116] The urgency calculation unit is used to calculate the scheduling urgency of nodes based on the statistical characteristics of the task queue.
[0117] A scheduling unit is generated to determine dynamic scheduling weights based on node health and scheduling urgency, and then generate a scheduling strategy.
[0118] The AI resource scheduling device for distributed architecture provided in this application embodiment can execute the AI resource scheduling method for distributed architecture provided in any embodiment of this application, and has the corresponding functional modules and beneficial effects of the execution method.
[0119] Please see Figure 4 As shown, it is a structural schematic diagram of an electronic device in this embodiment. The electronic device in this embodiment may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), vehicle terminals (such as vehicle navigation terminals), wearable electronic devices, etc., as well as fixed terminals such as digital TVs, desktop computers, smart home devices, etc. Figure 4 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.
[0120] like Figure 4 As shown, the electronic device includes: a processor 501, a memory 502, a communication interface 503, and a system bus 504. The processor includes at least one of a central processing unit (CPU), a graphics processing unit (GPU), or a field-programmable gate array (FPGA), configured to call computer programs and data stored in the memory and generate control instructions; the memory includes random access memory (RAM) and / or non-volatile memory (NVM), the NVM including flash memory, solid-state drive (SSD), or a combination thereof, used to store computer programs, process intermediate data, and historical data sets; the communication interface includes a wired communication module and a wireless communication module, the wired communication module supporting Ethernet or RS-485 protocols for connecting to sensor networks; the wireless communication module supporting LoRa, 5G, or satellite communication protocols for transmitting processing results to a remote server; the system bus adopts a PCI Express or AXI bus architecture to achieve high-speed data interaction and clock synchronization between the processor, memory, and communication interface.
[0121] This embodiment also provides a computer-readable storage medium, which physically stores computer-executable instructions. When the instructions are transmitted to the processing unit via the integrated circuit substrate, they are encapsulated and processed through the data channel of the bus system and then solidified into the non-volatile storage area of the storage module. The executable instructions are configured to implement the complete technical solution of the AI resource scheduling method applied to the distributed architecture when executed by the processor.
[0122] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description. It is impossible to exhaustively list all the implementation methods here. All obvious variations or modifications derived from the technical solutions of the present invention are still within the protection scope of the present invention.
Claims
1. An AI resource scheduling method applied to a distributed architecture, characterized in that, include: Based on CPU temperature, utilization, and task queue, basic node health characteristics are extracted, including temperature deviation characteristics, performance baseline deviation characteristics, and scheduling congestion characteristics. Hardware thermal stress characteristics are extracted based on node operating current and core voltage. The node health is determined based on the node's basic health characteristics and hardware thermal stress characteristics. The node health is negatively correlated with the average value of hardware thermal stress characteristics, temperature deviation characteristics, performance baseline deviation characteristics, and scheduling congestion characteristics in each real-time period. Calculate the node scheduling urgency based on the statistical characteristics of the task queue; Dynamic scheduling weights are determined based on node health and scheduling urgency, and then a scheduling strategy is generated.
2. The AI resource scheduling method applied to a distributed architecture according to claim 1, characterized in that, The extraction process for the temperature deviation feature is as follows: The average and maximum values T of the CPU temperature sequence within the evaluation window are calculated. Then, the temperature ratio factor and temperature peak penalty are calculated. Based on the temperature ratio factor and temperature peak penalty, the temperature deviation characteristics are determined. The temperature ratio factor is calculated using the average value of the CPU temperature sequence, the upper limit of the normal operating temperature of the node, and the stable idle temperature of the node. The temperature peak penalty is calculated using the maximum value of the CPU temperature sequence, the upper limit of the normal operating temperature of the node, and the critical temperature of the node. The temperature deviation feature is a weighted fusion value of the temperature ratio factor and the temperature peak penalty.
3. The AI resource scheduling method applied to a distributed architecture according to claim 1, characterized in that, The extraction process for the performance baseline deviation feature is as follows: Calculate the average CPU utilization and average memory usage within the evaluation window, and calculate the CPU deviation and memory stress factor to determine the performance baseline deviation characteristics. The CPU deviation is calculated by the average CPU utilization rate and the idle CPU utilization rate measured during node initialization. The memory pressure factor is positively correlated with the average memory occupancy rate. The performance baseline deviation feature is a weighted fusion value of the CPU deviation and the memory pressure factor.
4. The AI resource scheduling method applied to a distributed architecture according to claim 1, characterized in that, The analysis process for the scheduling congestion characteristics is as follows: Calculate the standard deviation and average queue length of the task queue within the evaluation window, and use the ratio of the standard deviation and average queue length of the task queue within the evaluation window as the queue fluctuation factor; Calculate the average waiting time of the oldest task, and then determine the waiting time factor, which is positively correlated with the average waiting time of the oldest task; The number of overload events within the statistical evaluation window is calculated, and the determination of the overload penalty factor is positively correlated with the number of overload events.
5. The AI resource scheduling method applied to a distributed architecture according to claim 4, characterized in that, An overload event is defined as follows: when the queue length rises from below the queue length overload threshold to above the queue length overload threshold, and remains above the queue length overload threshold for five consecutive sampling points, it is counted as an overload event; a new overload event is only counted when the queue length falls below the queue length overload threshold and the above condition is met again. The queue fluctuation factor, waiting time factor, and overload penalty factor are weighted and fused to determine the scheduling congestion characteristics.
6. The AI resource scheduling method applied to a distributed architecture according to claim 5, characterized in that, Within each real-time cycle, the node operating current In and core voltage Vc are processed: The instantaneous power consumption of a computing node is Pn = In × Vc; Calculate the power change ΔP = |Pn(t) - Pn(t-1)|, where t-1 represents the previous sampling time. Calculate the hardware thermal stress characteristic Fst, Fst=min(1,ΔP / ΔPmax), where ΔPmax is the preset power change benchmark.
7. The AI resource scheduling method applied to a distributed architecture according to claim 6, characterized in that, Within an evaluation window, the average waiting time Wa of all tasks passing through the node within the evaluation window is extracted to determine the node scheduling urgency, which is positively correlated with Wa.
8. The AI resource scheduling method applied to a distributed architecture according to claim 7, characterized in that, At the end of each evaluation window, the dynamic scheduling weight Wdn is calculated for all available nodes in the cluster: The dynamic scheduling weight Wdn is determined based on the node health Hn and the scheduling urgency Un, and Wdn = Hn × (1 - Un); During the next evaluation window, when new tasks arrive, the scheduler allocates tasks based on the dynamically calculated scheduling weights of each node: The dynamic scheduling weight of each node is divided by the sum of the weights of all nodes in the cluster to obtain the probability value of that node being selected. Then, the scheduler randomly selects a node according to this probability distribution and assigns the new task to that node.
9. An AI resource scheduling device for a distributed architecture, applied to the AI resource scheduling method for a distributed architecture as described in any one of claims 1-8, characterized in that, include: Extract the health base unit, which is used to extract the basic health characteristics of the node based on CPU temperature, utilization, and task queue; The thermal stress extraction unit is used to extract hardware thermal stress characteristics based on node operating current and core voltage; Determine the health unit, which is used to determine the health of a node based on its basic health characteristics and hardware thermal stress characteristics; The urgency calculation unit is used to calculate the scheduling urgency of nodes based on the statistical characteristics of the task queue. A scheduling unit is generated to determine dynamic scheduling weights based on node health and scheduling urgency, and then generate a scheduling strategy.
10. An electronic device, characterized in that, The electronic device includes: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the AI resource scheduling method for distributed architecture as described in any one of claims 1-8.