A machine learning resource regulation method, device, and medium based on fuzzy clustering

Abstract

This invention discloses a machine learning resource regulation method, device, and medium based on fuzzy clustering, relating to the field of resource scheduling technology. The method includes: extracting the occupancy intensity set of each operator's execution event from the operator execution sequence; synthesizing operator resource occupancy relationships; feeding these relationships into a fuzzy clustering model for membership partitioning to determine mixed-partition membership states; placing operator resource occupancy relationships with complementary resource demands into the same graphics processor execution window according to the mixed-partition membership states, forming complementary mixed-partition relationships; and when a mixed-partition failure occurs in a complementary mixed-partition relationship, reversing the operator resource occupancy relationship corresponding to the failure position to any of the following states: exclusive protection occupancy, off-peak execution occupancy, or prohibited mixed-partition occupancy. This invention improves the continuity, load stability, and regulation reliability of mixed-partition operation within the graphics processor resource pool by limiting the mixed-partition failure position.

Description

Technical Field

[0001] This invention relates to the field of resource scheduling technology, and in particular to a machine learning resource regulation method, device and medium based on fuzzy clustering. Background Technology

[0002] In recent years, the training scale and inference concurrency of machine learning models have continued to increase, making graphics processing unit (GPU) resource pools a key component in high-performance computing environments. Compared to traditional task-level resource allocation, operator-level runtime scheduling can reflect the execution behavior of computing tasks within GPUs with finer granularity and is gradually becoming an important direction in resource regulation research. During the execution of machine learning tasks, different operators exhibit different resource consumption patterns due to differences in network structure, tensor size, memory access methods, and execution flow arrangements. Related techniques typically combine GPU runtime logs, machine learning framework execution trajectories, and performance sampling data to analyze operator runtime sequence and resource consumption characteristics, thereby providing a data foundation for operator scheduling, window handling, and resource reuse.

[0003] However, the existing mixed-partition control of machine learning operators in the GPU resource pool still struggles to transform mixed-partition failure locations into subsequent window constraints: when operators are jointly carried within the same GPU execution window, the resource occupancy status changes with the running sampling boundary and execution flow connection relationship. Mixed-partition failures are often only treated as running anomalies within the current window and fail to be stably written back to the original operator resource occupancy relationship. As a result, subsequent GPU execution windows lack inheritable failure constraint basis and may continue to use the original mixed-partition candidate relationship, affecting the stability of continuous carrying of machine learning tasks in the resource pool. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a machine learning resource regulation method based on fuzzy clustering to solve the problem of difficulty in converting the failure positions of mixed parts into subsequent window constraints.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a machine learning resource regulation method based on fuzzy clustering, comprising, The operator execution events are intercepted in the graphics processor resource pool, and the execution succession relationship between the operator execution events is established according to the operator startup order to form an operator execution succession sequence. Extract the occupancy intensity set of each operator running event from the operator running sequence, synthesize the operator resource occupancy relationship, send the operator resource occupancy relationship into the fuzzy clustering model, classify the operator resource occupancy relationship according to the occupancy intensity set, and determine the mixed membership state of the operator resource occupancy relationship in computationally intensive occupancy, video memory bandwidth occupancy, video memory transport occupancy, and gap filling occupancy. According to the mixed-part affiliation status, the resource occupancy relationship of operators with complementary resource requirements is placed into the same graphics processor execution window to form a complementary mixed-part relationship of operators. When a mixed-part relationship of complementary operators fails, the resource occupancy relationship of the operator corresponding to the failed position is reversed to any of the following states: exclusive protection occupancy, off-peak execution occupancy, and prohibited mixed-part occupancy. The subsequent complementary mixed-part relationships of operators are then redefined according to the reversed operator resource occupancy relationship.

[0007] As a preferred embodiment of the machine learning resource regulation method based on fuzzy clustering described in this invention, wherein: the formation of the operator execution sequence specifically refers to... Extract operator execution events from the task execution events of the graphics processor resource pool, and label the operator execution events with machine learning task identifier, operator start position, operator end position, and graphics processor occupied position; The operator execution events are merged based on the same machine learning task identifier, and the merged operator execution events are arranged in order from earliest to latest according to the operator start position to form the operator start time sequence of the same machine learning task. In the operator startup sequence, the operator end position of the previous operator running event and the operator start position of the next operator running event define the running continuation segment. The running continuation segment is written between adjacent operator running events to form an execution succession relationship. Operator execution events in the same machine learning task are connected by execution succession relationships to form an operator execution succession sequence.

[0008] As a preferred embodiment of the machine learning resource regulation method based on fuzzy clustering described in this invention, the resource occupancy relationship of the synthesis operator is specifically as follows: Locate the execution sequence corresponding to a single operator execution event within the operator execution sequence, and delineate the resource occupation range from the operator start position to the operator end position; The occupation boundary is defined by the resource occupation range. Within the occupation boundary, the execution intensity item, transmission intensity item, and gap intensity item of the graphics processor occupation position are collected to form an occupation intensity set. Configure the occupancy intensity set ownership index for operator running events, and establish the occupancy ownership relationship between operator running events and occupancy intensity sets; Based on the position of the occupancy relationship and the succession segment, the occupancy relationship is added to the adjacent operator running events in the operator running succession sequence to synthesize the operator resource occupancy relationship.

[0009] As a preferred embodiment of the machine learning resource regulation method based on fuzzy clustering described in this invention, the determination of the operator resource occupancy relationship in the mixed membership state of computationally intensive occupancy, memory bandwidth occupancy, memory transport occupancy, and gap filling occupancy specifically involves: Based on the occupancy attribution relationship in the operator resource occupancy relationship, the occupancy intensity set associated with the operator running event is retrieved, and the occupancy intensity set is arranged according to the order of the operator running events in the operator running succession sequence to form the occupancy sequence to be divided; The occupancy sequence to be divided is fed into a fuzzy clustering model for membership division, resulting in membership division results; The execution intensity item, transfer intensity item, and gap intensity item are obtained from the occupancy intensity set. According to the membership partitioning results, the execution intensity item is associated with computationally intensive occupancy, the transfer intensity item is associated with memory bandwidth occupancy and memory transport occupancy, and the gap intensity item is associated with gap filling occupancy, forming an initial mixed membership state. Within the adjacent operator running events covered by the running continuation segment, the initial mixed membership states are concatenated according to the order of the operator running succession sequence to form mixed membership states; The mixed membership state is used to constrain the allocation of operator resource occupancy relationships among computationally intensive occupancy, video memory bandwidth occupancy, video memory transport occupancy, and gap filling occupancy, thereby determining the mixed membership state of operator resource occupancy relationships.

[0010] As a preferred embodiment of the machine learning resource regulation method based on fuzzy clustering described in this invention, wherein: the formation of the membership partitioning result specifically refers to, Within the occupancy sequence to be divided, the execution intensity items, transmission intensity items, and gap intensity items of the occupancy intensity set are arranged along the location of the operator running event to form the membership input sequence; The membership input sequence is fed into the fuzzy clustering model, and the occupancy reference cluster of the fuzzy clustering model is constructed by computationally intensive occupancy, memory bandwidth occupancy, memory transfer occupancy, and gap filling occupancy. The membership values ​​of the input sequence in each occupied reference cluster are determined to form fuzzy membership relationships; According to the fuzzy attribution relationship, the attribution positions of the occupied reference clusters in the membership input sequence are rearranged to form the membership cluster order result; The cluster order position of the membership cluster order result is backfilled into the occupancy sequence to be divided, and the occupancy sequence to be divided is marked as belonging to the division of computationally intensive occupancy, memory bandwidth occupancy, memory transport occupancy and gap filling occupancy, which is used as the membership partitioning result.

[0011] As a preferred embodiment of the machine learning resource regulation method based on fuzzy clustering described in this invention, wherein: the formation of complementary mixed-part relationships of operators specifically refers to... Different combinations of operator resource occupancy relationships are selected from the mixed membership states, and combinations of different operator resource occupancy relationships that coexist with computationally intensive occupancy and gap filling occupancy, and whose memory bandwidth occupancy and memory transport occupancy are staggered, are selected as candidate occupancy complementary relationships. The execution window is positioned for the complementary candidate relationship of resource occupation. The computationally intensive occupation and gap filling occupation corresponding to different operator resource occupation relationships in the complementary candidate relationship are placed into the same graphics processor execution window. The video memory bandwidth occupation and video memory transportation occupation are placed in different occupation positions in the graphics processor execution window to form a complementary resource placement relationship. The overlapping status of complementary resource placement relationships within the graphics processor execution window is verified. Complementary resource placement relationships with overlapping occupancy are removed, and the remaining complementary resource placement relationships are used as the window-borne verification relationships. The resource complementarity placement relationships that do not overlap in the window carrying the verification relationship are merged into the same graphics processor execution window, and the operator resource occupation relationships carried by the resource complementarity placement relationship are merged into operator complementarity mixed relationship.

[0012] As a preferred embodiment of the machine learning resource regulation method based on fuzzy clustering described in this invention, the re-definition of the complementary mixed-part relationship of subsequent operators specifically involves: Based on the operator complementary mixed-part relationship, the occupied positions of the resource complementary placement relationship are read along the execution window of the graphics processor, and the occupied positions where the occupancy overlaps are marked as mixed-part failure positions; Remove the complementary resource placement relationship that has overlapping occupancy from the window placement result of the graphics processor execution window where the mixed part fails, and mark the operator resource occupation relationship that has overlapping occupancy at the mixed part failure location as the failure attribution relationship; When the failure attribution relationship shows that the operator resource occupancy relationship continuously fills the graphics processor execution window, the operator resource occupancy relationship is reversed to exclusive protection occupancy. When the failure attribution relationship shows that the operator resource occupancy relationship only has local occupancy overlap at the mixed failure position, the operator resource occupancy relationship is reversed to off-peak execution occupancy. When the failure attribution relationship shows that the operator resource occupancy relationship continuously destroys the complementary resource placement relationship, the operator resource occupancy relationship is reversed to prohibit mixed occupancy. The exclusive protection occupancy is locked in a separate graphics processor execution window. The off-peak execution occupancy is moved out of the mixed failure position and re-entered into the window position. The prohibited mixed occupancy is excluded from the complementary mixed relationship of subsequent operators and the complementary mixed relationship of subsequent operators is redefined.

[0013] As a preferred embodiment of the machine learning resource regulation method based on fuzzy clustering described in this invention, the graphics processor resource pool is a set of resources composed of multiple graphics processors, video memory resources, computing resources and task scheduling relationships, which are shared by machine learning tasks for operation. The operator startup order is the execution sequence formed by arranging the operator running events in the same machine learning task from earliest to latest according to the operator startup time. The graphics processor execution window is a parallel runtime segment defined on the same graphics processor for operators with complementary resource requirements. The computationally intensive occupancy, memory bandwidth occupancy, memory transport occupancy, and gap filling occupancy are the ownership states of operator resource occupancy relationships in the mixed-part affiliation state, which are characterized by computational execution dominance, memory bandwidth dominance, memory transport dominance, and the availability of waiting gaps for occupancy.

[0014] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the machine learning resource regulation method based on fuzzy clustering as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the machine learning resource regulation method based on fuzzy clustering as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By reversing the failure position constraint, the closed-loop performance of machine learning operators in mixed-part operations is improved. After the failure position is transformed into an inheritable window constraint, the graphics processor execution window no longer relies solely on the initial complementary relationship to carry operators, but can correct subsequent carrying boundaries based on the failure attribution result; operators that continuously fill the window obtain independent carrying space, locally overlapping operators obtain staggered release margin, and operators that continuously disrupt complementary placement are excluded from the subsequent candidate range, enabling different failure types to form differentiated convergence paths in subsequent regulation. During the continuous operation of the machine learning task, the operator resource occupancy relationship can be continuously corrected with mixed-part feedback, reducing the transmission of erroneous mixed-part relationships across windows, maintaining stable coordination between computational load, memory access, and execution flow occupancy, and improving the continuity, load stability, and regulation reliability of mixed-part operation within the graphics processor resource pool. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0018] Fig. 1 This is a flowchart of a machine learning resource regulation method based on fuzzy clustering.

[0019] Fig. 2 A flowchart for determining the membership status of mixed departments.

[0020] Fig. 3 This is a flowchart of the sequence formed by the operation of the operator.

[0021] Fig. 4 A flowchart for reversing the process in case of a failed mixing operation. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figs. 1-4 This is one embodiment of the present invention, which provides a machine learning resource regulation method based on fuzzy clustering, comprising the following steps: S1. Intercept operator execution events in the graphics processor resource pool, and establish the execution succession relationship between operator execution events according to the operator startup order to form an operator execution succession sequence.

[0026] S1.1. Extract operator execution events from the task execution events of the graphics processor resource pool, and label the operator execution events with machine learning task identifiers, operator start positions, operator end positions, and graphics processor occupancy positions. Specifically, Read task execution events generated by the graphics processor resource pool from the graphics processor driver runtime log, machine learning framework execution log, and graphics processor performance sampling log, and filter out events containing operator name, machine learning task identifier, start timestamp, end timestamp, graphics processor number, video memory address range, and execution flow number as operator execution events.

[0027] The graphics processor driver runtime log is recorded by the graphics processor driver during the execution of machine learning tasks, including the graphics processor number, kernel start timestamp, kernel end timestamp, execution flow number, and memory access address range; the machine learning framework execution log is output by machine learning frameworks such as PyTorch and TensorFlow during training and inference tasks, including machine learning task identifier, operator name, operator call order, and tensor transfer relationship; the graphics processor performance sampling log is collected by sampling tools such as NVIDIA CUPTI and DCGM during the execution of the graphics processor, including computational utilization, memory usage, memory bandwidth usage, memory transfer status, and sampling timestamp.

[0028] Using the machine learning task identifier in the machine learning framework execution log as the attribution basis, the start timestamp is written to the operator start position, the end timestamp is written to the operator end position, and the graphics processor number, video memory address range, and execution flow number are merged and written to the graphics processor occupied position. When the operator name, machine learning task identifier, start timestamp, and end timestamp all exist and the start timestamp is earlier than the end timestamp, the calibrated operator execution event is kept under the same machine learning task. When the machine learning task identifier is missing, the timestamp is inverted, or the graphics processor number is blank, the operator execution event is kept as an interception anomaly marker, and finally, an operator execution event with machine learning task identifier, operator start position, operator end position, and graphics processor occupied position is formed.

[0029] The graphics processor resource pool is a collection of resources composed of multiple graphics processors, video memory resources, computing resources, and task scheduling relationships, which are shared by machine learning tasks for running.

[0030] S1.2. Merge operator execution events based on the same machine learning task identifier, and arrange the merged operator execution events in ascending order of operator start position to form the operator start sequence for the same machine learning task. Specifically, Operator execution events with machine learning task identifiers, operator start positions, operator end positions, and GPU occupancy positions are used as merging objects. The machine learning task identifiers in the operator execution events are read, and operator execution events with the same machine learning task identifier and no truncation anomaly flag are written under the same machine learning task name. For operator execution events with the same machine learning task identifier but missing operator start positions, the sorting position is supplemented according to the operator call order in the machine learning framework execution log, and the start position supplementation flag is retained. For operator execution events with inconsistent machine learning task identifiers, the same operator execution event being written under different machine learning task names, the same operator start position and the operator names not distinguishing the order of priority, they are retained as merging anomaly flags. The operator start order is the result of the operator start position from earliest to latest, and the merged operator execution events are arranged sequentially under the same machine learning task name to form the operator start sequence of the same machine learning task.

[0031] Among them, the operator startup order is the execution sequence formed by arranging the operator running events in the same machine learning task from earliest to latest according to the operator startup time; the same machine learning task is the object to which the operator running events belong when the same machine learning task is identified and the training task or inference task is executed in the graphics processor resource pool.

[0032] S1.3 In the operator startup sequence, the operator end position of the previous operator running event and the operator start position of the next operator running event define the execution continuation segment. The execution continuation segment is written between adjacent operator running events to form an execution succession relationship. Specifically, Along the operator startup sequence of the same machine learning task, the operator running event that comes first is taken as the starting point, and the operator running event that immediately follows the starting point is taken as the ending point. The operator end position is read from the starting point, and the operator start position is read from the ending point. The time interval between the operator end position and the operator start position is divided into the running continuation segment. When the operator end position is earlier than the operator start position, the running continuation segment indicates that there is a waiting interval between adjacent operator running events. When the operator end position is equal to the operator start position, the running continuation segment indicates that adjacent operator running events are continuous. When the operator end position is later than the operator start position, the overlapping position is written into the continuation overlap mark. The running continuation segment is embedded between the starting point and the ending point, and the starting point, the running continuation segment, and the ending point are connected in series to form the execution continuation relationship.

[0033] S1.4. Operator execution events in the same machine learning task are connected by execution succession relationships to form an operator execution succession sequence.

[0034] S2. Extract the occupancy intensity set of each operator running event from the operator running sequence, synthesize the operator resource occupancy relationship, send the operator resource occupancy relationship into the fuzzy clustering model, classify the operator resource occupancy relationship according to the occupancy intensity set, and determine the mixed membership state of the operator resource occupancy relationship in computationally intensive occupancy, video memory bandwidth occupancy, video memory transport occupancy, and gap filling occupancy.

[0035] S2.1 Locate the execution sequence corresponding to a single operator execution event within the operator execution sequence, and delineate the resource occupation range from the operator start position to the operator end position; S2.2. Define the occupation boundary based on the resource occupation range. Within the occupation boundary, collect the execution intensity, transmission intensity, and gap intensity items of the graphics processor's occupied position to form an occupation intensity set. Specifically, The operator start position within the resource occupancy range is taken as the starting point of the occupancy boundary, and the operator end position within the resource occupancy range is taken as the ending point of the occupancy boundary. Sampling records with sampling timestamps between the starting and ending points of the occupancy boundary are selected from the graphics processor performance sampling log, and matched with the graphics processor occupancy position of the operator running event according to the graphics processor number, execution flow number, and memory address range. When a match is found, the computational utilization is written to the execution intensity item, the memory bandwidth usage, memory transfer status, and memory usage are written to the transfer intensity item, and the waiting interval, continuous continuation, and continuation overlap markers in the running sequence are written to the gap intensity item. When a match is not found, the operator running event is retained as an occupancy aggregation anomaly marker. The execution intensity item, transfer intensity item, and gap intensity item are merged into the same operator running event name to form an occupancy intensity set.

[0036] S2.3 Configure the ownership index of the occupancy intensity set for the operator running event, and establish the occupancy ownership relationship between the operator running event and the occupancy intensity set. Specifically, The machine learning task identifier, operator name, operator start position, operator end position, graphics processor number, and execution flow number in the operator execution event are concatenated to form an attribution index. The occupancy intensity set is then attached to the attribution index name. When the execution intensity item, transit intensity item, and gap intensity item are all formed in the occupancy intensity set, the attribution index is written into the operator execution event. Through the attribution index, operator execution events with the same machine learning task identifier, operator name, operator start position, operator end position, graphics processor number, and execution flow number are bound one-to-one with the occupancy intensity set to form an occupancy attribution relationship. If any of the execution intensity item, transit intensity item, or gap intensity item is missing from the occupancy intensity set, a missing intensity item flag is retained. If the attribution index appears repeatedly and the operator start position, operator end position, graphics processor number, and execution flow number cannot distinguish the attribution, a conflict flag for the attribution index is retained.

[0037] S2.4. Based on the occupancy ownership relationship and the connection position between the operation sequence segments, add occupancy ownership relationships to adjacent operator operation events within the operator operation sequence to synthesize operator resource occupancy relationships. Specifically, Adjacent operator execution events that have already formed execution inheritance relationships in the operator execution inheritance sequence are used as supplementary objects. The inheritance start point and inheritance end point are read along the execution continuation segment, and the inheritance intensity set ownership index under the inheritance start point name and the inheritance intensity set ownership index under the inheritance end point name are searched in the occupation ownership relationship. When both the inheritance start point and the inheritance end point have occupation ownership relationships, the occupation ownership relationship of the inheritance start point is written to the front inheritance position of the execution continuation segment, and the occupation ownership relationship of the inheritance end point is written to the back inheritance position of the execution continuation segment. When the inheritance start point lacks an occupation ownership relationship, the inheritance end point lacks an occupation ownership relationship, or the execution continuation segment has a continuation overlap mark, the inheritance supplementation abnormal mark is retained. The occupation ownership relationships in the inheritance start point, the front inheritance position, the execution continuation segment, the back inheritance position, and the inheritance end point are merged to form the operator resource occupation relationship.

[0038] S2.5. Based on the occupancy attribution relationship in the operator resource occupancy relationship, retrieve the occupancy intensity set associated with the operator running events, and arrange the occupancy intensity sets according to the order of the operator running events in the operator running sequence to form the occupancy sequence to be divided. Specifically, Extract the occupancy attribution relationships from the operator resource occupancy relationships item by item, including the starting point, ending point, front-end and back-end positions. Find the occupancy intensity set bound to the operator running event according to the attribution index within the occupancy attribution relationship. Use the sequential position of the operator running events in the operator running sequence as the arrangement position, placing the occupancy intensity set associated with the starting point in the previous arrangement position and the occupancy intensity set associated with the ending point in the next arrangement position. When the same operator running event appears repeatedly in adjacent operator resource occupancy relationships, retain the original sequential position in the operator running sequence and merge the duplicate attribution markers. When the occupancy attribution relationship is missing, the attribution index does not find the occupancy intensity set, or the arrangement position of the occupancy intensity set is out of the operator running sequence, retain the arrangement anomaly marker. Concatenate the arranged occupancy intensity sets according to the sequential position of the operator running sequence to form the occupancy sequence to be divided.

[0039] S2.6. The occupancy sequence to be partitioned is fed into a fuzzy clustering model for membership partitioning, resulting in the membership partitioning results. Specifically, S2.6.1 Within the occupancy sequence to be partitioned, the execution intensity items, transit intensity items, and gap intensity items of the occupancy intensity set are arranged along the attribution position of the operator running events to form the membership input sequence, specifically, The occupancy intensity sets that have been arranged in the occupancy sequence to be divided are expanded sequentially according to their ownership positions. Execution intensity items under the same ownership position are placed in the execution input position, transmission intensity items are placed in the transmission input position, and gap intensity items are placed in the gap input position. When reading execution intensity items, transmission intensity items, and gap intensity items from the occupancy intensity set, the ownership position of the occupancy intensity set in the occupancy sequence to be divided is retained, ensuring that the execution intensity items, transmission intensity items, and gap intensity items are arranged in the same position as the operator running events. When all execution intensity items, transmission intensity items, and gap intensity items exist, the execution input position, transmission input position, and gap input position are merged into a corresponding input group and connected to the next ownership position. When there are missing intensity items, misaligned intensity items, or the occupancy intensity set is removed from the occupancy sequence to be divided at the ownership position, the membership input anomaly flag is retained. The corresponding input groups arranged continuously according to ownership positions are concatenated to form a membership input sequence.

[0040] Among them, the execution intensity term is characterized by the computational utilization and execution flow persistence of the operator running event on the graphics processor's computing resources; the transfer intensity term is characterized by the memory bandwidth usage, memory transfer status and memory usage of the operator running event during memory access and memory transfer; and the gap intensity term is characterized by the waiting interval, continuous continuation and continuation overlap mark in the running sequence segment, representing the gap state that can be occupied between adjacent operator running events.

[0041] S2.6.2. The membership input sequence is fed into the fuzzy clustering model. The occupancy reference cluster of the fuzzy clustering model is constructed based on computational occupancy, memory bandwidth occupancy, memory transfer occupancy, and gap filling occupancy. Specifically, After the membership input sequence is fed into the fuzzy clustering model, the fuzzy clustering model first reads the execution input bits, transmission input bits, and gap input bits in the same input group in the input sorting layer. The feature scale processing layer uses the sampling values ​​within the same occupancy boundary as the scale benchmark, converts the computation utilization rate in the execution input bits into the computation occupancy ratio, converts the video memory bandwidth occupancy in the transmission input bits into the bandwidth occupancy ratio, converts the video memory transport status in the transmission input bits into the transport occupancy ratio according to the transport duration sampling number, and converts the waiting interval and continuous position in the gap input bits into the gap carrying ratio according to the accommodation relationship between the running continuous segment duration and the execution duration of the subsequent operator running event. The computation occupancy ratio, bandwidth occupancy ratio, transport occupancy ratio, and gap carrying ratio constitute the clustering feature quantity with the same value scale.

[0042] The computational occupancy ratio is formed by sampling the computational utilization within the same occupancy boundary. The feature scale processing layer uses the upper limit of the computational capacity of the graphics processor execution window as a unified scale to convert the computational utilization sampling value into a computational occupancy ratio. The bandwidth occupancy ratio is formed by the amount of video memory bandwidth occupied in the transmitted input bits. The sampling interval in which the video memory bandwidth occupancy is continuously high within the same occupancy boundary is unified by the bandwidth capacity boundary of the graphics processor execution window to form the bandwidth occupancy ratio corresponding to the video memory bandwidth occupancy. The transport occupancy ratio is generated by the continuous sampling positions of the video memory transport state within the same occupancy boundary. The feature scale processing layer aligns the transport duration interval covered by the continuous sampling positions with the transport capacity boundary of the graphics processor execution window to form the transport occupancy ratio.

[0043] After clustering features enter the membership partitioning layer, the following clustering features are considered: those with consistently high computational utilization in the reference cluster receiving execution input bits; those with consistently high memory bandwidth occupancy in the reference cluster receiving transmission input bits; those with continuously occurring memory transport states in the reference cluster receiving transmission input bits; and those that pass the integer operator carrying verification in the reference cluster receiving gap input bits. Integer carrying verification refers to the following: the running continuation segment does not have a continuation overlap marker; the start and end boundaries of the waiting interval cover the operator start position to the operator end position of the subsequent operator running event; the memory address range does not form an occupational intersection with the subsequent operator running event; and the execution flow number can form a continuation with the subsequent operator running event. When the subsequent operator running event is entirely within the waiting interval, the corresponding clustering feature is retained as the input source for the gap filling reference cluster. When the subsequent operator running event crosses the waiting interval boundary, the subsequent operator running event is not split, and the corresponding input group is retained as a gap uncarryable marker.

[0044] The same input group belonging to the input anomaly marker does not participate in the generation of the occupancy reference cluster. The same input groups with complete execution input bits, transmission input bits, and gap input bits enter the computationally intensive occupancy reference cluster, memory bandwidth occupancy reference cluster, memory transport occupancy reference cluster, and gap filling occupancy reference cluster, respectively, forming the occupancy reference cluster of the fuzzy clustering model.

[0045] The training process for the fuzzy clustering model is as follows: The fuzzy clustering model includes an input organization layer, a feature scale processing layer, a membership partitioning layer, and an occupancy reference cluster output layer. The input organization layer receives input groups with clearly marked occupancy relationships from the historical operator execution sequence. The feature scale processing layer converts the execution intensity term, transmission intensity term, and gap intensity term into a single value scale. The membership partitioning layer uses a fuzzy C-means clustering algorithm. It first reads the distance values ​​between the input groups and the computationally intensive occupancy reference cluster, the memory bandwidth occupancy reference cluster, the memory transport occupancy reference cluster, and the gap filling occupancy reference cluster. Then, it assigns the membership values ​​of the input groups to the four occupancy reference clusters based on the relative magnitude of the distance values. Occupancy reference clusters with smaller distance values ​​receive higher membership values, thus forming the membership values ​​for the input groups entering the computationally intensive occupancy, memory bandwidth occupancy, memory transport occupancy, and gap filling occupancy reference clusters. The membership values ​​are determined by the output layer, which outputs four occupancy reference clusters. The training samples are taken from historical operator running events that have formed occupancy intensity sets in the graphics processor driver running log, machine learning framework execution log, and graphics processor performance sampling log. The training inputs are the execution intensity term, the transmission intensity term, and the gap intensity term. The training result is the membership values ​​of the same input group entering the four occupancy reference clusters. During training, the number of occupancy reference clusters is initially set to four, and the fuzzy coefficient adopts a common value greater than one, such as the example value of two. The loss function is the weighted sum of squared distances from the same input group to the center of the occupancy reference cluster. The membership division layer alternately updates the membership values ​​and the center of the occupancy reference cluster. The update stops when the change in the center of the occupancy reference cluster is less than one percent of the average distance of the historical operator running events, resulting in a fuzzy clustering model for membership division.

[0046] Among them, computationally intensive usage, memory bandwidth usage, memory transport usage, and gap filling usage are the ownership states of operator resource usage relationships in the mixed-part affiliation state, which are dominated by computation execution, memory bandwidth, memory transport, and waiting gaps.

[0047] S2.6.3. Determine the membership values ​​of the input sequence in each occupied reference cluster to form fuzzy membership relationships. Specifically, The peer input groups in the peer input sequence are sent to the peer partitioning layer according to their membership positions. The distance values ​​between the peer input group and the computationally intensive occupancy reference cluster, the memory bandwidth occupancy reference cluster, the memory transport occupancy reference cluster, and the gap filling occupancy reference cluster are read respectively. After the distance values ​​are read, the peer partitioning layer assigns membership values ​​in ascending order of distance values. The occupancy reference cluster with the earlier distance value is assigned a higher membership value, and the occupancy reference cluster with the later distance value is assigned a lower membership value. The sum of the four membership values ​​under the same peer input group is made to be one. When the peer input group has a membership input exception flag, it does not participate in the membership value allocation and retains the membership determination exception flag. When the execution input bit, transmission input bit, and gap input bit of the peer input group are all complete, the membership values ​​of the peer input group entering the computationally intensive occupancy, memory bandwidth occupancy, memory transport occupancy, and gap filling occupancy are written to the membership positions to form a fuzzy membership relationship.

[0048] S2.6.4. According to the fuzzy attribution relationship, rearrange the attribution positions of the occupied reference clusters in the membership input sequence to form the membership cluster order result, specifically, The system reads the membership positions item by item along the membership input sequence and reads the membership values ​​corresponding to the same membership position in the fuzzy membership relationship. Using the membership position as the rearrangement position and the membership value as the rearrangement basis, it adjusts the order of the computationally intensive reference clusters, memory bandwidth reference clusters, memory transport reference clusters, and gap filling reference clusters under the same membership position, forming an arrangement of the occupancy reference clusters corresponding to the membership position. When a membership position has an abnormal membership determination flag, the cluster order abnormality flag is retained, and the rearrangement of the occupancy reference clusters corresponding to the membership position is stopped. The rearranged occupancy reference clusters are then matched with the membership positions in the membership input sequence to form a membership cluster order result containing the cluster order position and the occupancy reference clusters.

[0049] S2.6.5. Backfill the cluster order positions of the membership cluster order results into the occupancy sequence to be partitioned, indicating the partitioning assignment of the occupancy sequence to be partitioned into computationally intensive occupancy, memory bandwidth occupancy, memory transport occupancy, and gap filling occupancy, as the membership partitioning result. Specifically, The cluster order position and occupied reference cluster in the membership cluster order result are read item by item. According to the correspondence between the cluster order position and the operator running event position in the occupancy sequence to be divided, the occupied reference cluster is backfilled into the occupancy sequence to be divided. When the cluster order position is consistent with the operator running event position, the order of the occupied reference clusters under the same membership position is retained, and the front-end occupied reference cluster is used as the partitioning membership of the occupancy sequence to be divided in the same membership position. When there is a cluster order abnormality mark, a missing cluster order position, or a cluster order position that cannot correspond to the occupancy sequence to be divided, the membership position is retained as a partitioning backfilling abnormal mark. After the backfilling is completed, the occupancy sequence to be divided obtains the partitioning membership of entering computationally intensive occupancy, video memory bandwidth occupancy, video memory transport occupancy, and gap filling occupancy, forming the membership partitioning result.

[0050] S2.7. Obtain the execution intensity item, transfer intensity item, and gap intensity item from the occupancy intensity set. According to the membership partitioning results, associate the execution intensity item with computationally intensive occupancy, associate the transfer intensity item with memory bandwidth occupancy and memory transport occupancy, and associate the gap intensity item with gap filling occupancy to form the initial mixed membership state. Specifically, The attribution position without a partition backfilling anomaly marker in the membership partitioning result is taken as the state association position, and the execution intensity item, transfer intensity item, and gap intensity item are read from the occupancy intensity set corresponding to the state association position. Within the state association position, based on the partitioning attribution of the same attribution position, the execution intensity item is associated with the computationally intensive occupancy, the memory bandwidth occupancy in the transfer intensity item is associated with the memory bandwidth occupancy, the memory transport status in the transfer intensity item is associated with the memory transport occupancy, and the gap intensity item is associated with the gap filling occupancy. When the attribution position in the membership partitioning result has a partition backfilling anomaly marker, the corresponding attribution position is retained as the initial association anomaly marker. When the occupancy intensity set corresponding to the state association position has either an intensity item missing marker or an attribution index conflict marker, the state association position is retained as the initial association anomaly marker. After completing the state association between the intensity item and the occupancy status within the state association position, the computationally intensive occupancy association status and the memory bandwidth occupancy association status are combined into a front-end association status, and the memory transport occupancy association status and the gap filling occupancy association status are combined into a back-end association status. The front-end association status and the back-end association status form the initial mixed membership status.

[0051] S2.8. Within the adjacent operator running events covered by the running continuation segment, the initial mixed membership states are concatenated according to their order in the operator running succession sequence to form mixed membership states, specifically as follows: Along the starting and ending points covered by the running continuation segment, the initial mixed-part membership status under the name of the starting point is placed at the front end of the running continuation segment, and the initial mixed-part membership status under the name of the ending point is placed at the back end of the running continuation segment; the waiting interval, continuous continuation, and continuation overlap markers in the running continuation segment are retained between the front end and the back end; the serial connection positions with initial classification exception markers are retained as mixed-part serial connection exception markers; the front end, running continuation segment, and back end position are serialized according to the order of the operator running continuation sequence to form the mixed-part membership status.

[0052] S2.9. The mixed membership state is used to constrain the allocation of operator resource usage relationships among computationally intensive usage, video memory bandwidth usage, video memory transport usage, and gap filling usage, thereby determining the mixed membership state of operator resource usage relationships. Specifically, The front-end position, running continuation segment, and back-end position that have been connected in the mixed-part membership state are merged into the same operator resource occupancy relationship name; the front-end position and the back-end position retain the computationally intensive occupancy membership value, the video memory bandwidth occupancy membership value, the video memory transport occupancy membership value, and the gap filling occupancy membership value, respectively; the running continuation segment retains the waiting interval, continuous continuation, and continuation overlap markers; the mixed-part connection abnormal marker is retained with the original membership position; after the membership values ​​of the operator resource occupancy relationship in the four occupancy memberships are merged, the mixed-part membership state of the operator resource occupancy relationship is determined.

[0053] S3. According to the mixed membership status, place the resource occupancy relationship of operators with complementary resource requirements into the same graphics processor execution window to form a complementary mixed relationship of operators.

[0054] S3.1. From the mixed membership states, different combinations of operator resource occupancy relationships are selected, and combinations of different operator resource occupancy relationships where computationally intensive occupancy and gap-filling occupancy coexist and memory bandwidth occupancy and memory transport occupancy are staggered are selected as candidate relationships for complementary occupancy relationships. Specifically, Select distinct first and second operator resource occupancy relationships from the mixed membership states. Extract the computationally intensive occupancy membership values, memory bandwidth occupancy membership values, memory transport occupancy membership values, and gap filling occupancy membership values ​​under the names of the first and second operator resource occupancy relationships respectively. Compare the occupancy at the front end of the first operator resource occupancy relationship, the running continuation segment of the second operator resource occupancy relationship, and the back end of the second operator resource occupancy relationship. If the computationally intensive occupancy membership value of the first operator resource occupancy relationship is at the top of the four membership values ​​at the front end, and the gap filling occupancy membership value of the second operator resource occupancy relationship is at the top of the four membership values ​​at the back end, and if the second operator resource occupancy relationship retains a waiting interval between the front end and the back end that can accommodate the first operator resource occupancy relationship, then mark the first and second operator resource occupancy relationships as having a window carrying relationship. Within the combination of operator resource occupancy relationships with a shared window carrying relationship, the attribution position of memory bandwidth occupancy and the attribution position of memory transport occupancy are compared. If the attribution positions of memory bandwidth occupancy and memory transport occupancy are distributed under different operator execution event names, and the first position of the membership value of memory bandwidth occupancy does not coincide with the first position of the membership value of memory transport occupancy, then memory bandwidth occupancy and memory transport occupancy are marked as misaligned. If the first operator resource occupancy relationship and the second operator resource occupancy relationship are the same, either operator resource occupancy relationship has a mixed concatenation anomaly mark, the first position of memory bandwidth occupancy and memory transport occupancy coincides, or the first operator resource occupancy relationship and the second operator resource occupancy relationship do not form a shared window carrying relationship, then the corresponding operator resource occupancy relationship combination will not enter the occupancy complementarity candidate relationship. Operator resource occupancy relationship combinations that simultaneously satisfy the shared window carrying relationship and the misalignment of memory bandwidth occupancy and memory transport occupancy are screened out as occupancy complementarity candidate relationships.

[0055] Among them, the resource occupancy relationship of a single operator is selected as a complete scheduling object and is not split into computational carrying object and gap carrying object in the complementary occupancy candidate relationship; computationally intensive occupancy and gap filling occupancy have the same window carrying relationship, which means that the resource occupancy relationships of different operators within the same graphics processor execution window respectively present the first position of computationally intensive occupancy and the first position of gap filling occupancy, and the carrying boundary is provided by the waiting interval in the running continuation segment.

[0056] S3.2. For the complementary resource allocation candidate relationships, the computationally intensive resource allocation and gap-filling resource allocation corresponding to different operator resource allocation relationships are placed into the same graphics processor execution window. Furthermore, the video memory bandwidth allocation and video memory transport allocation are placed in different allocation positions within the graphics processor execution window, thus forming a complementary resource allocation relationship. Specifically… The resource occupancy relationships of the first and second operators, as well as the execution continuation segments within the second operator resource occupancy relationship, are brought into the graphics processor execution window. First, the same graphics processor execution window is selected based on the graphics processor number corresponding to both the first and second operator resource occupancy relationships. Then, the computationally intensive occupancy in the first operator resource occupancy relationship is placed in the computational load position within the graphics processor execution window, and the gap-filling occupancy in the second operator resource occupancy relationship is placed in the gap-filling position corresponding to the waiting interval within the graphics processor execution window. This ensures that the computationally intensive occupancy and gap-filling occupancy in different operator resource occupancy relationships are within the same graphics processor execution window; memory bandwidth. The bandwidth occupancy position is written to the graphics processor execution window, and the memory transport occupancy position is written to the graphics processor execution window. The bandwidth occupancy position and the transport occupancy position are distinguished by the sampling timestamp, memory address range, and execution flow number. When the complementary candidate relationship lacks a graphics processor number, the graphics processor number of the first operator resource occupancy relationship and the second operator resource occupancy relationship are inconsistent, the running continuation segment of the second operator resource occupancy relationship has a continuation overlap mark, or the bandwidth occupancy position and the transport occupancy position coincide, the window placement abnormal mark is retained. After the placement of the calculation carrying position, gap carrying position, bandwidth occupancy position, and transport occupancy position within the graphics processor execution window is completed, a resource complementary placement relationship is formed.

[0057] The graphics processor execution window is a parallel runtime segment defined on the same graphics processor for the resource occupancy relationship of operators with complementary resource requirements. The candidate relationship for complementary occupancy is formed by combining different operator resource occupancy relationships. When the window is placed, only the computationally intensive occupancy and gap filling occupancy corresponding to different operator resource occupancy relationships are placed into the same graphics processor execution window. No individual operator resource occupancy relationship is split and placed.

[0058] S3.3. Verify the overlapping status of complementary resource placement relationships within the graphics processor execution window, remove resource complementary placement relationships with overlapping occupancy, and use the remaining resource complementary placement relationships as the window-carrying verification relationships. Specifically, The computational carrying positions, gap carrying positions, bandwidth occupying positions, and transport occupying positions in the resource complementarity placement relationship are arranged into the window occupancy table of the same graphics processor execution window. The window occupancy table uses the sampling timestamp, memory address range, and execution flow number to indicate the start and end boundaries of each occupied position. The computational carrying positions and gap carrying positions are kept in the same graphics processor execution window, while the bandwidth occupying positions and transport occupying positions are listed in different occupied positions. Resource complementarity placement relationships where the sampling timestamps intersect, the memory address ranges intersect, and the execution flow numbers fall into the same occupied position are marked as overlapping occupancy relationships and removed from the graphics processor execution window. Resource complementarity placement relationships without overlapping occupancy relationships and window placement anomaly markers are kept in the graphics processor execution window as window carrying verification relationships.

[0059] S3.4. Merge the resource complementarity placement relationships that do not overlap in the window-carrying verification relationships into the same graphics processor execution window, and combine the operator resource occupation relationships carried by the resource complementarity placement relationships into operator complementary mixed relationships. Specifically, Resource complementary placement relationships without overlapping or abnormal window placement markers in the window carrying verification relationships are grouped into the same graphics processor execution window according to the graphics processor number, the start and end boundaries of the graphics processor execution window, and the execution flow number. Within the graphics processor execution window, the placement order of computational carrying position, gap carrying position, bandwidth occupancy position, and transport occupancy position is retained, and the operator resource occupancy relationship carried by the resource complementary placement relationship is written into its respective carrying position. Resource complementary placement relationships with overlapping, abnormal window placement markers, and abnormal acceptance / filling markers are not merged into the graphics processor execution window and are retained as mixed-part candidate exclusion markers. After window merging is completed, the operator resource occupancy relationships within the same graphics processor execution window are merged according to the carrying relationships of computationally intensive occupancy accepting gap filling occupancy and memory bandwidth occupancy avoiding memory transport occupancy, forming operator complementary mixed-part relationships.

[0060] S4. When a mixed-part failure occurs in the operator complementary mixed-part relationship, the operator resource occupancy relationship corresponding to the mixed-part failure position is reversed to any of the following states: exclusive protection occupancy, off-peak execution occupancy, and prohibited mixed-part occupancy. The subsequent operator complementary mixed-part relationships are then redefined according to the reversed operator resource occupancy relationship.

[0061] S4.1. Based on the complementary mixing relationship of operators, read the occupied positions of the complementary placement relationship of resources along the execution window of the graphics processor, and mark the occupied positions where overlapping occurs as mixing failure positions. Specifically, The resource complementarity placement relationships that have already merged in the operator complementarity mixed-part relations are loaded into the mixed-part verification table of the graphics processor execution window. The mixed-part verification table retains the graphics processor number, the start and end boundaries of the graphics processor execution window, the computational carrying position, the gap carrying position, the bandwidth occupied position, the transport occupied position, the sampling timestamp, the video memory address range, and the execution flow number. The computational carrying position, gap carrying position, bandwidth occupied position, and transport occupied position within the same graphics processor execution window are expanded according to the sampling timestamp range, and the video memory address range and the execution flow number are appended to the corresponding occupied position. Occupied positions where the sampling timestamp ranges overlap and the video memory address ranges overlap, or occupied positions where the sampling timestamp ranges overlap and the execution flow number falls into the same execution flow, are marked with an occupation overlap mark. Resource complementarity placement relationships that do not have an occupation overlap mark are retained in the operator complementarity mixed-part relations, and occupied positions with an occupation overlap mark are written to the mixed-part failure position to form a mixed-part failure position.

[0062] S4.2. Remove the resource complementarity placement relationship that causes overlapping occupancy from the window placement result of the graphics processor execution window where the mixing failure location is located, and mark the operator resource occupancy relationship that participates in the overlapping occupancy at the mixing failure location as the failure attribution relationship. Specifically, Retrieve the graphics processor number, graphics processor execution window start and end boundaries, occupied location name, sampling timestamp, video memory address range, and execution flow number from the failed location. Then, locate the computational carrying location, gap carrying location, bandwidth occupying location, and transport occupying location with the occupation overlap mark from the resource complementary placement relationship already incorporated in the same graphics processor execution window. Delete the resource complementary placement relationship carried by the occupied location with the occupation overlap mark from the graphics processor execution window, while the resource complementary placement relationship that did not participate in the occupation overlap remains in the graphics processor execution window. Paste the failed location back to the name of the operator resource occupation relationship that participated in the occupation overlap, and write the graphics processor number, graphics processor execution window start and end boundaries, occupation overlap mark, and the deleted resource complementary placement relationship into the operator resource occupation relationship that participated in the occupation overlap to form a failure attribution relationship.

[0063] S4.3 When the failure attribution relationship shows that the operator resource occupancy relationship continuously fills the graphics processor execution window, the operator resource occupancy relationship is reversed to exclusive protection occupancy. When the failure attribution relationship shows that the operator resource occupancy relationship only has local occupancy overlap at the mixed failure position, the operator resource occupancy relationship is reversed to off-peak execution occupancy. When the failure attribution relationship shows that the operator resource occupancy relationship continuously destroys the resource complementary placement relationship, the operator resource occupancy relationship is reversed to prohibit mixed occupation.

[0064] Among them, the failure attribution relationship is characterized by the operator resource occupancy relationship continuously filling the graphics processor execution window. This means that the same operator resource occupancy relationship has an overlapping occupancy mark between the start boundary and the end boundary of the graphics processor execution window, and continuously occupies the main bearing position among the computing bearing position, bandwidth bearing position, and transport bearing position. Exclusive protection occupancy means that the operator resource occupancy relationship that continuously fills the graphics processor execution window is kept separately in an independent graphics processor execution window and is no longer shared with other operator resource occupancy relationships.

[0065] The failure attribution relationship is characterized by the operator resource occupancy relationship only having local overlap at the mixed failure position. This means that the same operator resource occupancy relationship only has an overlap mark within the sampling timestamp interval, video memory address interval, and execution flow number range corresponding to the mixed failure position, while other carrying positions within the graphics processor execution window remain in a mixed carrying state. The off-peak execution occupancy is to move the operator resource occupancy relationship that only has local overlap at the mixed failure position out of the overlapping sampling timestamp interval and reschedule it to a carrying position within the same graphics processor execution window where no overlap occurs.

[0066] Failure attribution indicates that the continuous disruption of the resource complementarity placement relationship by the operator resource occupancy relationship is caused by the same operator resource occupancy relationship repeatedly forming overlapping occupancy marks in multiple adjacent graphics processor execution windows, resulting in the continuous deletion of the resource complementarity placement relationship from the graphics processor execution window; prohibiting mixed occupancy means excluding the operator resource occupancy relationship that continuously disrupts the resource complementarity placement relationship from the subsequent occupancy complementarity candidate relationship, and no longer participating in the mixed placement of the subsequent graphics processor execution window.

[0067] S4.4. Lock the exclusive protection occupancy to a separate graphics processor execution window, move the off-peak execution occupancy out of the mixed-partition failure position and re-participate in the window position, exclude the prohibited mixed-partition occupancy from the subsequent operator complementary mixed-partition relationship, and re-limit the subsequent operator complementary mixed-partition relationship. Specifically, The exclusive protection occupancy, off-peak execution occupancy, and prohibited mixed-partition occupancy that have been reversed in the failure attribution relationship are imported into the subsequent window limitation table. The subsequent window limitation table retains the operator resource occupancy relationship, graphics processor number, graphics processor execution window start and end boundaries, mixed-partition failure position, occupancy overlap mark, and occupancy status after reversal. The exclusive protection occupancy is locked to a separate graphics processor execution window, and the computational bearing position, bandwidth occupancy position, and transport occupancy position of the operator resource occupancy relationship are retained in the separate graphics processor execution window. The separate graphics processor execution window does not accept other occupancy complementary candidate relationships. The off-peak execution occupancy is moved out of the sampling timestamp interval, video memory address interval, and execution flow number corresponding to the mixed-partition failure position, and transferred to the bearing position without occupancy overlap mark in the same graphics processor execution window, and participates in window placement again. The prohibited mixed-partition occupancy is written into the subsequent mixed-partition exclusion list, and the operator resource occupancy relationship corresponding to the prohibited mixed-partition occupancy is excluded in the subsequent occupancy complementary candidate relationship screening. After completing the exclusive locking, off-peak rearrangement, and prohibition exclusion, the subsequent operator complementary mixed-partition relationship only retains the operator resource occupancy relationship that can continue to be mixed-partitioned, and the subsequent operator complementary mixed-partition relationship is re-limited.

[0068] The subsequent window constraint table is used to accept the results of the failed attribution relationship reversal and to limit whether exclusive protection occupancy, off-peak execution occupancy, and prohibited mixed occupancy continue to participate in mixed placement in the subsequent graphics processor execution window; the subsequent operator complementary mixed placement relationship is a set of operator resource occupancy relationships that can still be jointly carried in the graphics processor execution window in a resource complementary manner after being re-filtered according to the subsequent window constraint table.

[0069] This embodiment also provides a computer device applicable to the machine learning resource regulation method based on fuzzy clustering, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the machine learning resource regulation method based on fuzzy clustering as proposed in the above embodiment.

[0070] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0071] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the machine learning resource control method based on fuzzy clustering as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0072] In summary, this invention improves the closed-loop performance of mixed-partition operations in machine learning by reversing the failure position constraint. After the failure position is transformed into an inheritable window constraint, the GPU execution window no longer relies solely on the initial complementary relationship to carry operators, but can correct subsequent carrying boundaries based on the failure attribution result. Operators that continuously fill the window gain independent carrying space, locally overlapping operators gain staggered release margin, and operators that continuously disrupt complementary placements are excluded from the subsequent candidate range, enabling different failure types to form differentiated convergence paths in subsequent regulation. During continuous operation of the machine learning task, the operator resource occupancy relationship can be continuously corrected with mixed-partition feedback, reducing the propagation of erroneous mixed-partition relationships across windows, maintaining stable coordination between computational load, memory access, and execution flow occupancy, and improving the continuity, load stability, and regulation reliability of mixed-partition operation within the GPU resource pool.

[0073] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A machine learning resource regulation method based on fuzzy clustering, characterized in that: include, The operator execution events are intercepted in the graphics processor resource pool, and the execution succession relationship between the operator execution events is established according to the operator startup order to form an operator execution succession sequence. Extract the occupancy intensity set of each operator running event from the operator running sequence, synthesize the operator resource occupancy relationship, send the operator resource occupancy relationship into the fuzzy clustering model, classify the operator resource occupancy relationship according to the occupancy intensity set, and determine the mixed membership state of the operator resource occupancy relationship in computationally intensive occupancy, video memory bandwidth occupancy, video memory transport occupancy, and gap filling occupancy. According to the mixed-part affiliation status, the resource occupancy relationship of operators with complementary resource requirements is placed into the same graphics processor execution window to form a complementary mixed-part relationship of operators. When a mixed-part relationship of complementary operators fails, the resource occupancy relationship of the operator corresponding to the failed position is reversed to any of the following states: exclusive protection occupancy, off-peak execution occupancy, and prohibited mixed-part occupancy. The subsequent complementary mixed-part relationships of operators are then redefined according to the reversed operator resource occupancy relationship.

2. The machine learning resource regulation method based on fuzzy clustering as described in claim 1, characterized in that: The formation of the operator execution sequence is specifically as follows: Extract operator execution events from the task execution events of the graphics processor resource pool, and label the operator execution events with machine learning task identifier, operator start position, operator end position, and graphics processor occupied position; The operator execution events are merged based on the same machine learning task identifier, and the merged operator execution events are arranged in order from earliest to latest according to the operator start position to form the operator start time sequence of the same machine learning task. In the operator startup sequence, the operator end position of the previous operator running event and the operator start position of the next operator running event define the running continuation segment. The running continuation segment is written between adjacent operator running events to form an execution succession relationship. Operator execution events in the same machine learning task are connected by execution succession relationships to form an operator execution succession sequence.

3. The machine learning resource regulation method based on fuzzy clustering as described in claim 2, characterized in that: The resource occupancy relationship of the synthesis operator is specifically as follows: Locate the execution sequence corresponding to a single operator execution event within the operator execution sequence, and delineate the resource occupation range from the operator start position to the operator end position; The occupation boundary is defined by the resource occupation range. Within the occupation boundary, the execution intensity item, transmission intensity item, and gap intensity item of the graphics processor occupation position are collected to form an occupation intensity set. Configure the occupancy intensity set ownership index for operator running events, and establish the occupancy ownership relationship between operator running events and occupancy intensity sets; Based on the position of the occupancy relationship and the succession segment, the occupancy relationship is added to the adjacent operator running events in the operator running succession sequence to synthesize the operator resource occupancy relationship.

4. The machine learning resource regulation method based on fuzzy clustering as described in claim 3, characterized in that: The determination of operator resource occupancy relationships in the mixed membership state of computationally intensive occupancy, memory bandwidth occupancy, memory transport occupancy, and gap filling occupancy is specifically as follows: Based on the occupancy attribution relationship in the operator resource occupancy relationship, the occupancy intensity set associated with the operator running event is retrieved, and the occupancy intensity set is arranged according to the order of the operator running events in the operator running succession sequence to form the occupancy sequence to be divided; The occupancy sequence to be divided is fed into a fuzzy clustering model for membership division, resulting in membership division results; The execution intensity item, transfer intensity item, and gap intensity item are obtained from the occupancy intensity set. According to the membership partitioning results, the execution intensity item is associated with computationally intensive occupancy, the transfer intensity item is associated with memory bandwidth occupancy and memory transport occupancy, and the gap intensity item is associated with gap filling occupancy, forming an initial mixed membership state. Within the adjacent operator running events covered by the running continuation segment, the initial mixed membership states are concatenated according to the order of the operator running succession sequence to form mixed membership states; The mixed membership state is used to constrain the allocation of operator resource occupancy relationships among computationally intensive occupancy, video memory bandwidth occupancy, video memory transport occupancy, and gap filling occupancy, thereby determining the mixed membership state of operator resource occupancy relationships.

5. The machine learning resource regulation method based on fuzzy clustering as described in claim 4, characterized in that: The formation of the membership division result is specifically as follows: Within the occupancy sequence to be divided, the execution intensity items, transmission intensity items, and gap intensity items of the occupancy intensity set are arranged along the location of the operator running event to form the membership input sequence; The membership input sequence is fed into the fuzzy clustering model, and the occupancy reference cluster of the fuzzy clustering model is constructed by computationally intensive occupancy, memory bandwidth occupancy, memory transfer occupancy, and gap filling occupancy. The membership values ​​of the input sequence in each occupied reference cluster are determined to form fuzzy membership relationships; According to the fuzzy attribution relationship, the attribution positions of the occupied reference clusters in the membership input sequence are rearranged to form the membership cluster order result; The cluster order position of the membership cluster order result is backfilled into the occupancy sequence to be divided, and the occupancy sequence to be divided is marked as belonging to the division of computationally intensive occupancy, memory bandwidth occupancy, memory transport occupancy and gap filling occupancy, which is used as the membership partitioning result.

6. The machine learning resource regulation method based on fuzzy clustering as described in claim 5, characterized in that: The formation of complementary hybrid part relations of operators specifically refers to... Different combinations of operator resource occupancy relationships are selected from the mixed membership states, and combinations of different operator resource occupancy relationships that coexist with computationally intensive occupancy and gap filling occupancy, and whose memory bandwidth occupancy and memory transport occupancy are staggered, are selected as candidate occupancy complementary relationships. The execution window is positioned for the complementary candidate relationship of resource occupation. The computationally intensive occupation and gap filling occupation corresponding to different operator resource occupation relationships in the complementary candidate relationship are placed into the same graphics processor execution window. The video memory bandwidth occupation and video memory transportation occupation are placed in different occupation positions in the graphics processor execution window to form a complementary resource placement relationship. The overlapping status of complementary resource placement relationships within the graphics processor execution window is verified. Complementary resource placement relationships with overlapping occupancy are removed, and the remaining complementary resource placement relationships are used as the window-borne verification relationships. The resource complementarity placement relationships that do not overlap in the window carrying the verification relationship are merged into the same graphics processor execution window, and the operator resource occupation relationships carried by the resource complementarity placement relationship are merged into operator complementarity mixed relationship.

7. The machine learning resource regulation method based on fuzzy clustering as described in claim 6, characterized in that: The re-definition of the complementary mixed-part relations of subsequent operators specifically refers to... Based on the operator complementary mixed-part relationship, the occupied positions of the resource complementary placement relationship are read along the execution window of the graphics processor, and the occupied positions where the occupancy overlaps are marked as mixed-part failure positions; Remove the complementary resource placement relationship that has overlapping occupancy from the window placement result of the graphics processor execution window where the mixed part fails, and mark the operator resource occupancy relationship that has overlapping occupancy at the mixed part failure location as the failure attribution relationship; When the failure attribution relationship shows that the operator resource occupancy relationship continuously fills the graphics processor execution window, the operator resource occupancy relationship is reversed to exclusive protection occupancy. When the failure attribution relationship shows that the operator resource occupancy relationship only has local occupancy overlap at the mixed failure position, the operator resource occupancy relationship is reversed to off-peak execution occupancy. When the failure attribution relationship shows that the operator resource occupancy relationship continuously destroys the complementary resource placement relationship, the operator resource occupancy relationship is reversed to prohibit mixed occupancy. The exclusive protection occupancy is locked in a separate graphics processor execution window. The off-peak execution occupancy is moved out of the mixed-part failure position and re-entered into the window position. The prohibited mixed-part occupancy is excluded from the complementary mixed-part relationships of subsequent operators and the complementary mixed-part relationships of subsequent operators are redefined.

8. The machine learning resource regulation method based on fuzzy clustering as described in claim 1, characterized in that: The graphics processor resource pool is a collection of resources composed of multiple graphics processors, video memory resources, computing resources, and task scheduling relationships, which are shared by machine learning tasks for running. The operator startup order is the execution sequence formed by arranging the operator running events in the same machine learning task from earliest to latest according to the operator startup time. The graphics processor execution window is a parallel runtime segment defined on the same graphics processor for operators with complementary resource requirements. The computationally intensive occupancy, memory bandwidth occupancy, memory transport occupancy, and gap filling occupancy are the ownership states of operator resource occupancy relationships in the mixed-part affiliation state, which are characterized by computational execution dominance, memory bandwidth dominance, memory transport dominance, and the availability of waiting gaps for occupancy.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the machine learning resource regulation method based on fuzzy clustering as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the machine learning resource regulation method based on fuzzy clustering as described in any one of claims 1 to 8.

Images