A memory intelligent scheduling method, device and electronic device
By acquiring the static and dynamic characteristics of tasks in the intelligent computing center, fusing them into memory access mode parameters, filtering candidate nodes and calculating matching degree and conflict intensity, and using a fuzzy rule base for decision-making, the problem of scheduling decisions being disconnected from memory domain states is solved, achieving efficient utilization of memory resources and improved task execution efficiency.
Patent Information
- Application Number
- CN202611270915.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-20
- Publication Date
- 2026-11-17
AI Technical Summary
In existing technologies for memory resource scheduling in intelligent computing centers, scheduling decisions are severely out of sync with the actual state of the memory domain and the memory access requirements of tasks, resulting in a global imbalance between resource utilization and task execution efficiency. In particular, in multi-level non-uniform memory domain environments, the rigid logic of prioritizing computing power leads to memory bandwidth saturation and mismatch of latency-sensitive tasks.
By acquiring the static and dynamic characteristics of the task and fusing them into memory access mode parameters, candidate nodes are screened and the matching degree and conflict intensity are calculated. A fuzzy rule base is used to make decisions and generate a memory scheduling scheme to ensure the matching degree between the task and the memory domain and the optimization of bandwidth utilization.
It achieves efficient utilization of memory resources in the intelligent computing center, avoids latency-sensitive task mismatch and bandwidth contention, and improves the overall resource utilization and task execution efficiency of the system.
Smart Images

Figure CN123114741A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer technology, and specifically relates to a memory intelligent scheduling method, device, and electronic device. Background Technology
[0002] With the explosive growth of emerging computing paradigms such as large language model training, AIGC (Artificial Intelligence Generated Content) inference, and multimodal data processing, intelligent computing centers are undergoing a profound transformation from homogeneous general-purpose computing to heterogeneous high-performance computing. The hardware architecture of a single intelligent computing server is becoming increasingly complex: the number of CPU (Central Processing Unit) cores has exceeded several hundred, and the Non-Uniform Memory Access (NUMA) layer is constantly deepening; memory pooling technology, represented by CXL (Compute Express Link), has broken through the capacity and bandwidth walls of traditional DDR (Double Data Rate) memory; heterogeneous accelerators such as GPU (Graphics Processing Unit) / NPU (Neural Processing Unit) are widely deployed, and their onboard high-bandwidth memory (HBM) constitutes a dedicated storage layer independent of the host memory.
[0003] Against this backdrop, the memory resources of modern intelligent computing centers are no longer a single, uniform "large pool," but have evolved into a multi-level, non-uniform memory domain where local DDR near-memory, CXL pooled mid-memory, and GPU video memory coexist in a heterogeneous manner. These different memory domains exhibit significant differences in access latency, bandwidth limits, capacity costs, and compatibility with the physical computing power of computing devices. Furthermore, different types of computing tasks also exhibit drastically different characteristics in their memory access patterns: some are extremely sensitive to latency, some rely on high-bandwidth streaming reads and writes, and some only require large-capacity sequential access.
[0004] These trends have collectively given rise to a key need: intelligent computing centers require a fine-grained intelligent scheduling mechanism that can sense differences in memory domain topology, match task memory access characteristics, and dynamically adapt to load changes. This mechanism aims to maximize the utilization of expensive heterogeneous memory resources while ensuring the Quality of Service (QoS), thereby truly reducing the unit operating cost of inclusive computing power.
[0005] In existing technologies, mainstream resource scheduling schemes generally adopt the following approach in memory allocation: first, select idle computing devices based on the CPU / GPU computing power requirements of the task, and then allocate memory to the nearest device's local NUMA node as the target domain. When multiple tasks compete for the same memory domain resource, they can only passively block and wait without any active intervention capability. This fixed logic essentially treats memory as a subordinate and auxiliary resource of computing power devices, and memory allocation decisions are completely dependent on the computing power selection results.
[0006] However, while the above-mentioned general approach is simple to implement, in the multi-level non-uniform memory domain environment of new intelligent computing centers, its fixed logic of prioritizing computing power and attaching memory leads to a serious disconnect between scheduling decisions and the actual state of the memory domain and the memory access needs of tasks. Specifically, this manifests as idle computing power but saturated memory bandwidth, latency-sensitive tasks being misallocated to remote memory (such as remote memory domains like CXL memory pools with access latency much higher than local DDR), and the system having no proactive guidance capability when multiple tasks compete for bandwidth. This results in a serious global imbalance between the overall resource utilization of the system and the efficiency of task execution. Summary of the Invention
[0007] To address the aforementioned problems in the prior art, the present invention provides a memory intelligent scheduling method, apparatus, and electronic device.
[0008] The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides a memory intelligent scheduling method applied to the management node of an intelligent computing center, the method comprising: S1. Obtain the memory requirement and static characteristics of the task to be scheduled, including access type, access continuity and latency sensitivity; S2. Predict the dynamic characteristics of the task to be scheduled based on historical task execution records. The dynamic characteristics include memory access locality strength, read-write ratio, and peak bandwidth demand. Fuse the static characteristics with the dynamic characteristics to obtain the memory access mode parameters of the task to be scheduled. S3. Select memory domain nodes with available memory not less than the memory requirement as candidate nodes, and calculate the matching degree of each candidate node according to the memory access mode parameters and the attribute information of each candidate node; the matching degree is used to characterize the comprehensive adaptability of the candidate node to the task to be scheduled in multiple dimensions, including the memory resource status dimension determined according to the memory domain topology and the task adaptation dimension determined according to the memory access mode parameters. S4. Based on the peak bandwidth demand and the upper limit of the memory controller bandwidth of each candidate node, calculate the predicted bandwidth utilization of each candidate node, and calculate the conflict intensity based on the predicted bandwidth utilization; the conflict intensity is used to characterize the memory bandwidth contention pressure of the candidate nodes after the task to be scheduled is added. S5. Query the matching degree and conflict intensity of each candidate node into a preset fuzzy rule base, and determine the target candidate node and the corresponding decision instruction based on the query results; the fuzzy rule base is a preset set of mapping rules used to determine the decision instruction based on the combination of matching degree and conflict intensity. S6. Generate a memory scheduling scheme based on the decision instructions.
[0009] Secondly, the present invention provides a memory intelligent scheduling device applied to the management node of an intelligent computing center, the device comprising: The acquisition module is used to acquire the memory requirements and static characteristics of the task to be scheduled, including access type, access continuity and latency sensitivity. The prediction module is used to predict the dynamic characteristics of the task to be scheduled based on historical task execution records. The dynamic characteristics include at least memory access locality strength, read-write ratio, and peak bandwidth demand. The static characteristics are fused with the dynamic characteristics to obtain the memory access mode parameters of the task to be scheduled. The first calculation module is used to filter memory domain nodes with available memory not less than the memory requirement as candidate nodes, and calculate the matching degree of each candidate node according to the memory access mode parameters and the attribute information of each candidate node. The matching degree is used to characterize the comprehensive adaptability of the candidate node to the task to be scheduled in multiple dimensions, including the memory resource status dimension determined according to the memory domain topology and the task adaptability dimension determined according to the memory access mode parameters. The second calculation module is used to calculate the predicted bandwidth utilization of each candidate node based on the peak bandwidth demand and the upper limit of the memory controller bandwidth of each candidate node, and to calculate the conflict intensity based on the predicted bandwidth utilization; the conflict intensity is used to characterize the memory bandwidth competition pressure of the candidate nodes after the task to be scheduled is added. The query module is used to query a preset fuzzy rule base for the matching degree and conflict intensity of each candidate node, and determine the target candidate node and the corresponding decision instruction based on the query results; the fuzzy rule base is a preset set of mapping rules used to determine the decision instruction based on the combination of matching degree and conflict intensity. The generation module is used to generate a memory scheduling scheme based on the decision instructions.
[0010] Thirdly, the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor, when executing a program stored in memory, implements the above-described intelligent memory scheduling method.
[0011] The intelligent memory scheduling method provided by this invention uses the requirement that available memory is not less than the task's demand as a hard threshold for selecting candidate nodes, fundamentally reversing the rigid logic of selecting computing power first and then allocating memory. This eliminates the root cause of mismatch where computing power is idle but memory bandwidth is saturated. Furthermore, by fusing static task characteristics with historically driven dynamic prediction characteristics into a memory access mode parameter, and using this parameter as the basis for matching degree calculation, the scheduling decision can perceive the differentiated needs of tasks for latency, bandwidth, and access modes, preventing latency-sensitive tasks from being misallocated to distant memory. Simultaneously, this invention pre-calculates and quantifies the predicted bandwidth utilization rate as conflict intensity based on the peak bandwidth demand of the task and the bandwidth limit of each candidate node, enabling the scheduler to grasp the bandwidth competition of each node before the task actually runs. This approach transforms passive blocking and waiting into proactive prediction of risks. Based on this, matching degree and conflict intensity are input into the fuzzy rule base as two parallel dimensions for joint decision-making. This ensures that nodes with high matching degree but saturated bandwidth are determined to migrate, while nodes with moderate matching degree but sufficient bandwidth are determined to be directly scheduled, thus avoiding decision biases from single-dimensional sorting. Finally, based on the decision instructions output by the fuzzy rule base, a memory scheduling scheme including direct allocation or split migration is proactively generated. This allows tasks to obtain executable memory allocation and computing power binding schemes without being trapped in a blocking waiting queue. This achieves a shift from passive blocking to proactive guidance in the execution mechanism, solving the global imbalance caused by the severe disconnect between scheduling decisions and the actual state of the memory domain and task memory access needs in existing technologies. Attached Figure Description
[0012] Figure 1 This is a schematic diagram of the system architecture of the intelligent computing center where the intelligent memory scheduling method provided by this invention is located; Figure 2 This is a flowchart illustrating a memory intelligent scheduling method provided by the present invention; Figure 3 This is a simplified flowchart of the intelligent memory scheduling method provided by the present invention; Figure 4 This is a comparison of the cross-domain memory access rate of the present invention with that of existing mainstream scheduling schemes under different typical loads; Figure 5This is a comparative diagram showing the ratio of task performance degradation as concurrency increases in a scenario where multiple tasks compete for bandwidth on NUMA nodes, using the present invention and the Kubernetes scheduling scheme. Figure 6 This is a schematic diagram of the structure of an electronic device provided by the present invention. Detailed Implementation
[0013] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.
[0014] To address the global imbalance caused by a severe disconnect between scheduling decisions and the actual state of the memory domain and task memory access requirements in existing technologies, this invention provides an intelligent memory scheduling method applied to the management nodes of an intelligent computing center. See also... Figure 1 As shown, the intelligent computing center comprises a computing node cluster, a computing node resource area, and a management node. The computing node cluster consists of a large number of computing server nodes providing general-purpose computing power. The computing node resource area represents the core underlying hardware of the intelligent computing center, including NUMA nodes, CXL memory pools, and GPU memory domains—these are the target hardware resources under actual control. The management node monitors the capacity, bandwidth, and link status of each NUMA node, CXL memory pool, and GPU memory domain in real time. Combined with received user task requirements, it executes scheduling logic to generate memory allocation sequences and computing power binding schemes, which are then distributed to the node agents of the computing nodes for distributed execution.
[0015] Figure 2 This is a flowchart illustrating a memory intelligent scheduling method provided by the present invention. Figure 3 This is a simplified flowchart illustrating the intelligent memory scheduling method provided by this invention, as shown below. Figure 2 and Figure 3 As shown, the intelligent memory scheduling method provided by this invention includes the following steps: S1. Obtain the memory requirements and static characteristics of the task to be scheduled.
[0016] Here, static characteristics include access type, access continuity, and latency sensitivity.
[0017] Specifically, memory requirements can be directly obtained from the task description file of the task to be scheduled. If not explicitly declared in the task description file, a default value can also be obtained from historical statistics based on the task type. For example, for a task identified as "Transformer training subtask", the management node retrieves the historical records of all similar tasks in the past 30 days. After removing outliers caused by abnormal interruptions, the historical records show that the memory requirements of 12 similar tasks are distributed between 14GB and 22GB, with an average of 16.4GB and a median of 15.8GB. The management node can then use this average or median as the memory requirement for the "Transformer training subtask".
[0018] In this invention, access type is used to characterize the task's tendency to read and write memory during execution, such as read-intensive, write-intensive, or read-write balanced. Access continuity is used to characterize the pattern of memory address access by the task, such as streaming access or highly random access patterns. Latency sensitivity is used to characterize the task's tolerance for memory access latency, such as high latency sensitivity, medium latency sensitivity, or insensitivity to latency.
[0019] Specifically, in practical applications, the access type, access continuity, and latency sensitivity of the task to be scheduled can be obtained by the user or application management platform through explicit declaration in the job description file (such as Kubernetes Pod YAML or Slurm job script) during the task submission stage. At the same time, the management node can combine the historical task execution records of the intelligent computing center with offline static code analysis tools to parse the task's running logs, automatically verify and correct the static characteristics declared by the user. In addition, for latency sensitivity, it is also possible to make auxiliary inferences based on the application type of the task (such as online inference, model training, or offline data processing) through preset mapping relationships, thereby obtaining the static characteristics of the task to be scheduled and providing a data foundation for subsequent scheduling decisions.
[0020] S2. Predict the dynamic characteristics of the task to be scheduled based on historical task execution records, and fuse the static and dynamic characteristics to obtain the memory access mode parameters of the task to be scheduled.
[0021] The dynamic characteristics include memory access locality intensity, read-write ratio and peak bandwidth demand, while the memory access mode parameters include: access type factor, continuity factor and latency sensitivity factor.
[0022] In step S2, predicting the dynamic characteristics of the task to be scheduled based on historical task execution records includes: acquiring historical task execution records, which contain execution logs, page table access records, and performance monitoring unit (PMU) data for each historical task; training a time-series prediction model using the task type identifier of each historical task and the PMU data sequence from multiple sampling periods before task execution as input, and using the memory access locality strength, read / write ratio, and peak bandwidth demand extracted from the page table access records and PMU data of each historical task as labels; and inputting the task type identifier of the task to be scheduled and the PMU data sequence from multiple sampling periods before task execution into the trained time-series prediction model to obtain the dynamic characteristics of the task to be scheduled.
[0023] Specifically, the management node trains a two-layer Long Short-Term Memory (LSTM) network offline. During training, it uses the task type identifier of historical tasks and the PMU (Performance Monitoring Unit) data sequence from multiple sampling periods before execution as input. The memory access locality strength, read / write ratio, and peak bandwidth demand extracted after the task's complete execution cycle are used as labels. After training, the model is deployed to the management node's online inference engine. When a new task arrives, the management node inputs the task type identifier and the initial PMU sampling data into the LSTM model. The model outputs predicted memory access locality strength, read / write ratio, and peak bandwidth demand as dynamic features. Furthermore, if the task type has insufficient historical samples (e.g., less than 10) or is in a cold start phase, it reverts to a preset empirical value based on the task type as the default value for the dynamic features.
[0024] The management node can retrain the LSTM network and update the model version using newly added historical data at fixed intervals (such as daily or weekly). This ensures that the model can continuously adapt to changes in task memory access patterns while keeping training overhead within a controllable range, without affecting the real-time performance of online scheduling.
[0025] The training sample data required for training the LSTM network can be obtained as follows: The management node maintains a historical task execution record database. Each record contains a task type identifier, hardware event counts collected by the Performance Monitoring Unit (PMU) during execution, and execution logs after the task ends. The raw counts provided by the PMU include: the number of memory read transactions, the number of memory write transactions, the number of local memory accesses, the number of remote memory accesses, and the instantaneous memory bandwidth usage sampled every 100ms. The management node calculates the memory locality strength, read-write ratio, and peak bandwidth demand based on the raw counts provided by the PMU.
[0026] The read-write ratio is calculated as: number of bytes written / (number of bytes read + number of bytes written), reflecting the proportion of memory write operations performed by the task. The locality of memory access strength is calculated as: number of local accesses / (number of local accesses + number of remote accesses), where the number of local and remote accesses is directly provided by the PMU's local / remote access event count. The peak bandwidth requirement is calculated as the maximum instantaneous memory bandwidth usage across all sampling points during task execution. The instantaneous memory bandwidth usage at each sampling point is calculated as: (number of bytes read within the sampling period + number of bytes written within the sampling period) / sampling period duration.
[0027] The memory access pattern parameters of the task to be scheduled are obtained by fusing static and dynamic features, including: processing both static and dynamic features into vectors and concatenating the vectors to obtain a joint vector; inputting the joint vector into a pre-trained prediction model, which then predicts the memory access pattern parameters based on the joint vector.
[0028] Specifically, firstly, the access type, access continuity, and latency sensitivity in the static features are converted into values between 0 and 1 according to a preset mapping relationship, forming static numerical vectors. Similarly, the memory access locality strength, read / write ratio, and peak bandwidth demand in the dynamic features are also processed into dynamic numerical vectors. Then, the static and dynamic numerical vectors are concatenated into a six-dimensional joint vector. The six-dimensional joint vector is input into a pre-trained prediction model, which outputs memory access mode parameters.
[0029] Here, the prediction model employs a shallow neural network, taking the six-dimensional joint vector corresponding to the historical task as input and the actual memory access pattern parameters extracted from the performance monitoring unit (PMU) data after the historical task is completed as labels. The model is trained offline on the management node. The memory access pattern parameters output by the prediction model are actually a three-dimensional vector, with its three dimensions corresponding to the access type factor, continuity factor, and latency sensitivity factor, all taking values in the range [0,1].
[0030] Understandably, access type, access continuity, and latency sensitivity in static features are coarse-grained qualitative judgments. They tell the scheduler roughly which category the task belongs to (e.g., read-intensive, streaming, high latency sensitive), but they cannot distinguish the fluctuations in actual memory access behavior within the same category due to differences in model structure, input size, operator implementation, etc. In contrast, memory access locality strength, read-write ratio, and peak bandwidth demand in dynamic features are fine-grained quantitative measured data. They accurately describe the specific memory access behavior of the task in historical execution, but their reliability depends on the coverage of historical data. For task types that have never appeared before or cold start scenarios, dynamic predictions may be inaccurate. In this invention, a predictive model learns the mapping rules between static feature labels and dynamic feature values combined with actual memory access pattern parameters. This predictive model is then used to fuse static and dynamic features, so that the fused memory access pattern parameters retain prior information about the static category while incorporating fine calibration from dynamic data. This refines coarse-grained qualitative judgments into continuous quantitative factors that carry specific deviations.
[0031] Specifically, the predictive model learns the mapping relationship of "what the actual memory access pattern parameters are when a certain type of static feature label and a certain set of dynamic feature values appear simultaneously" through a large number of historical task samples. For example, for tasks with read-intensive labels, when the memory access locality strength is 0.9 and the read-write ratio is 0.15, its actual access type factor is close to 0.88; while when the memory access locality strength is only 0.6 and the read-write ratio is 0.35, its actual access type factor is close to 0.72. The model automatically captures this pattern during training. Static labels provide class priors, and dynamic values provide fine distinctions within the class. The model fuses the two to output a continuous factor value that simultaneously carries the class affiliation and the specific degree of deviation. This allows the fused memory access pattern parameters to retain both the prior knowledge of the task type and reflect the actual behavioral characteristics of the task based on historical data, enabling subsequent matching degree calculations to be accurately evaluated in a unified continuous numerical space.
[0032] S3. Select memory domain nodes with available memory not less than the memory requirement as candidate nodes, and calculate the matching degree of each candidate node based on the memory access mode parameters and the attribute information of each candidate node.
[0033] Here, the matching degree is used to characterize the comprehensive adaptability of candidate nodes to the scheduled tasks in multiple dimensions, including the memory resource status dimension determined based on the memory domain topology and the task adaptation dimension determined based on the memory access mode parameters.
[0034] In step S3, memory domain nodes with available memory not less than the memory requirement are selected as candidate nodes. This includes: the management node obtains the available memory capacity of each memory domain node from the maintained memory domain topology, compares the available memory capacity with the memory requirement, and retains memory domain nodes with available memory capacity greater than or equal to the memory requirement as candidate nodes. Additionally, GPU memory domain nodes are only included in the candidate set when the task to be scheduled is a GPU-accelerated task and the system supports virtualized memory isolation allocation.
[0035] Specifically, the management node maintains a memory domain topology map when the system starts up or periodically. This topology map contains attribute information of all memory domain nodes in the system, including the available memory capacity of the candidate nodes, the current bandwidth utilization, the link quality between the candidate nodes and the management node, the historical task execution success rate, and the current load of the adjacent computing devices.
[0036] In practice, the management node collects data through various operating system interfaces and hardware drivers to build and update the topology: For NUMA nodes, the available memory capacity is obtained by reading standard interfaces such as / sys / devices / system / node / nodeX / meminfo, the list of adjacent CPU cores is obtained by reading / sys / devices / system / node / nodeX / cpumap (or cpulist), and the memory controller bandwidth limit and current bandwidth utilization are read by the uncore_imc counter of the performance monitoring unit (PMU). For CXL memory pool nodes, the available capacity and bandwidth limit are obtained through the CXL device driver or cxl-cli management interface, and the access latency parameters are read through the CXL device registers; for GPU memory domain nodes, the available memory capacity and bandwidth limit are read through the GPU driver interface (such as NVIDIA Management Library, NVML). The compute affinity of memory domain nodes is determined using PCIe topology information. Here, compute affinity refers to the physical proximity of memory domain nodes and compute devices, which can be determined using PCIe topology information. Specifically, this proximity is determined by the PCIe bus connection relationship and can be represented by the NUMA node number. The management node reads ` / sys / bus / pci / devices / ` <bdf>The ` / numa_node` method retrieves the NUMA node number to which the GPU belongs, and then matches this number with the list of computing devices adjacent to each NUMA node in the topology diagram to determine which computing devices are adjacent to the GPU memory domain node. For the communication links between nodes, the uplink / downlink traffic byte count is obtained through the PCIe / CXL controller statistics register, the access delay jitter variance is collected through the PMU, and the highest transmission rate is periodically measured to calculate the link quality indicators.
[0037] Specifically, the link quality score is determined by the product of three factors: the first factor is the link peak rate factor, which is the ratio of the highest transmission rate observed in the current sampling period to the link's historical highest transmission rate, reflecting the degree to which the link's peak transmission capacity at the current moment maintains its historical best level; the second factor is the link idleness factor, which is 1 minus the proportion of the link's used bandwidth to the total bandwidth, reflecting the link's remaining available bandwidth resources. The higher the proportion of used bandwidth, the closer this factor is to 0, indicating that the link is more congested; the third factor is the link stability factor, which is 1 divided by (1 + the ratio of the current delay jitter variance to the link's historical maximum jitter variance). The larger the delay jitter, the closer this factor is to 0, indicating that the link is more unstable. All three factors take values in the range [0,1]. After multiplying them together, they are multiplied by the topology coefficient correction factor (default 1.2) to obtain the final link quality score. The higher the score, the higher the overall link quality, and it is recorded in the topology graph for subsequent scheduling decisions.
[0038] Based on the collected data, the following information can be recorded for each memory domain node in the memory domain topology: node type, available memory capacity, memory controller bandwidth limit, current bandwidth utilization, access latency, list of nearby computing devices, and link quality indicators with the management node. The management node updates the information in the topology at fixed intervals during system operation to ensure the real-time nature of the data used for scheduling decisions.
[0039] In step S3, the matching degree of each candidate node is calculated based on the memory access mode parameters and the attribute information of each candidate node, including: Capacity suitability is determined based on the ratio of available memory capacity to memory requirement of candidate nodes; bandwidth suitability is determined based on the current bandwidth utilization and link quality of candidate nodes; a pre-defined mapping table is consulted based on the continuity factor and latency sensitivity factor in the memory access mode parameters to obtain the memory access suitability baseline value of candidate nodes; the memory access suitability baseline value is adjusted based on the access type factor in the memory access mode parameters to obtain the memory access suitability; reliability suitability is determined based on the success rate of candidate nodes in historical task execution; and computing power affinity is determined based on the current load of computing power devices adjacent to candidate nodes. Matching degree; a five-dimensional vector is constructed based on the capacity matching degree, bandwidth matching degree, memory access matching degree, reliability matching degree, and computing power affinity matching degree. Specifically, these matching degrees are first normalized, and then a five-dimensional vector is constructed. Using capacity matching degree, bandwidth matching degree, memory access matching degree, reliability matching degree, and computing power affinity as five reference dimensions, the positive ideal solution and negative ideal solution corresponding to the five reference dimensions are determined. The distance from the five-dimensional vector of the candidate node to the positive ideal solution and the negative ideal solution is calculated, and the matching degree of the candidate node is determined based on the distance to the positive ideal solution and the negative ideal solution.
[0040] Understandably, among the five reference dimensions, capacity adaptability, bandwidth adaptability, and computing power affinity correspond to the memory resource status dimensions determined based on the memory domain topology; while memory access adaptability and reliability adaptability correspond to the task adaptability dimensions determined based on memory access mode parameters.
[0041] In calculating the matching degree of candidate nodes, the formula for calculating capacity fit is as follows: ; in, This indicates the capacity fit; a larger value indicates that the candidate node has more available memory. The current available memory capacity of the candidate node is obtained from the memory domain topology; This indicates the memory requirements of the task to be scheduled.
[0042] Furthermore, if the candidate node is a CXL memory pool node, since the CXL memory pool belongs to a remote memory domain, it incurs additional access latency overhead compared to local memory. Therefore, the management node can multiply its capacity suitability by a discount factor, that is: ; in, The range of values for the above discount coefficient is: , The value should not exceed 0.5 to prevent the capacity fit of CXL nodes from being excessively penalized. For local DDR memory access latency, The access latency for the CXL memory pool is obtained from the memory domain topology.
[0043] Furthermore, if the candidate node is a GPU memory domain node, its capacity fit is fixed at 1.0. This is because GPU memory is dedicated heterogeneous memory, independent of host memory. As long as the available memory meets the task requirements, the capacity dimension is considered to be fully fitted. .
[0044] In calculating the matching degree of candidate nodes, the formula for calculating bandwidth suitability is as follows: ; in: This indicates bandwidth adaptability; a larger value indicates that the candidate node has more abundant bandwidth resources and better link quality. This represents the current bandwidth utilization of the candidate node's memory controller, which is recorded and dynamically updated by the memory domain topology. This represents the link quality score between the management node and the candidate node, obtained from the memory domain topology. The larger the value, the stronger the effective transmission capability of the link.
[0045] Furthermore, if the candidate node is a local NUMA node, then The maximum value of the link quality scores across all links in the system represents the absence of link loss during local access.
[0046] In calculating the matching degree of candidate nodes, the memory access fitness degree is obtained in the following way: The management node first extracts the continuity factor and latency sensitivity factor from the memory access mode parameters, and then uses these two factors as indexes to query a pre-defined memory access adaptation mapping table. This mapping table is pre-calibrated according to the physical characteristics of different memory domain types. Local DDR has low latency characteristics and is highly adaptable to tasks with high latency sensitivity; CXL pool has large capacity but relatively high latency characteristics and is highly adaptable to large-capacity access tasks with low latency sensitivity; GPU memory has ultra-high bandwidth characteristics and is highly adaptable to streaming access tasks. Each memory domain type corresponds to a baseline adaptation value in the mapping table determined by both the continuity factor and the latency sensitivity factor. The management node retrieves the corresponding baseline adaptation value from the mapping table based on the type of the candidate node (recorded by the memory domain topology).
[0047] Then, the management node adjusts the baseline adaptation value based on the access type factor in the memory access mode parameters. Read-intensive tasks have high bandwidth requirements, and their access type factor is close to 1. In this case, the baseline adaptation values for GPU memory and CXL pool are increased, while the baseline adaptation value for local DDR is decreased, according to a preset adjustment step (0.2 for GPU memory, 0.1 for CXL pool, and 0.1 for local DDR). Write-intensive tasks have high requirements for latency and cache consistency, and their access type factor is close to 0. In this case, the baseline adaptation value for local DDR is increased, while the baseline adaptation values for GPU memory and CXL pool are decreased. Read-write balanced tasks fall between these two categories, and the adjustment magnitude is correspondingly reduced. The final adjusted value is the memory access adaptation degree of the candidate node, ranging from [0,1]. A larger value indicates that the memory domain type of the candidate node matches the memory access behavior of the task to be scheduled more closely. The baseline value and the adjustment coefficient of the access type factor in the mapping table are determined offline based on historical scheduling experience.
[0048] The formula for calculating the reliability fit during the matching degree calculation of candidate nodes is as follows: ; in: This indicates the reliability adaptability; a larger value indicates that the candidate node is more stable when performing similar tasks in the past. This indicates the number of tasks that a candidate node has successfully completed in its historical records, representing the number of tasks with the same memory access pattern. This represents the total number of times a candidate node has executed tasks of the same memory access pattern in its historical records. If ,but Use the default value (e.g., 0.8). Here, "same type memory access mode" refers to historical task records that are consistent with the access type, access continuity, and latency sensitivity of the current task.
[0049] The computational affinity fit is obtained in the process of calculating the matching degree of candidate nodes as follows: The management node first obtains the computing power requirement type (CPU core requirement or GPU device requirement) of the task to be scheduled from the task description. Then, it obtains a list of computing power devices adjacent to the candidate node from the memory domain topology. For NUMA nodes, the adjacent computing power devices are the list of CPU cores under that node; for GPU memory domain nodes, the adjacent computing power devices are the CPU cores under the NUMA nodes that the GPU is affinity-attached to, and the GPU itself. If the computing power devices adjacent to the candidate node can meet the computing power requirement type of the task, the computing power affinity fit is 1; if not, the computing power affinity fit is 0. Simultaneously, the management node further adjusts the fit based on the current load of the adjacent computing power devices: if the computing power type matches and the current load of the adjacent computing power devices is low, the fit is close to 1; if the load of the adjacent computing power devices is close to saturation, the fit is reduced accordingly.
[0050] After calculating the capacity fit, bandwidth fit, memory access fit, reliability fit, and computing power affinity fit of the candidate nodes, the positive and negative ideal solutions corresponding to the five reference dimensions are determined. Specifically, this includes: statistically analyzing the fit distribution of each reference dimension in historical scheduling based on historical task execution records; determining the historical reference upper limit and historical reference lower limit for each reference dimension based on the statistical results; obtaining the theoretical upper limit and theoretical lower limit of each dimension of the system hardware; determining the positive ideal solution based on the common constraint of the historical reference upper limit and the theoretical upper limit of each reference dimension; and determining the negative ideal solution based on the common constraint of the historical reference lower limit and the theoretical lower limit of the hardware.
[0051] Specifically, the positive ideal solution is determined by the common constraint of the historical reference upper limit and the theoretical upper limit of each reference dimension, and the negative ideal solution is determined by the common constraint of the historical reference lower limit and the hardware theoretical lower limit of each reference dimension. Specifically, the management node takes the smaller value between the historical reference upper limit and the theoretical upper limit as the positive ideal solution value of the dimension, and takes the larger value between the historical reference lower limit and the hardware theoretical lower limit as the negative ideal solution value of the dimension.
[0052] After determining the positive and negative ideal solutions corresponding to the five reference dimensions, calculate the distance from the five-dimensional vector of the candidate node to the positive and negative ideal solutions, and obtain the matching degree of the candidate node based on the distance to the positive and negative ideal solutions.
[0053] S4. Calculate the predicted bandwidth utilization of each candidate node based on the peak bandwidth demand and the upper limit of the memory controller bandwidth of each candidate node, and calculate the conflict intensity based on the predicted bandwidth utilization.
[0054] Here, conflict intensity is used to characterize the memory bandwidth contention pressure of candidate nodes after the task to be scheduled is added.
[0055] In step S4, the predicted bandwidth utilization rate of each candidate node is calculated based on the peak bandwidth demand and the upper limit of the memory controller bandwidth of each candidate node. This includes: for each candidate node, dividing the peak bandwidth demand predicted in step S2 by the upper limit of the memory controller bandwidth of the candidate node to obtain the additional bandwidth occupancy ratio of the candidate node after the task is added; adding the additional bandwidth occupancy ratio to the current bandwidth utilization rate of the candidate node recorded in the topology diagram to obtain the predicted bandwidth utilization rate.
[0056] Then, the conflict intensity is calculated based on the predicted bandwidth utilization, including: for each candidate node, obtaining the historical bandwidth utilization distribution of that candidate node; determining the cumulative probability corresponding to the predicted bandwidth utilization based on the historical bandwidth utilization distribution; and determining the conflict intensity based on the cumulative probability. Among these, the cumulative probability is positively correlated with the conflict intensity, and in practice, the cumulative probability can be directly used as the conflict intensity.
[0057] S5. Query the pre-set fuzzy rule base for the matching degree and conflict intensity of each candidate node, and determine the target candidate node and the corresponding decision instruction based on the query results.
[0058] Here, the fuzzy rule base is a pre-defined set of mapping rules used to determine decision instructions based on a combination of matching degree and conflict intensity.
[0059] In step S5, for each candidate node, its matching degree and conflict intensity are input into the fuzzy rule base to query the candidate decision instruction corresponding to the candidate node; the candidate decision instruction includes a direct scheduling instruction or a migration instruction; the candidate node with the highest matching degree score in the query results is determined as the target candidate node, and its corresponding candidate decision instruction is determined as the decision instruction; if the decision instruction is a migration instruction, the candidate node with the matching degree score lower than the target candidate node is determined as the suboptimal node, and the target candidate node and the suboptimal node are jointly used as the allocation objects in the scheduling scheme.
[0060] In this invention, the fuzzy rule base is a set of mapping rules pre-installed in the management node, used to determine the decision instruction corresponding to a candidate node based on the combination of the matching degree and conflict intensity. Each rule in the fuzzy rule base defines the correspondence between the matching degree interval, the conflict intensity interval, and the decision instruction. Table 1 illustrates the specific form of the fuzzy rule base as an example: Table 1
[0061] In Table 1, DST represents Direct Scheduling Instruction, which means all tasks are allocated to the target candidate node, and MIG represents Migration Instruction, which means tasks are split and allocated to the target candidate node and the second-best node according to the conflict intensity.
[0062] In the above rules, the boundary between the matching degree and conflict intensity is a preset empirical value, which can also be adjusted offline by the management node based on the system's operating status and historical scheduling effects. The input to each rule is the matching degree and conflict intensity of a single candidate node, and the output is the candidate decision instruction corresponding to that node. After querying each candidate node one by one, the management node selects the one with the highest matching degree from all candidate nodes as the target candidate node, and its corresponding candidate decision instruction becomes the final decision instruction.
[0063] Optionally, in one implementation, before performing the query operation in step S5, the method of the present invention may further include: first determining the total number of candidate nodes; if the total number of candidate nodes is 1, then skipping the fuzzy rule base query step, directly determining the unique candidate node as the target candidate node, and forcibly generating a direct scheduling instruction (DST) as a decision instruction.
[0064] S6. Generate a memory scheduling scheme based on the decision instructions.
[0065] Specifically, when the decision instruction is DST, the management node determines the target candidate node as the allocation node, uses the memory requirement as the allocation capacity of the node, and determines the computing power devices adjacent to the node as the bound computing power devices. The scheduling scheme contains only one allocation entry, and all memory allocation for the task is completed on this node. Computing power binding also only involves the computing power devices adjacent to this node. When the decision instruction is MIG, the management node determines the target candidate node as the primary node and determines the determined second-best node (the candidate node with a matching score lower than the target candidate node) as the secondary node. The allocation capacity of the primary node is max(1 Γ, 0.5) × Q, where the allocated capacity of the secondary node is Q minus the allocated capacity of the primary node, and Γ represents the conflict intensity of the target candidate node. The computing power devices adjacent to the primary and secondary nodes are designated as bound computing power devices. The scheduling scheme includes two allocation entries; the task's memory is split among the two nodes according to the conflict intensity ratio, and the computing power is also bound to the respective allocated nodes.
[0066] After the management node completes the above allocation, the generated memory scheduling scheme includes the following information: the memory domain nodes allocated to the tasks to be scheduled and their corresponding allocated capacities (in GB), as well as the binding relationships between each memory domain node and its adjacent computing devices. The management node distributes the memory scheduling scheme to the resource manager (such as a Kubernetes APIServer or Slurm controller). The resource manager is the resource management infrastructure of the intelligent computing center (such as a Kubernetes APIServer or Slurm controller), responsible for converting the scheduling scheme into specific resource allocation instructions and distributing them to the node agents on each compute node for execution. After receiving the instructions from the resource manager, the node agent (kubelet or slurmd) completes memory allocation, computing power binding, and task startup locally, ultimately realizing the implementation and execution of the scheduling scheme.
[0067] The intelligent memory scheduling method provided by this invention uses the requirement that available memory is not less than the task's demand as a hard threshold for selecting candidate nodes, fundamentally reversing the rigid logic of selecting computing power first and then allocating memory. This eliminates the root cause of mismatch where computing power is idle but memory bandwidth is saturated. Furthermore, by fusing static task characteristics with historically driven dynamic prediction characteristics into a memory access mode parameter, and using this parameter as the basis for matching degree calculation, the scheduling decision can perceive the differentiated needs of tasks for latency, bandwidth, and access modes, preventing latency-sensitive tasks from being misallocated to distant memory. Simultaneously, this invention pre-calculates and quantifies the predicted bandwidth utilization rate as conflict intensity based on the peak bandwidth demand of the task and the bandwidth limit of each candidate node, enabling the scheduler to grasp the bandwidth competition of each node before the task actually runs. This approach transforms passive blocking and waiting into proactive prediction of risks. Based on this, matching degree and conflict intensity are input into the fuzzy rule base as two parallel dimensions for joint decision-making. This ensures that nodes with high matching degree but saturated bandwidth are determined to migrate, while nodes with moderate matching degree but sufficient bandwidth are determined to be directly scheduled, thus avoiding decision biases from single-dimensional sorting. Finally, based on the decision instructions output by the fuzzy rule base, a memory scheduling scheme including direct allocation or split migration is proactively generated. This allows tasks to obtain executable memory allocation and computing power binding schemes without being trapped in a blocking waiting queue. This achieves a shift from passive blocking to proactive guidance in the execution mechanism, solving the global imbalance caused by the severe disconnect between scheduling decisions and the actual state of the memory domain and task memory access needs in existing technologies.
[0068] To verify the effectiveness of the NUMA / CXL-aware intelligent memory scheduling method proposed in this invention, a smart computing center hardware platform was built in a simulation environment. Various representative workloads were selected for testing, and the method was compared with existing mainstream scheduling schemes. The experimental environment configuration is as follows: A dual-socket NUMA system simulation environment was built based on the QEMU / KVM virtualization platform, consisting of two NUMA nodes, each equipped with 128GB of DDR5 local memory and a memory controller bandwidth limit of 204.8GB / s; a CXL memory pool (512GB capacity, 150ns access latency, and a bandwidth limit of 51.2GB / s) was configured; and two GPUs (each with 80GB of HBM memory) were configured. Four representative workloads were selected for testing: Transformer inference tasks (read-intensive, streaming, and high latency sensitive), large language model training tasks (read / write balanced, streaming, and medium latency sensitive), graph database query tasks (read / write balanced, random, and high latency sensitive), and batch data cleaning tasks (write-intensive, sequential, and low latency sensitive). Two existing scheduling schemes were selected as baselines: the Linux kernel's default NUMA local allocation strategy (autonuma) and the current mainstream Kubernetes default scheduler.
[0069] The experimental content and results are as follows: I. Cross-domain access rate comparison: The cross-domain memory access ratio (i.e., the proportion of tasks accessing non-local NUMA nodes, CXL pools, or non-affinity GPU memory) was statistically analyzed under different scheduling schemes for various workloads. Cross-domain access rate = (number of requests for non-local NUMA node memory by tasks) / (total number of memory access requests by tasks) × 100%; where non-local NUMA nodes refer to memory nodes that do not belong to the same NUMA domain as the computing power device bound to the task, including remote NUMA nodes, CXL memory pools, and non-affinity GPU memory domains. The experiment was run 10 times for each workload under full load conditions, and the average cross-domain access rate was taken.
[0070] Experimental results are as follows Figure 4 As shown, the cross-domain access rates for the four workloads in this invention are as follows: Transformer inference task 9.7%, large language model training task 14.8%, graph database query task 15.2%, and batch data cleaning task 13.5%, with an average of 13.3%. Under the same conditions, Kubernetes' cross-domain access rates for the four workloads are 39.6%, 45.3%, 46.7%, and 37.9%, with an average of 42.4%; and autonuma's are 24.1%, 29.8%, 31.5%, and 29.2%, with an average of 28.7%. The graph database query task, due to its random memory access characteristics, has a slightly higher cross-domain access rate than other tasks, which is consistent with its expected memory access pattern. This demonstrates the effectiveness of this invention's approach of using the memory domain as an independent scheduling entity and determining memory access suitability based on memory access pattern parameters to match task and memory domain types.
[0071] II. Performance degradation comparison in multi-task bandwidth contention scenarios: A series of progressive load scenarios were designed, gradually increasing the number of concurrent tasks on the NUMA0 node (from 1 to 8). The performance degradation ratio of the average execution time of each task relative to the single-task runtime was measured. The performance degradation ratio is calculated as: [(Average runtime of task in multi-task concurrency - Single-task exclusive runtime) / Single-task exclusive runtime] × 100%. Here, the single-task exclusive runtime is the time taken for the task to execute alone without other tasks competing for memory bandwidth, and the average runtime of task in multi-task concurrency is the average runtime of the same task under concurrent scenarios. The test was run 3 times for each concurrency level, and the average value was taken.
[0072] Experimental results are as follows Figure 5 As shown, when there is only one task, neither this invention nor Kubernetes exhibits any performance degradation. When multiple tasks are concurrent, the average performance degradation of each task in this invention is less than that in Kubernetes, and the more tasks there are, the more significant the difference becomes. This demonstrates the effectiveness of this invention in proactively making decisions through conflict intensity prediction and fuzzy rule base.
[0073] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.
[0074] Based on the same inventive concept, this invention also provides a memory intelligent scheduling device applied to the management node of an intelligent computing center, the device comprising: The acquisition module is used to acquire the memory requirements and static characteristics of the task to be scheduled, including access type, access continuity and latency sensitivity. The prediction module is used to predict the dynamic characteristics of the task to be scheduled based on historical task execution records. The dynamic characteristics include at least memory access locality strength, read-write ratio, and peak bandwidth demand. The static characteristics are fused with the dynamic characteristics to obtain the memory access mode parameters of the task to be scheduled. The first calculation module is used to filter memory domain nodes with available memory not less than the memory requirement as candidate nodes, and calculate the matching degree of each candidate node according to the memory access mode parameters and the attribute information of each candidate node. The matching degree is used to characterize the comprehensive adaptability of the candidate node to the task to be scheduled in multiple dimensions, including the memory resource status dimension determined according to the memory domain topology and the task adaptability dimension determined according to the memory access mode parameters. The second calculation module is used to calculate the predicted bandwidth utilization of each candidate node based on the peak bandwidth demand and the upper limit of the memory controller bandwidth of each candidate node, and to calculate the conflict intensity based on the predicted bandwidth utilization; the conflict intensity is used to characterize the memory bandwidth competition pressure of the candidate nodes after the task to be scheduled is added. The query module is used to query a preset fuzzy rule base for the matching degree and conflict intensity of each candidate node, and determine the target candidate node and the corresponding decision instruction based on the query results; the fuzzy rule base is a preset set of mapping rules used to determine the decision instruction based on the combination of matching degree and conflict intensity. The generation module is used to generate a memory scheduling scheme based on the decision instructions.
[0075] Optionally, the prediction module predicts the dynamic characteristics of the task to be scheduled based on historical task execution records, including: Obtain historical task execution records, which include execution logs, page table access records, and performance monitoring unit (PMU) data for each historical task; The time series prediction model is trained by taking the task type identifier in the execution log of each historical task as input and the memory access locality strength, read-write ratio and bandwidth demand peak extracted from the page table access records and PMU data of each historical task as labels. The task type identifier of the task to be scheduled is input into the trained time-series prediction model to obtain the dynamic features of the task to be scheduled.
[0076] Optionally, the prediction module fuses the static features with the dynamic features to obtain the memory access pattern parameters of the task to be scheduled, including: The access type, access continuity, and latency sensitivity in the static features are mapped to numerical values according to a preset mapping relationship to obtain a static numerical vector. Map the memory access locality strength, read-write ratio and bandwidth demand peak in the dynamic features to dynamic numerical vectors; The static numerical vector and the dynamic numerical vector are input into a pre-trained prediction model, and the memory access mode parameters are output through the prediction model.
[0077] Optionally, the memory access mode parameters include an access type factor, a continuity factor, and a latency sensitivity factor; the attribute information includes the available memory capacity of the candidate node, the current bandwidth utilization rate, the link quality between the candidate node and the management node, the historical task execution success rate, and the current load of the adjacent computing devices.
[0078] Optionally, the first calculation module calculates the matching degree of each candidate node based on the memory access mode parameters and the attribute information of each candidate node, including: The capacity fit is determined based on the ratio of the available memory capacity of the candidate nodes to the memory requirement. The bandwidth suitability is determined based on the current bandwidth utilization of the candidate nodes and the link quality. The memory access mode parameters are used to query a preset mapping table to obtain the memory access adaptation baseline value of the candidate node; the memory access mode parameters are used to correct the memory access adaptation baseline value to obtain the memory access adaptation degree. The reliability fit is determined based on the success rate of the candidate nodes in historical task execution. The computing power affinity is determined based on the current load of the computing power devices adjacent to the candidate node. A five-dimensional vector is constructed based on the capacity adaptability, bandwidth adaptability, memory access adaptability, reliability adaptability, and computing power affinity adaptability. Using capacity adaptability, bandwidth adaptability, memory access adaptability, reliability, and computing power affinity as five reference dimensions, the positive and negative ideal solutions corresponding to the five reference dimensions are determined. Calculate the distance from the five-dimensional vector of the candidate node to the positive ideal solution and the negative ideal solution, and determine the matching degree of the candidate node based on the distance to the positive ideal solution and the negative ideal solution.
[0079] Optionally, the first calculation module determines the positive and negative ideal solutions corresponding to the five reference dimensions, including: Based on historical task execution records, the fit distribution of each reference dimension in historical scheduling is statistically analyzed, and the historical reference upper limit and historical reference lower limit of each reference dimension are determined based on the statistical results. Obtain the theoretical upper and lower limits of each dimension of the system hardware; Positive ideal solutions are determined by the common constraints of the historical reference upper bound and the theoretical upper bound of each reference dimension, and negative ideal solutions are determined by the common constraints of the historical reference lower bound and the hardware theoretical lower bound of each reference dimension.
[0080] Optionally, the second calculation module calculates the conflict intensity based on the predicted bandwidth utilization, including: for each candidate node, obtaining the historical bandwidth utilization distribution of the candidate node; determining the cumulative probability corresponding to the predicted bandwidth utilization based on the historical bandwidth utilization distribution; and determining the conflict intensity based on the cumulative probability; wherein the cumulative probability is positively correlated with the conflict intensity.
[0081] Optionally, the query module queries a pre-set fuzzy rule base for the matching degree and conflict intensity of each candidate node, and determines the target candidate node and corresponding decision instruction based on the query results, including: For each candidate node, its matching degree and conflict intensity are input into the fuzzy rule base, and the candidate decision instruction corresponding to the candidate node is obtained by querying; the candidate decision instruction includes a direct scheduling instruction or a migration instruction; The candidate node with the highest matching score in the query results is determined as the target candidate node, and its corresponding candidate decision instruction is determined as the decision instruction. If the decision instruction is a migration instruction, then the candidate node with a matching score lower than the target candidate node is determined as the suboptimal node, and the target candidate node and the suboptimal node are jointly used as the allocation objects in the scheduling scheme.
[0082] This invention also provides an electronic device, such as... Figure 6 As shown, it includes a processor 601, a communication interface 602, a memory 603, and a communication bus 604, wherein the processor 601, the communication interface 602, and the memory 603 communicate with each other through the communication bus 604. Memory 603 is used to store computer programs; When the processor 601 executes the computer program stored in the memory 603, it implements the steps of any of the above-mentioned intelligent memory scheduling methods.
[0083] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus.
[0084] The communication interface is used for communication between the aforementioned electronic devices and other devices.
[0085] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.
[0086] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0087] It should be noted that, for the device / electronic device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments. Furthermore, since the device and electronic device in the embodiments of the present invention are devices and electronic devices applying the above-described intelligent memory scheduling method, all embodiments of the above-described intelligent memory scheduling method are applicable to the device and electronic device, and can achieve the same or similar beneficial effects.
[0088] It should be noted that the terms "first," "second," etc., 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 so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention.
[0089] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In the description of the invention, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.
[0090] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.< / bdf>
Claims
1. A memory intelligent scheduling method, characterized in that, The method, applied to the management node of an intelligent computing center, includes: S1. Obtain the memory requirement and static characteristics of the task to be scheduled, including access type, access continuity and latency sensitivity; S2. Predict the dynamic characteristics of the task to be scheduled based on historical task execution records. The dynamic characteristics include memory access locality strength, read-write ratio, and peak bandwidth demand. Fuse the static characteristics with the dynamic characteristics to obtain the memory access mode parameters of the task to be scheduled. S3. Select memory domain nodes with available memory not less than the memory requirement as candidate nodes, and calculate the matching degree of each candidate node according to the memory access mode parameters and the attribute information of each candidate node; the matching degree is used to characterize the comprehensive adaptability of the candidate node to the task to be scheduled in multiple dimensions, including the memory resource status dimension determined according to the memory domain topology and the task adaptation dimension determined according to the memory access mode parameters. S4. Based on the peak bandwidth demand and the upper limit of the memory controller bandwidth of each candidate node, calculate the predicted bandwidth utilization of each candidate node, and calculate the conflict intensity based on the predicted bandwidth utilization; the conflict intensity is used to characterize the memory bandwidth contention pressure of the candidate nodes after the task to be scheduled is added. S5. Query the matching degree and conflict intensity of each candidate node into a preset fuzzy rule base, and determine the target candidate node and the corresponding decision instruction based on the query results; the fuzzy rule base is a preset set of mapping rules used to determine the decision instruction based on the combination of matching degree and conflict intensity. S6. Generate a memory scheduling scheme based on the decision instructions.
2. The memory intelligent scheduling method according to claim 1, characterized in that, The step of predicting the dynamic characteristics of the task to be scheduled based on historical task execution records includes: Obtain historical task execution records, which include execution logs, page table access records, and performance monitoring unit (PMU) data for each historical task; The time series prediction model is trained by taking the task type identifier of each historical task and the PMU data sequence of multiple sampling periods before the task execution as input, and the memory access locality strength, read-write ratio and bandwidth demand peak extracted from the page table access records and PMU data of each historical task as labels. The task type identifier of the task to be scheduled and the PMU data sequence of multiple sampling periods before the task execution are input into the trained time series prediction model to obtain the dynamic characteristics of the task to be scheduled.
3. The memory intelligent scheduling method according to claim 1, characterized in that, The static features and dynamic features are fused to obtain the memory access mode parameters of the task to be scheduled, including: The access type, access continuity, and latency sensitivity in the static features are mapped to numerical values according to a preset mapping relationship to obtain a static numerical vector. Map the memory access locality strength, read-write ratio and bandwidth demand peak in the dynamic features to dynamic numerical vectors; The static numerical vector and the dynamic numerical vector are input into a pre-trained prediction model, and the memory access mode parameters are output through the prediction model.
4. The memory intelligent scheduling method according to claim 1, characterized in that, The memory access mode parameters include access type factor, continuity factor, and latency sensitivity factor; the attribute information includes the available memory capacity of the candidate node, current bandwidth utilization, link quality with the management node, historical task execution success rate, and current load of the adjacent computing devices.
5. The memory intelligent scheduling method according to claim 4, characterized in that, The step of calculating the matching degree of each candidate node based on the memory access mode parameters and the attribute information of each candidate node includes: The capacity fit is determined based on the ratio of the available memory capacity of the candidate nodes to the memory requirement. The bandwidth suitability is determined based on the current bandwidth utilization of the candidate nodes and the link quality. The memory access mode parameters are used to query a preset mapping table to obtain the memory access adaptation baseline value of the candidate node; the memory access mode parameters are used to correct the memory access adaptation baseline value to obtain the memory access adaptation degree. The reliability fit is determined based on the success rate of the candidate nodes in historical task execution. The computing power affinity is determined based on the current load of the computing power devices adjacent to the candidate node. A five-dimensional vector is constructed based on the capacity adaptability, bandwidth adaptability, memory access adaptability, reliability adaptability, and computing power affinity adaptability. Using capacity adaptability, bandwidth adaptability, memory access adaptability, reliability, and computing power affinity as five reference dimensions, the positive and negative ideal solutions corresponding to the five reference dimensions are determined. Calculate the distance from the five-dimensional vector of the candidate node to the positive ideal solution and the negative ideal solution, and determine the matching degree of the candidate node based on the distance to the positive ideal solution and the negative ideal solution.
6. The memory intelligent scheduling method according to claim 5, characterized in that, Determining the positive and negative ideal solutions corresponding to the five reference dimensions includes: Based on historical task execution records, the fit distribution of each reference dimension in historical scheduling is statistically analyzed, and the historical reference upper limit and historical reference lower limit of each reference dimension are determined based on the statistical results. Obtain the theoretical upper and lower limits of each dimension of the system hardware; Positive ideal solutions are determined by the common constraints of the historical reference upper limit and the theoretical upper limit for each reference dimension, and negative ideal solutions are determined by the common constraints of the historical reference lower limit and the theoretical lower limit for each reference dimension.
7. The memory intelligent scheduling method according to claim 1, characterized in that, The step of calculating the conflict intensity based on the predicted bandwidth utilization includes: for each candidate node, obtaining the historical bandwidth utilization distribution of that candidate node; determining the cumulative probability corresponding to the predicted bandwidth utilization based on the historical bandwidth utilization distribution; and determining the conflict intensity based on the cumulative probability; wherein the cumulative probability is positively correlated with the conflict intensity.
8. The memory intelligent scheduling method according to claim 1, characterized in that, The step of querying a pre-set fuzzy rule base for the matching degree and conflict intensity of each candidate node, and determining the target candidate node and corresponding decision instruction based on the query results, includes: For each candidate node, its matching degree and conflict intensity are input into the fuzzy rule base, and the candidate decision instruction corresponding to the candidate node is obtained by querying; the candidate decision instruction includes a direct scheduling instruction or a migration instruction; The candidate node with the highest matching score in the query results is determined as the target candidate node, and its corresponding candidate decision instruction is determined as the decision instruction. If the decision instruction is a migration instruction, then the candidate node with a matching score lower than the target candidate node is determined as the suboptimal node, and the target candidate node and the suboptimal node are jointly used as the allocation objects in the scheduling scheme.
9. A memory intelligent scheduling device, characterized in that, The device, used as a management node in an intelligent computing center, includes: The acquisition module is used to acquire the memory requirements and static characteristics of the task to be scheduled, including access type, access continuity and latency sensitivity. The prediction module is used to predict the dynamic characteristics of the task to be scheduled based on historical task execution records. The dynamic characteristics include at least memory access locality strength, read-write ratio, and peak bandwidth demand. The static characteristics are fused with the dynamic characteristics to obtain the memory access mode parameters of the task to be scheduled. The first calculation module is used to filter memory domain nodes with available memory not less than the memory requirement as candidate nodes, and calculate the matching degree of each candidate node according to the memory access mode parameters and the attribute information of each candidate node. The matching degree is used to characterize the comprehensive adaptability of the candidate node to the task to be scheduled in multiple dimensions, including the memory resource status dimension determined according to the memory domain topology and the task adaptability dimension determined according to the memory access mode parameters. The second calculation module is used to calculate the predicted bandwidth utilization of each candidate node based on the peak bandwidth demand and the upper limit of the memory controller bandwidth of each candidate node, and to calculate the conflict intensity based on the predicted bandwidth utilization; the conflict intensity is used to characterize the memory bandwidth competition pressure of the candidate nodes after the task to be scheduled is added. The query module is used to query a preset fuzzy rule base for the matching degree and conflict intensity of each candidate node, and determine the target candidate node and the corresponding decision instruction based on the query results; the fuzzy rule base is a preset set of mapping rules used to determine the decision instruction based on the combination of matching degree and conflict intensity. The generation module is used to generate a memory scheduling scheme based on the decision instructions.
10. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The method steps for implementing the above-described intelligent memory scheduling method when a processor executes a computer program stored in memory.
