A method and system for dynamic scheduling of heterogeneous computing power based on multi-level cache awareness

By dynamically identifying memory access variation ranges and cache behavior in a multi-level cached computing system and adjusting the task allocation order, the problem of cache access relationship fluctuations is solved, and the stability and continuity of task execution are achieved.

CN123114774APending Publication Date: 2026-11-17DXT (BEIJING) TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611577240.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-13
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing computing systems with multi-level caching structures, task memory access behavior is prone to stage differences as data access paths change, leading to frequent fluctuations in cache access relationships and affecting execution coordination and task continuity.

Method used

By obtaining the task number, memory access address, and access time of the CPU core task queue, an access time series is established, access change intervals are identified, memory access segments are divided, and cache hit and replacement records are combined to identify cache behavior identifiers, filter schedulable time intervals without cache conflicts, and dynamically adjust the task allocation order.

Benefits of technology

It enhances the ability to identify data access patterns during task execution, reduces the impact of data contention during parallel access of heterogeneous cores, maintains the continuity of cache usage and the stability of memory access paths, and reduces the risk of data migration disturbances and execution fluctuations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114774A_ABST
    Figure CN123114774A_ABST
Patent Text Reader

Abstract

This invention relates to the field of computer system resource scheduling technology, specifically to a method and system for dynamic scheduling of heterogeneous computing power based on multi-level cache awareness. The method extracts task numbers, memory access addresses, and time records; identifies changes in access locations to determine memory access segment positions; analyzes cache hit and replacement behavior and distribution span to determine stage identifiers; associates time intervals and compares heterogeneous core cache line occupancy to identify conflict distribution; filters conflict-free intervals and allocates computing cores based on behavioral characteristics; and sequentially organizes the execution queue and binds it to form a heterogeneous scheduling execution record. This invention achieves dynamic hierarchical expression of behavior by dividing memory access intervals, identifies memory access stages based on cache hit and replacement characteristics and enhances pattern recognition capabilities, combines cache line matching to filter schedulable intervals to reduce heterogeneous core competition, and completes core aggregation and sequential organization based on stage differences, thereby maintaining cache occupancy continuity and reducing the migration and fluctuation risks of cross-core scheduling.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of computer system resource scheduling technology, and in particular to a method and system for dynamic scheduling of heterogeneous computing power based on multi-level cache awareness. Background Technology

[0002] The field of computer system resource scheduling technology includes operating system kernel scheduling mechanisms, processor task allocation strategies, storage hierarchy management, and heterogeneous computing resource collaborative control. Its core lies in the unified modeling and allocation of various computing resources. By managing the task mapping relationships of different computing units such as CPUs, GPUs, and dedicated accelerators, and combining cache hierarchy structures, memory access paths, and data locality characteristics, it organizes and constrains the relationship between task execution order and resource occupancy. At the same time, this technology also involves thread scheduling strategies, load balancing mechanisms, cache consistency management, and multi-level storage access control to support resource coordination in complex computing environments.

[0003] The multi-level cache-aware heterogeneous computing power dynamic scheduling method refers to a method for task allocation and execution order control of CPU cores, GPU units, and other computing units in a computing system containing a multi-level cache structure. It mainly involves task partitioning based on cache level information, data access path analysis related to cache hit rate, setting mapping rules between tasks and computing units, and adjustment strategies for task queues during operation. Its implementation includes prioritizing tasks based on the data residency status of different cache levels, grouping and allocating tasks according to computing unit type, adjusting the task scheduling order based on cache access overhead, and controlling data transmission paths between different computing units to form a scheduling process oriented towards a multi-level cache structure.

[0004] Existing technologies focus on organizing heterogeneous computing power allocation based on static cache hierarchy relationships and task mapping rules. In long-term continuous operation scenarios, task memory access behavior is prone to stage differences as data access paths change. However, there is a lack of continuous segmentation capability for memory access change trajectories during operation, making it difficult to distinguish the data access characteristics corresponding to different stages. When multiple tasks alternately occupy cache resources, cache access relationships are prone to frequent fluctuations. For example, when related tasks alternate between CPU and GPU execution, data access paths may continuously change, which can easily lead to unstable task execution order and unbalanced cache occupancy, thereby affecting the execution coordination and task operation continuity in heterogeneous computing environments. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide a method and system for dynamic scheduling of heterogeneous computing power based on multi-level cache awareness. The technical solution is as follows:

[0006] A method for dynamic scheduling of heterogeneous computing power based on multi-level cache awareness includes the following steps:

[0007] S1: Obtain the task number, memory access address, and access time of the CPU core task queue, establish the correspondence between the three, form a memory access address sequence according to the access time, identify adjacent address changes and continuous change intervals, set segmentation boundary points, form a set of access change markers, and obtain a set of memory access segment locations.

[0008] S2: Extract the boundary point positions from the memory access segment location set, divide the memory access address sequence, read the first and second level cache hit records to summarize the access hit performance, obtain replacement records to form replacement performance, identify the address distribution span, establish corresponding relationships and mark stage identifiers to obtain a cache behavior identifier set.

[0009] S3: Call the cache behavior identifier set, associate memory access segments with access time, divide continuous time intervals according to start and end time, extract CPU core memory access addresses and convert them into a cache line number set, obtain the GPU core concurrently occupied number set, match common cache line numbers and positions, and obtain cache conflict distribution records.

[0010] S4: Based on the cache conflict distribution record, filter the schedulable time intervals with no common cache lines, correspond to the time of the cache behavior identifier set, extract the task number, stage order, access hit and replacement performance, and combine the computing power core number to determine the core allocation type to obtain the computing power allocation sequence table.

[0011] S5: Based on the computing power allocation sequence table, the task numbers are collected according to the computing power core number to form CPU core and GPU core execution queues. Combining the schedulable time interval and stage identifier with the stage order of the cache behavior identifier set, the queue execution order is sorted and bound to the record to obtain the heterogeneous computing power scheduling execution record.

[0012] A heterogeneous computing power dynamic scheduling system based on multi-level cache awareness, the system comprising:

[0013] The sequence partitioning module obtains the task number, memory access address, and access time of the CPU core task queue, establishes the correspondence between the three, forms a memory access address sequence according to the access time, identifies adjacent address changes and continuous change intervals, sets segmentation boundary points, forms a set of access change markers, and obtains a set of memory access segment locations.

[0014] The status recognition module extracts the boundary point positions from the memory access segment position set, divides the memory access address sequence, reads the first and second level cache hit records to summarize the access hit performance, obtains replacement records to form replacement performance, identifies the address distribution span, establishes corresponding relationships and marks stage identifiers to obtain a cache behavior identifier set.

[0015] The data contention detection module calls the cache behavior identifier set, associates memory access segments with access time, divides continuous time intervals according to start and end time, extracts CPU core memory access addresses and converts them into a cache line number set, obtains the GPU core concurrent occupancy number set, matches common cache line numbers and positions, and obtains cache conflict distribution records.

[0016] The intelligent scheduling module, based on the cache conflict distribution record, filters the schedulable time intervals with no common cache lines, corresponds to the time of the cache behavior identifier set, extracts the task number, stage order, access hit and replacement performance, and determines the core allocation type by combining the computing power core number to obtain the computing power allocation sequence table.

[0017] The queue management module, based on the computing power allocation sequence table, gathers task numbers according to the computing power core number to form CPU core and GPU core execution queues. Combining the schedulable time interval, stage identifier, and the stage order of the cache behavior identifier set, it organizes the queue execution order and binds records to obtain heterogeneous computing power scheduling execution records.

[0018] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0019] In this invention, by dividing the position change trajectory in the memory access address sequence into continuous intervals and forming a stage association relationship with the access time, a dynamic hierarchical expression of task memory access behavior can be realized. Based on the cache hit status, cache replacement status and address span characteristics corresponding to different memory access stages, stage identification is completed, which can enhance the ability to identify data access patterns during task execution. By combining the cache line number matching results within a continuous time interval to filter schedulable intervals, the impact of data competition generated during parallel access of heterogeneous cores can be reduced. Furthermore, by completing the aggregation of computing power cores and the reorganization of execution order based on the differences in stage behavior, the continuity of cache occupancy and the stability of memory access paths during task execution can be maintained, and the risks of data migration disturbances and execution fluctuations generated during cross-core scheduling can be reduced. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention;

[0021] Figure 2 This is a flowchart illustrating the acquisition process of S1 in this invention;

[0022] Figure 3 This is a flowchart illustrating the acquisition process of S2 in this invention;

[0023] Figure 4 This is a flowchart illustrating the acquisition process of S3 in this invention;

[0024] Figure 5 This is a flowchart illustrating the acquisition process of S4 in this invention;

[0025] Figure 6 This is a flowchart of the acquisition process for S5 of the present invention. Detailed Implementation

[0026] The following detailed explanation illustrates the specific implementation methods:

[0027] The basic implementation examples are as follows: Figure 1 As shown: A method for dynamic scheduling of heterogeneous computing power based on multi-level cache awareness includes the following steps:

[0028] S1: Obtain the task number of each task in the CPU core execution task queue, as well as the memory access address and access time record corresponding to the memory access instructions generated by each task. Establish the correspondence between the task number, memory access address and access time record, arrange them in the order of access time to form a memory access address sequence, and sequentially determine whether the access position of adjacent memory access addresses has changed. Record the sequence position where the access position has changed as a candidate change position. Perform continuous identification on the candidate change positions, and determine the interval where the access position changes continuously occur as the access change interval. Set segment boundary points at the boundary of the access change interval, organize each segment boundary point in the order of access time to form an access change mark set, and determine the corresponding position of each segment boundary point in the memory access address sequence based on the access change mark set to obtain the memory access segment position set.

[0029] Please see Figure 2 Specifically, S1 includes:

[0030] S11: Obtain the task number of each task in the CPU core execution task queue, as well as the memory access address and access time record corresponding to the memory access instructions generated by each task. Register the task number, memory access address, and access time record according to the memory access instruction attribution, and organize the registered memory access addresses according to the access time order to generate a memory access address sequence; as follows:

[0031] The system retrieves the task IDs of each task in the CPU core's task queue, along with the memory access addresses and timestamps corresponding to the memory access instructions generated by each task. It then calls the CPU's internal monitoring registers to extract the set of task identifiers for the current execution cycle, targeting the task IDs. to To determine if a value is non-zero, existing hardware instruction tracing techniques (such as IntelPT enabled by configuring the IA32_RTIT_CTLMSR register, or ARM CoreSight architecture with ATB bus configuration) are used to collect the raw memory access data stream triggered by the memory management unit's PMI (Performance Monitoring Interrupt) overflow when the buffer is full. The 64-bit virtual address signal is extracted and converted to a hexadecimal representation of the physical memory access address. Simultaneously, a high-precision event timer (HPET) or a clock timestamp counter (such as the RDTSC instruction) is used to obtain the clock pulse value at the time each memory access instruction is issued as the access time record. For example, a task number is set. The corresponding memory access address is And the access time is the first The timestamp will be used to assign the task number. The generated memory access addresses and access times are encapsulated into structured data. The task number, memory access address, and access time records are sequentially registered according to the memory access instruction's attribution. An index mapping table is then established, hashing the physical register address with the logical task number. By retrieving the bus snooping records from the cache controller, the traces of each memory access instruction at each stage of the pipeline are obtained. The address bus level state of each memory access instruction during execution is extracted and converted into the corresponding address value. All the obtained memory access address data is stored in a temporary buffer. A sorting operator is called to sort the data in ascending order according to the access time value. If an access time exists... and Then corresponding address Ranked corresponding address Previously, when instructions with the same access time were encountered, they were sorted a second time according to their order in the instruction pipeline. All sorted memory access addresses were then linearly concatenated according to their time order to form a continuous one-dimensional data vector, generating a sequence of memory access addresses.

[0032] S12: Based on the memory access address sequence, sequentially read the preceding and following memory access addresses corresponding to adjacent access sequences. Compare the access positions of the preceding and following memory access addresses to determine whether there is a change in access position between adjacent access sequences. Mark the sequence positions where access positions have changed to obtain candidate change positions; specifically as follows:

[0033] Based on the memory access address sequence, the preceding and following memory access addresses corresponding to adjacent access sequences are read sequentially, and the sequence index pointer is set to start from the beginning. Start moving, read position memory access address and location memory access address The previous memory access address is compared with the next memory access address, and a subtraction operation is performed to calculate the absolute value of the numerical offset between the two adjacent memory access addresses. (Unit: byte) is used to measure the span of adjacent memory access instructions in physical memory space. The unit of measurement is bytes. For example, retrieving the previous address and the next address Calculate the difference as Right now A length of 1 byte, calling a pre-set address space mutation judgment benchmark value. This benchmark value is set with reference to the standard cache line size of the processor's L3 cache in existing technology, and is set to a value of [value missing]. Bytes, for the difference Compared with the benchmark value Perform size comparison and set the system's maximum addressing span constant. (Based on x86_64 architecture) ),like In the open interval If a change in access position occurs between adjacent access sequences, it indicates that the memory access flow has jumped across regions at this point, violating spatial locality; conversely, if... In a closed interval If the access is determined to be a sequential access within spatial locality, it means that no out-of-bounds jump occurred during the memory access. No splitting is needed; the original continuous state is directly preserved. The sequence position where the access location changes is marked, specifically at the corresponding sequence index. Add a Boolean flag and assign it a value. If no change has occurred, then assign a value. For example, when the first The and the first The difference between the addresses is When the index position is 1 byte, Mark the jump point, traverse the entire memory access address sequence until the last element, and set all the marked bits to 0. The index number is extracted into a new dynamic array to obtain the candidate change position.

[0034] S13: Call the candidate change locations, identify the sequence intervals where access locations change consecutively according to the access time order, determine the sequence intervals where access locations change consecutively as access change intervals, and set segmentation boundary points at the boundaries of the access change intervals. Associate these segmentation boundary points with the corresponding positions in the memory access address sequence to obtain the memory access segment location set; specifically as follows:

[0035] The candidate change positions are called, and the sequence intervals in which the access positions change consecutively are identified according to the access time order. The jump point index numbers stored in the dynamic array output by S12 are extracted, and the index interval distance between adjacent candidate change positions is calculated. ,in Indicates the first The position indices of each transition point in the original sequence are used, with the unit being the number of transition points, and a continuity threshold is set. This threshold is determined by retrieving the average number of basic block instructions for the current task from the CPU front-end decoder cache. Perform dynamic calculation settings (set the calculation formula as: when hour ,otherwise In this example, it is set to a fixed value. (For ease of explanation), for each index interval distance With continuity discrimination threshold Perform numerical comparison, if In a closed interval If the changes in positions before and after are determined to belong to the same continuous change process, it indicates that there is a dense random hash access in the region. Candidate changes that satisfy the continuity condition are then set up and merged until a match is found. Greater than In this case, the merging of the current set is stopped and records in the new set are started. This indicates that there is a long sequential access segment between the jump points, which does not belong to the same hash access behavior. The sequence interval where the access position changes continuously is determined as the access change interval, for example, when finding an index. The addresses at all locations experienced jumps exceeding the baseline value, with adjacent intervals of all [missing information]. And it is within the range Within, the interval Define the access change interval as a range, and set segmentation boundary points at the boundaries of the access change interval, i.e., the starting index of the interval. Previous and end indexes Then, insert the segment identifiers and call the segment position mapping function. ,in The index number representing the input segmentation identifier. The corresponding physical memory pointer address of the output is represented by the sequence offset corresponding to the segment identifier, which is converted into a physical memory pointer address. The segment boundary point is associated with the corresponding position in the memory access address sequence. The start address pointer and end address pointer of each segment are recorded to form a list consisting of multiple address offset pairs, thus obtaining the set of memory access segment positions.

[0036] S2: Extract the sequence positions corresponding to the segment boundary points in the memory access segment location set, divide the memory access address sequence into multiple memory access segments, read the corresponding L1 cache hit records and L2 cache hit records in each memory access segment, summarize them to form access hit performance, obtain the corresponding cache replacement records in the memory access segment to form replacement performance, identify the distribution span of memory access addresses in the memory access segment, establish a correspondence between the access hit performance, replacement performance and distribution span corresponding to the same memory access segment, arrange each correspondence according to the order of each memory access segment in the memory access address sequence, mark each memory access segment with a stage identifier according to the arranged correspondence, summarize each stage identifier in order to obtain a cache behavior identifier set;

[0037] Please see Figure 3 Specifically, S2 includes:

[0038] S21: Extract the sequence positions corresponding to the segment boundary points in the memory access segment location set, call the access order adjacent to the sequence positions in the memory access address sequence, take the segment boundary points as splitting nodes, determine the memory address range between adjacent splitting nodes, divide the memory access address sequence according to the access time order, register each memory access segment and its corresponding sequence position, and obtain the segment boundary position table; as detailed below:

[0039] Extract the sequence positions corresponding to the segment boundary points in the memory access segment location set, call the access order adjacent to the sequence positions in the memory access address sequence, and retrieve the boundary point index values ​​stored in the dynamic array in S13. and The segmentation boundary points are used as splitting nodes to logically split the memory access address sequence, and a loop step variable is set. from The total number of points increasing from the dividing point In each loop, read the first... Each dividing point location With the Each dividing point location Determine the range of memory access addresses between adjacent split nodes, and for those within closed intervals... For each memory access instruction within the memory, the address is extracted;

[0040] If the current dividing point Then the index in the original sequence will be changed from arrive The four memory access addresses are grouped into the same independent subset, at which point the first memory access address is extracted. Bit address , No. Bit address , No. Bit address Passing the exam Bit address The memory access address sequence is divided according to the access time. The standard memory copy instructions in the operating system kernel mode are used to transfer the above address subsets belonging to different intervals to the contiguous storage space allocated on the stack. Each address subset is assigned a unique segment number. For example, index to The address set is marked as , index to The address set is marked as Each memory access segment is registered with its corresponding sequence location. A mapping table is created in the kernel-mode memory block, and the segment numbers are recorded. As the key, the starting offset of this segment in the original sequence and termination offset The value is stored as a reference, and the base address of each segment in physical memory is calculated using the address pointer. The total number of memory access instructions contained in each segment is also calculated. By performing cumulative statistics, a segment boundary position table is obtained.

[0041] S22: Based on the segment boundary bit order table, read the corresponding L1 cache hit record, L2 cache hit record, and cache replacement record within each memory access segment. Summarize the L1 and L2 cache hit records into an access hit representation, and organize the cache replacement records into a replacement representation. Compare the attribution relationship between the hit representation and replacement representation within the same memory access segment to obtain the cache representation identifier table; details are as follows:

[0042] Based on the segment boundary bit table, the corresponding L1 cache hit record, L2 cache hit record, and cache replacement record are read within each memory access segment. Using a CPU performance monitoring unit (PMU) from existing technology, specific hardware counters (such as MEM_LOAD_RETIRED.L1_HIT and L2_HIT events) are read for... Within each segment, the sequence of memory access instructions is processed, and the query feedback status of each address in L1DataCache and L2Cache is extracted one by one. A level 1 cache hit counter is then set. and L2 cache hit counter If the address When querying L1, the response is Hit. Perform an increment operation; if it is Miss in L1 and Hit in L2, then... Perform an increment operation to obtain the statistics. The total number of hits within the cache is calculated by summing the L1 cache hit records and L2 cache hit records into an access hit performance, and then performing a summation operation to obtain the total number of hits for the current segment. For example, targeting middle One instruction, if and ,but Values The unit of measurement is the number of times. The PMU is synchronously called to extract the L2_LINES_OUT.SILENT cache line swapping record triggered during the execution of this segment, and the cache replacement counter is set. Read the total number of evicted cache lines generated during the LRU replacement algorithm execution, if any occurred during this period. The next cache line update will cause The cache replacement records are organized into replacement behavior, and the instruction execution cycle count in the performance monitoring unit is extracted and compared with... The system performs correlation analysis, comparing the hit and replacement behavior within the same memory access segment, and then performs numerical comparison operations to determine the relationship. and The size relationship determines the hit dominance coefficient. This measures the effective retention rate of the cache system when processing this segment. It is dimensionless and is calculated by substituting the data. Call the hit detection benchmark value The baseline value is set to a ratio of hit to replacement under typical system load. , calculate and Perform interval matching, if In the range ,because This indicates that the cache replacement frequency is higher than the hit frequency, resulting in low data reuse rate. Therefore, the current segment is considered replacement-sensitive, indicating severe cache thrashing. In the range If the result is determined to be of the sensitive type, it indicates a high data reuse rate. The determination result is then hashed and bound to the segment number to obtain the cache performance identifier table.

[0043] S23: Call the cache performance identifier table to identify the distribution span of memory addresses within the memory access segment. Establish a correspondence between access hit performance, replacement performance, and distribution span for the same memory access segment. Arrange the correspondence according to the order of each memory access segment in the memory address sequence. Determine the stage to which each memory access segment belongs and mark the stage identifier. Summarize the stage identifiers in order to obtain the cache behavior identifier set; as detailed below:

[0044] The cache performance identifier table is invoked to identify the distribution span of memory access addresses within the memory access segment, and the segment boundary bit sequence table generated in S21 is extracted. Corresponding address set Perform the maximum value selection operation to obtain Perform the minimum value selection operation to obtain The span of address space distribution is calculated by subtraction. The obtained value is Right now The physical space span of a byte is used to establish a correspondence between access hit behavior, replacement behavior, and distribution span according to the same memory access segment, and the data obtained in S22 is used to... , And the current calculation Encapsulate the relationships into a structure, arrange the relationships according to the order of each memory access segment in the S11 memory access address sequence, and read... , to The structure sequence is used to determine the stage to which each memory access segment belongs and to mark the stage identifier, and a distribution dispersion judgment threshold is set. This threshold is set with reference to the standard operating system memory page size and can be dynamically configured to be an integer multiple of the system page size (to accommodate large page memory scenarios such as 2MB). In this example, the default setting is [missing value]. byte, for of and Perform a comparison, if In the range Furthermore, the hit dominance coefficient of S22 output In the range If the task involves extensive data fetching and a high cache miss rate, then this stage is designated as the "Data Streaming Access Stage" and marked in hexadecimal. ,like In the range and In the range If the task is concentrated within a single page and the data reuse is extremely high, it is labeled as a "local computation phase" and marked as such. If other combinations occur, they are marked as "mixed transition phase" and assigned a value. ,For example because and Marked as The identifiers of each stage are summarized in sequence, and the identifier codes corresponding to each segment are pushed into the result queue in a linear order on the time axis to form a time series vector containing the micro-features of each task execution, thus obtaining a set of cache behavior identifiers.

[0045] S3: Call the stage identifiers in the cache behavior identifier set, associate the memory access segments corresponding to each stage identifier with the access time records in the memory access address sequence, divide the memory access execution process of each task in the CPU core execution task queue into continuous time intervals according to the start and end access times corresponding to each memory access segment, extract the memory access addresses generated by the CPU core in each continuous time interval, convert them into a cache line number set according to the cache line ownership relationship corresponding to the memory access address, obtain the cache line occupation number set of the GPU core in the same continuous time interval, match and compare the cache line number set and the cache line occupation number set according to the cache line number, identify the common cache line numbers and their corresponding positions, mark the common cache line number situation in each continuous time interval, and obtain the cache conflict distribution record;

[0046] Please see Figure 4 Specifically, S3 includes:

[0047] S31: Invoke the stage identifiers in the cache behavior identifier set, associate the memory access segment corresponding to each stage identifier with the access time record in the memory access address sequence, read the start and end access times of the memory access segment, determine the time connection status of adjacent memory access segments, divide the memory access execution process of each task in the CPU core execution task queue into continuous time intervals, and obtain the memory access time interval set; as follows:

[0048] The system retrieves the stage identifiers from the cache behavior identifier set, associates the memory access segment corresponding to each stage identifier with the access time record in the memory access address sequence, and extracts the markers. The "data streaming access phase" identifier, which is associated with the segmentation. Retrieve the original timestamp data corresponding to the start memory access instruction of this segment, which is recorded in the hardware TSC register in S11. and the original timestamp data corresponding to the end memory access instruction. Read the previous adjacent segment The last access time record is Perform a subtraction operation to calculate the idle time difference between adjacent memory access segments:

[0049] ;

[0050] Substituting specific numerical values ​​into the calculation yields the following results. Call the system-defined time continuity judgment benchmark value This benchmark value is based on the average cycle time loss of a processor performing a complete context switch, and is adaptively adjusted according to the L1 cache hit rate and pipeline depth of different processor architectures. Perform numerical comparison operations to determine the value. and The interval to which it belongs, when the calculation result In a closed interval If the time interval between adjacent memory access segments is determined to be continuous, it means that the execution of the two tasks was not suspended or interrupted by the system. and The timeline is linearly spliced ​​together, and the maximum idle waiting time limit supported by the system is set. If read subsequently Start time and The end time difference is Because it is in the interval If the connection status is determined to be interrupted, it means that the task has been switched out of the core by the system scheduler. For interrupted states, the current timeline is truncated at the interrupt point and a new interval record is started. The memory access execution process of each task in the CPU core execution task queue is divided into continuous time intervals. At this time, the task is... The execution process is divided into the first continuous time interval. Its time span is from Continue until The above association, subtraction, comparison and splicing operations are repeated for the entire memory access address sequence. All generated discrete continuous time intervals are stored in the time interval index table in chronological order to obtain the memory access time interval set.

[0051] S32: Based on the memory access time interval set, extract the memory access addresses generated by the CPU core within each consecutive time interval, call the attribution relationship between memory access addresses and cache line numbers, map the memory access addresses to cache line numbers, determine the cache line number aggregation status within the same consecutive time interval, register the cache line numbers according to the consecutive time interval, and generate a cache line number set; specifically as follows:

[0052] Based on the set of memory access time intervals, the memory access addresses generated by the CPU core are extracted within each consecutive time interval, and the time interval is called. Retrieve all memory access addresses falling within the specified interval from the segment boundary bit sequence table, and extract the hexadecimal address sequence:

[0053] This function establishes the relationship between memory access addresses and cache line numbers, and sets the byte width offset of hardware cache lines. This value is based on the standard cache line size in existing technologies. byte( The following settings are used to perform a right shift operation on each retrieved memory access address to obtain the unique cache line number:

[0054] ;

[0055] in, Indicates the generated cache line number. Represents the input hexadecimal physical memory address. This represents a bitwise logical right shift operation. Representing a displacement constant, the physical meaning of this formula lies in eliminating the low-order values ​​in the physical address used to indicate the offset of bytes within a line. The high-order cache line label is retained, and the result is calculated by substituting the data:

[0056] ;

[0057] ;

[0058] ;

[0059] ;

[0060] Map the memory access address to the cache line number, determine the cache line number aggregation status within the same continuous time interval, create a temporary hash storage container and attempt to push the generated numbers onto it. During the pushing process, perform an equality check; if the container already contains the same value, perform a deduplication and discard operation, indicating that repeated accesses to data in the same region belong to the same cache line hit. Count the total number of non-duplicate cache lines in the current interval and record their physical index sequence. The extracted set of numbers is The cache line numbers are registered according to consecutive time intervals, and a dedicated bitmap storage block is allocated in the memory management unit to identify the intervals. Map and bind it to the corresponding number sequence, and synchronously transfer the distribution span in S23. The bytes are appended as attribute information to the record entry to generate a set of cache line numbers.

[0061] S33: Based on the cache line number set, obtain the cache line occupancy number set of the GPU core within the same continuous time interval. Compare the cache line number set with the cache line occupancy number set according to the cache line number to determine the co-occurrence status, record the common cache line number and its corresponding position, and mark them in the order of the continuous time interval to obtain the cache conflict distribution record; as detailed below:

[0062] Based on the cache line number set, the set of cache line occupancy numbers of the GPU core within the same continuous time interval is obtained. Cross-architecture clock sources are synchronized via APIC (Advanced Programmable Interrupt Controller) to calibrate the memory access timestamps of the CPU and GPU. When the GPU supports the Resizable BAR base address register or NVIDIA GPUDirectP2P technology, existing technologies such as PCIe bus listening or IOMMU (Input / Output Memory Management Unit) address translation table fetching methods are used to read the cache line occupancy numbers within the time interval. The shared memory sampling addresses involved in the graphics rendering task being executed by the internal GPU core are extracted, and the memory access address sequence generated by the GPU is extracted and a right shift is performed in the same way as in S32. Bitwise operations convert it into a cache line number;

[0063] Assume the obtained GPU cache line number sequence is as follows:

[0064] The CPU cache line number set is compared with the extracted GPU cache line occupancy number set according to the cache line number. A step-by-step logical equality matching operation is performed to determine whether an element in the CPU set exists in the GPU set, identifying the cache line numbers that appear in both sets. and Determine the co-occurrence status and perform a counting operation to obtain the number of cache lines currently in conflict. for Retrieve the total number of unique cache lines within the current CPU range, as calculated in S32. for ;

[0065] Calculate the local conflict rate Call the conflict determination benchmark ratio coefficient This coefficient is set based on an empirical value for the contention tolerance of heterogeneous processors sharing the final-level cache (this empirical value is derived from the contention threshold when bus bandwidth throughput decreases by 5% under multiple typical heterogeneous test benchmarks). (Right now Perform a numerical range matching operation on the calculation result. Perform interval matching judgment; if the result falls within the interval... This time interval is then marked as a high-risk conflict state, indicating that there is a serious data eviction phenomenon between the CPU and GPU during this period. If the cache is determined to be in a conflict-free and safe state, the common cache line number and its corresponding location are recorded, and the conflict numbers in the high-risk conflict state are recorded. Index bit in CPU sequence and the index bit in the GPU sequence Perform combination encapsulation, mark according to continuous time intervals, and record in the scheduling status record table. The row containing the Boolean value As a conflict marker, it is filled in if the state is safe. The process involves iterating through subsequent intervals within the memory access time interval set and repeating the above comparison process to obtain a cache conflict distribution record.

[0066] S4: Based on the common cache line number marking results in the cache conflict distribution record, filter out continuous time intervals without common cache line number markings to generate a set of schedulable time intervals. Match the set of schedulable time intervals with the stage identifiers in the cache behavior identifier set. Extract the task number, stage order, access hit performance, and replacement performance associated with each schedulable time interval. Obtain the computing power core number corresponding to the CPU core and GPU core. Determine the core allocation type corresponding to the task number based on the access hit performance and replacement performance. Assign stage tasks with access hit performance as the main factor to the CPU core and stage tasks with replacement performance as the main factor to the GPU core. Organize the correspondence between task number and computing power core number according to the schedulable time interval order to obtain the computing power allocation sequence table.

[0067] Please see Figure 5 Specifically, S4 includes:

[0068] S41: Based on the common cache line number marking result in the cache conflict distribution record, read the continuous time interval marking field, filter the continuous time intervals that do not have a common cache line number marking, sort the filtered time intervals by start and end access time, determine the overlap status of adjacent time intervals, register the non-overlapping time intervals with the task number, and generate a set of schedulable time intervals; as detailed below:

[0069] Based on the common cache line number marking result in the cache conflict distribution record, the scheduling status record table generated in S33 is called to extract the Boolean conflict flag bit corresponding to each line record. A logical inversion operation is performed on the flag bit to remove high-risk conflict intervals with a flag value of 1. Continuous time intervals without common cache line number markings are then selected, i.e., conflict-free safe intervals with a flag value of 0 are selected as the allocation objects. Simultaneously, the task numbers and corresponding memory access data within the high-risk conflict intervals with a flag value of 1 are pushed into the delayed retry queue, and a system backoff cycle duration is appended to their timestamps. These retry data will wait for the next global scheduling cycle to re-enter the original data stream of S11 for secondary scheduling evaluation, thus avoiding task starvation and loss, such as when a time interval is read. The flag bit is 0, and its corresponding start timestamp is extracted. and end timestamp A doubly linked list structure is allocated in the dynamic memory area to store the filtered time intervals. The filtered time intervals are arranged according to the start and end access times. The sorting operator is called to perform ascending sorting according to the size of the start timestamp value. The sorted linked list is traversed, and adjacent time intervals are read in turn. and The time boundary value is used to perform a subtraction operation to calculate the overlap offset. If the previous interval End time is The next interval Start time is ;

[0070] Calculation Call the overlap determination benchmark value This benchmark value is set with reference to processor clock jitter deviation and bus latency. Set the maximum pre-overlap constant that the system can tolerate. Determine the overlap between adjacent time intervals, and perform interval comparison based on the calculation results. If... In a closed interval Then, the overlapping state of adjacent time intervals is determined to be non-overlapping, if the value is in an open interval. (e.g., the start time of the next interval is) ,but If the intervals overlap, it is determined that there is a time overlap, which means that the parallel execution of instructions will cause bus resource contention interference. For intervals with overlapping states, a boundary pruning operation is performed to extract the overlapping part and take... and The mean is used as the new dividing point to modify the original boundary value. The non-overlapping time intervals and task numbers are registered as fields, and the task numbers stored in S11 are retrieved. The process involves filtering and correcting the start and end values ​​of non-overlapping time intervals to ensure they do not interfere with each other, along with their corresponding task numbers, and then packaging and registering them into a data structure. This data structure is then written into the offset address register of the operating system's scheduling control block to generate a set of schedulable time intervals.

[0071] S42: Call the set of schedulable time intervals, match the start and end access times of the schedulable time intervals with the time ranges of the stage identifiers in the cache behavior identifier set, extract the task number, stage order, access hit performance, and replacement performance, determine the overlap between the stage identifier and the interval time, and register the corresponding content to the stage field to obtain the stage task mapping table; as detailed below:

[0072] Call the set of schedulable time intervals and extract non-overlapping time intervals from the linked list structure. Boundary parameters (initial) ,termination ( ), synchronously retrieve the cache behavior identifier set, match the start and end access times of the schedulable time interval with the time ranges of the stage identifiers in the cache behavior identifier set, read the time range records corresponding to each stage identifier, perform an inclusion matching judgment of the time range, and for the interval Iterate through all memory access segments recorded in S23, if the segment... The start time of access is recorded as And the last access time is recorded as Calculate the initial difference and the difference at the end Determine if the stage marker coincides with the interval time; when both the initial difference and the final difference are within the closed interval. When the stage identifier and the interval time coincide, it is considered a complete containment. If one is positive and the other is negative, it is considered a partial intersection and truncation. The task number, stage arrangement order, access hit performance, and replacement performance are extracted. For the case of complete containment, the task number associated with that segment is directly extracted. The stage arrangement order index number and the access hit performance calculated in S22 and replacement performance Assuming that the data is extracted from record S22 and The structure assignment instruction is invoked to push the above parameters into the attribute field of the current scheduling time slice. If it is determined to be a partial cross-truncation and multiple segmented fragments are detected within the time interval, an accumulation and aggregation operation is performed to merge the parameters, and the interval comprehensive performance of the overlapping area is obtained through a formula:

[0073] ;

[0074] In the formula, This represents the calculated interval weighted performance value (dimensions are equivalent frequencies). Indicates the first The raw performance statistics (hit count or replacement count) of each cut segment within this time range. Indicates the first The duration of each segment within the current interval (in nanoseconds). This indicates the total length of the currently schedulable time interval. To determine the total number of overlapping segments, the cross-boundary performance is smoothly transitioned according to the time proportion. During this process, boundary crossing detection is performed. If the sum of the durations of multiple segments exceeds the total length of the currently schedulable time interval (i.e., ...), the cross-boundary performance will be checked. When, for weighting factors A normalization operator is introduced to scale the values ​​proportionally to a total of 1, preventing abnormal overflow calculations of the performance values. The smoothed comprehensive hit value is obtained by inputting specific truncated data. The content corresponding to the overlapping state is registered in the stage field. A behavior descriptor covering time and performance attributes is assigned to each task number in the system global scheduling table, resulting in a stage task correspondence table.

[0075] S43: Based on the stage task mapping table, read the access hit performance and replacement performance corresponding to the task number, and obtain the computing power core number corresponding to the CPU core and GPU core. Compare the access hit performance record with the replacement performance record to determine the core allocation type corresponding to the task number. Establish a correspondence between the task number and the computing power core number according to the schedulable time interval order to obtain the computing power allocation sequence table; as detailed below:

[0076] Based on the phase task mapping table, read the access hit performance and replacement performance corresponding to the task number, and then read the task number from the system global scheduling table. Access hit performance items associated within the schedulable interval and replacement items The system compares the access hit performance record with the replacement performance record and performs a subtraction operation to obtain the core allocation discrimination difference. Call the allocation decision benchmark threshold The benchmark threshold is set based on the power consumption loss coefficient of task switching and bus bandwidth transfer under different system architectures. Regarding the discriminant difference Execute interval judgment logic and set the maximum gain upper limit constant. Determine the core allocation type corresponding to the task number. In the open interval If the cache hit rate is significantly higher than the replacement rate, the core allocation type corresponding to the task number is determined to be compute-intensive affinity, suitable for execution on a CPU with a complex, multi-level, high-capacity cache. In this case, the CPU core number is extracted from the operating system's kernel resource pool. Binding is achieved by modifying the process's kernel affinity mask through a system call function such as `sched_setaffinity()` in the Linux kernel, which is a low-level operating system function. In a closed interval This indicates that the cache is frequently replaced, classifying the allocation type as memory bandwidth sensitive (memory access dominant). This is suitable for GPU execution using a parallel architecture to hide memory access latency. In this case, the GPU core number should be extracted. Binding is performed for the current computational example. The result will be the task The computing core number is marked as For another segment in S22 Execute the same logic, and if it hits, the result is... And the replacement is manifested as Calculate the difference Because it is in the interval Mark its associated computing core number as GPU. The task number is associated with the computing core number according to the schedulable time interval. Then, according to the chronological order of the schedulable time interval, the kernel-mode memory copy instruction is called to copy the task. Core Physics The time interval and stage identifier are packaged and stored in a static allocation vector to obtain the computing power allocation sequence table.

[0077] S5: Based on the correspondence between task numbers and computing core numbers arranged in the schedulable time interval order in the computing power allocation sequence table, the task numbers are grouped into core queues according to the computing core numbers, forming CPU core execution queues and GPU core execution queues respectively. Based on the schedulable time interval and stage identifier associated with the task number, combined with the stage arrangement order in the cache behavior identifier set, the execution order of the task numbers in each core execution queue is sorted. The sorted task numbers, computing core numbers, execution queue types and queue execution order are bound and recorded to obtain heterogeneous computing power scheduling execution records.

[0078] Please see Figure 6 Specifically, in S5, this includes:

[0079] S51: Based on the correspondence between task numbers and computing core numbers arranged in the schedulable time interval order in the computing power allocation sequence table, read the task number, computing core number, and schedulable time interval order. Group task numbers with the same computing core number into the same core queue field. Determine the queue type corresponding to the core number, and separately group the CPU core execution queue and GPU core execution queue to obtain the core task aggregation table; details are as follows:

[0080] Based on the correspondence between task numbers and computing core numbers arranged in the schedulable time interval order in the computing power allocation sequence table, the starting address of the computing power allocation sequence table stored in the operating system's memory management unit is called to read the task numbers, computing core numbers, and the schedulable time interval order, and the task numbers are extracted from them. Computing core number and the starting value of the schedulable time interval With termination value Read the task number from another record. Computing core number and corresponding time intervals to Two independent, contiguous storage spaces are allocated in the system's dynamic random access memory as the core queue fields, named respectively. and The task numbers corresponding to the same computing power core number are grouped into the same core queue field. For each record in the computing power allocation sequence table, a core number extraction action is performed, and the extracted number value is compared with the preset core type constant. Perform a logical equality check; if the check is true, assign a number to the task. The associated timestamp is encapsulated as a structure and pushed into the stack. At the end of the queue pointer, if the core number value matches the GPU type constant... If they match, assign the task number. deposit In the linked list node of the queue, determine the queue type corresponding to the core number by retrieving the function bit field in the core descriptor set by the underlying hardware. When the function bit value is in a closed interval... The time is determined to be a scalar calculation queue mainly based on serial instructions, when the function bit value is in a closed interval. The time frame is determined to be a vector computation queue primarily based on single instruction multiple data streams, which is then aggregated into the CPU core execution queue and the GPU core execution queue, and counted using system-level counters. Total number of tasks and Total number of tasks If the current task If a CPU is assigned to a queue, its queue type identifier in the aggregation table is updated to [value]. (Scalar), if If it is assigned to the GPU queue, the identifier is updated. (Vector) Execute the memory refresh instruction to write the logical base addresses of two physically isolated queues into the system scheduling list, and obtain the core task aggregation table.

[0081] S52: Call the core task aggregation table, read the schedulable time interval and stage identifier associated with the task number, combine the stage arrangement order in the cache behavior identifier set, arrange the task numbers in the same core execution queue according to the schedulable time interval, determine the corresponding position of the stage identifier in the stage arrangement order, register the task number and queue execution order, and generate a queue order tag table; as detailed below:

[0082] The core task aggregation table is invoked to read the schedulable time range and stage identifier associated with the task number, and the task number is obtained. Associated schedulable time intervals and the stage identifiers defined in S23 (Data streaming access phase), combining the phase arrangement order in the cache behavior identifier set, retrieve the original memory access time record sequence generated in S11 and the global phase arrangement order vector extracted from the cache behavior identifier set, and arrange the task numbers in the same core execution queue according to the schedulable time interval. Sort all task numbers by execution time in chronological order and read... start time With another task in the same queue start time Perform subtraction to calculate the time difference. The order of tasks is determined by the sign of the difference, prioritizing tasks with smaller time values. The time priority is assigned a time level. Determine the task Assign a grade to the second-best item Determine the position of the stage identifier in the stage sequence and identify the stage identifier. Physical storage index location in the global behavior vector The numerical significance of this physical index position lies in the fact that the earlier the index position, the greater its data preparation role in the global computation flow, and the smaller the corresponding penalty value in sorting. This is reflected in the extracted index value. The queue sorting operator is invoked to calculate the overall execution sequence number of each task. Substitute directly into this formula The value is used as a penalty for execution order; the smaller the overall sequence number, the higher the priority it should be in the pipeline.

[0083] against Pick And by substituting the corresponding penalty terms, we can calculate the result. ,against Pick And the penalty item is Calculated Perform a numerical comparison operation to determine the value of the result. The value is in a closed interval The task is registered as the first execution priority, and The value is in a closed interval The task is registered as the second priority, and its task number and queue execution order are recorded. The task is then sequentially filled into the process control block (PCB) and queue control block of the operating system. The priority value and estimated execution time are used to generate a queue order label table with strict execution order constraints.

[0084] S53: Based on the queue order tag table, read and organize the task number, computing core number, execution queue type, and queue execution order. Bind the task number to the computing core number, determine the matching status between the execution queue type and the core number, and write the task number, computing core number, execution queue type, and queue execution order into the scheduling record field to obtain the heterogeneous computing power scheduling execution record; details are as follows:

[0085] Based on the queue order label table, read and organize the task number, computing core number, execution queue type, and queue execution order to extract the task number. Computing core number Execution queue type and queue execution order The system retrieves the idle status parameters of the current physical core from the Resource Mapping Table, binds the task number to the corresponding computing core number, and writes the task mount bit in the control register of the selected core using memory-mapped I / O technology. The logical starting address is used to determine the matching status between the execution queue type and the core number, and then the core physical number is extracted. The corresponding 8-bit factory hardware signature and queue type The corresponding 8-bit compilation protocol code Perform a bitwise XOR logical operation to calculate the matching deviation value. The significance of this XOR operation lies in detecting whether there is a mismatch in characteristic bits between the instruction set architecture required by the task and the inherent execution architecture of the hardware. The value is If the value is within a closed interval, the match is considered to be completely identical. If a conflict is identified, it indicates hardware incompatibility. In this case, an exception rollback strategy is executed, suspending the task and allowing the system bus to schedule it back to the general-purpose processing core. If a complete match is determined, the task number, computing core number, execution queue type, and queue execution order are written to the scheduling record field. The underlying system kernel's write function is called, specifically allocating contiguous memory in the scheduling log area of ​​non-volatile memory. The byte-level record space is sequentially written with a hexadecimal-formatted task identifier, core physical address index, queue property code, and a four-byte aligned execution sequence number. The bitstream of the current record is extracted, a Cyclic Redundancy Check (CRC) code is calculated, and appended to the end of the field to ensure data tamper-proofing. Tasks in The above extraction, binding, matching, and writing actions are performed synchronously to ensure the GPU core number is accurate. With queue type The XOR matching result also meets the zero deviation requirement. All scheduling instruction sets that have passed the core verification are uniformly encapsulated and sent to the underlying arbiter to obtain the heterogeneous computing power scheduling execution record.

[0086] A heterogeneous computing power dynamic scheduling system based on multi-level cache awareness, the system includes:

[0087] The sequence partitioning module obtains the task number, memory access address, and access time of the CPU core task queue, establishes the correspondence between the three, forms a memory access address sequence according to the access time, identifies adjacent address changes and continuous change intervals, sets segmentation boundary points, forms a set of access change markers, and obtains a set of memory access segment locations.

[0088] The status recognition module extracts the boundary point positions in the memory access segment location set, divides the memory access address sequence, reads the first and second level cache hit records to summarize the access hit performance, obtains replacement records to form replacement performance, identifies the address distribution span, establishes corresponding relationships and marks stage identifiers, and obtains a cache behavior identifier set.

[0089] The data contention detection module calls the cache behavior identifier set, associates memory access segments with access time, divides continuous time intervals according to start and end times, extracts CPU core memory access addresses and converts them into a cache line number set, obtains the GPU core concurrently occupied number set, matches common cache line numbers and positions, and obtains cache conflict distribution records.

[0090] The intelligent scheduling module, based on the cache conflict distribution record, filters the schedulable time intervals with no common cache lines, corresponds to the cache behavior identifier set time, extracts the task number, stage order, access hit and replacement performance, and determines the core allocation type by combining the computing power core number, thus obtaining the computing power allocation sequence table.

[0091] The queue management module, based on the computing power allocation sequence list, gathers task numbers according to the computing power core number to form CPU core and GPU core execution queues. Combining the schedulable time interval, stage identifier, and cache behavior identifier set, it organizes the queue execution order and binds records to obtain heterogeneous computing power scheduling execution records.

[0092] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for dynamic scheduling of heterogeneous computing power based on multi-level cache awareness, characterized in that, Includes the following steps: S1: Obtain the task number, memory access address, and access time of the CPU core task queue, establish the correspondence between the three, form a memory access address sequence according to the access time, identify adjacent address changes and continuous change intervals, set segmentation boundary points, form a set of access change markers, and obtain a set of memory access segment locations. S2: Extract the boundary point positions from the memory access segment location set, divide the memory access address sequence, read the first and second level cache hit records to summarize the access hit performance, obtain replacement records to form replacement performance, identify the address distribution span, establish corresponding relationships and mark stage identifiers to obtain a cache behavior identifier set. S3: Call the cache behavior identifier set, associate memory access segments with access time, divide continuous time intervals according to start and end time, extract CPU core memory access addresses and convert them into a cache line number set, obtain the GPU core concurrently occupied number set, match common cache line numbers and positions, and obtain cache conflict distribution records. S4: Based on the cache conflict distribution record, filter the schedulable time intervals with no common cache lines, correspond to the time of the cache behavior identifier set, extract the task number, stage order, access hit and replacement performance, and combine the computing power core number to determine the core allocation type to obtain the computing power allocation sequence table.

2. The heterogeneous computing power dynamic scheduling method based on multi-level cache awareness according to claim 1, characterized in that: The method further includes: S5: Based on the computing power allocation sequence table, the task numbers are collected according to the computing power core number to form CPU core and GPU core execution queues. Combining the schedulable time interval and stage identifier with the stage order of the cache behavior identifier set, the queue execution order is sorted and bound to the record to obtain the heterogeneous computing power scheduling execution record.

3. The heterogeneous computing power dynamic scheduling method based on multi-level cache awareness according to claim 1, characterized in that: The steps for obtaining S1 are as follows: S11: Obtain the task number of each task in the CPU core execution task queue, as well as the memory access address and access time record corresponding to the memory access instruction generated by each task. Register the task number, memory access address and access time record according to the memory access instruction ownership relationship, organize the memory access addresses according to the access time order, and generate a memory access address sequence. S12: Based on the memory access address sequence, read the previous memory access address and the next memory access address corresponding to the adjacent access order in sequence, compare the access position of the previous memory access address and the next memory access address, determine the access position change status between adjacent access orders, write the changed sequence position into the candidate tag field, and obtain the candidate changed position. S13: Call the candidate change position, identify the sequence intervals in which the access position changes continuously according to the access time order, determine the sequence interval as the access change interval, set the segment boundary point at the boundary of the access change interval, organize the segment boundary point, associate the segment boundary point with the corresponding position of the memory access address sequence, and obtain the memory access segment position set.

4. The heterogeneous computing power dynamic scheduling method based on multi-level cache awareness according to claim 1, characterized in that: The steps for obtaining S2 are as follows: S21: Extract the sequence position corresponding to each segment boundary point in the memory access segment location set, call the adjacent access order in the memory access address sequence, take the segment boundary point as the splitting node, determine the range of memory access addresses between adjacent splitting nodes, divide the memory access address sequence according to the access time order, and obtain the segment boundary position table. S22: Based on the segment boundary bit order table, read the first-level cache hit record, second-level cache hit record and cache replacement record in each memory access segment, summarize the first-level cache hit record and second-level cache hit record into access hit performance, organize the cache replacement record into replacement performance, and obtain the cache performance identifier table. S23: Call the cache performance identifier table, identify the distribution span of memory access addresses within the memory access segment, establish a correspondence between access hit performance, replacement performance and distribution span according to the same memory access segment, arrange the correspondence according to the order of memory access address sequence, mark the stage identifier of each memory access segment, and obtain a cache behavior identifier set.

5. The heterogeneous computing power dynamic scheduling method based on multi-level cache awareness according to claim 1, characterized in that: The steps for obtaining S3 are as follows: S31: Call the stage identifiers in the cache behavior identifier set, associate the memory access segment corresponding to the stage identifier with the access time record in the memory access address sequence, read the start and end access time of the memory access segment, determine the time connection status of adjacent memory access segments, divide the CPU core task memory access execution process into continuous time intervals, and obtain the memory access time interval set. S32: Based on the memory access time interval set, extract the memory access address generated by the CPU core in each continuous time interval, call the memory access address and cache line number attribution relationship, assign the memory access address to the cache line number, determine the number collection status in the same continuous time interval, register according to the continuous time interval, and generate a cache line number set. S33: Based on the cache line number set, obtain the cache line occupancy number set of the GPU core in the same continuous time interval, compare the cache line number set with the cache line occupancy number set according to the cache line number, determine the common occurrence status, record the common cache line number and corresponding position, mark according to the continuous time interval, and obtain the cache conflict distribution record.

6. The heterogeneous computing power dynamic scheduling method based on multi-level cache awareness according to claim 1, characterized in that: The steps for obtaining S4 are as follows: S41: Based on the common cache line number marking result in the cache conflict distribution record, read the continuous time interval marking field, filter the continuous time intervals that do not have the common cache line number marking, arrange them according to the start and end access time, determine the overlapping status of adjacent intervals, register the non-overlapping intervals with the task number, and generate a set of schedulable time intervals. S42: Call the set of schedulable time intervals, read the start and end access times of the schedulable time intervals and the time range of the stage identifiers in the set of cache behavior identifiers, match the interval times with the stage times, determine the time overlap status, extract the task number, stage arrangement order, access hit performance and replacement performance, and obtain the stage task correspondence table. S43: According to the stage task correspondence table, read the access hit performance and replacement performance corresponding to the task number, and obtain the computing power core number corresponding to the CPU core and GPU core. Compare the access hit performance record item with the replacement performance record item, determine the core allocation type corresponding to the task number, and register the task number and computing power core number in the order of the schedulable time interval to obtain the computing power allocation sequence table.

7. The heterogeneous computing power dynamic scheduling method based on multi-level cache awareness according to claim 2, characterized in that: The steps for obtaining S5 are as follows: S51: Based on the correspondence between task number and computing core number in the computing power allocation sequence table, read the schedulable time interval order, determine the ownership of the computing core number, and collect the task number into the CPU core execution queue and GPU core execution queue to obtain the core task collection table. S52: Call the core task collection table, read the schedulable time interval and stage identifier associated with the task number, combine the stage arrangement order in the cache behavior identifier set, determine the order relationship of task numbers in the same core execution queue, register the task number and queue execution order, and generate a queue order tag table. S53: Based on the queue order tag table, read the task number, computing core number, execution queue type and queue execution order, determine the matching status of the execution queue type and computing core number, bind the task number, computing core number, queue type and execution order to obtain the heterogeneous computing power scheduling execution record.

8. A heterogeneous computing power dynamic scheduling system based on multi-level cache awareness, characterized in that, The system is used in the heterogeneous computing power dynamic scheduling method based on multi-level cache awareness as described in any one of claims 1-7, the system comprising: The sequence partitioning module obtains the task number, memory access address, and access time of the CPU core task queue, establishes the correspondence between the three, forms a memory access address sequence according to the access time, identifies adjacent address changes and continuous change intervals, sets segmentation boundary points, forms a set of access change markers, and obtains a set of memory access segment locations. The status recognition module extracts the boundary point positions from the memory access segment position set, divides the memory access address sequence, reads the first and second level cache hit records to summarize the access hit performance, obtains replacement records to form replacement performance, identifies the address distribution span, establishes corresponding relationships and marks stage identifiers to obtain a cache behavior identifier set. The data contention detection module calls the cache behavior identifier set, associates memory access segments with access time, divides continuous time intervals according to start and end time, extracts CPU core memory access addresses and converts them into a cache line number set, obtains the GPU core concurrent occupancy number set, matches common cache line numbers and positions, and obtains cache conflict distribution records. The intelligent scheduling module, based on the cache conflict distribution record, filters the schedulable time intervals with no common cache lines, corresponds to the time of the cache behavior identifier set, extracts the task number, stage order, access hit and replacement performance, and determines the core allocation type by combining the computing power core number to obtain the computing power allocation sequence table. The queue management module, based on the computing power allocation sequence table, gathers task numbers according to the computing power core number to form CPU core and GPU core execution queues. Combining the schedulable time interval, stage identifier, and the stage order of the cache behavior identifier set, it organizes the queue execution order and binds records to obtain heterogeneous computing power scheduling execution records.