A heterogeneous computing server system based on a reconfigurable dataflow architecture

By using a heterogeneous computing server system based on a reconfigurable data flow architecture, and dynamically selecting migration or partial reconfiguration strategies, the problems of excessive latency and low resource utilization under the static mapping strategy are solved, thus achieving efficient heterogeneous computing resource management.

CN123114657APending Publication Date: 2026-11-17BEIJING SHENWEI DIGITAL TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611275771.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing heterogeneous computing systems deployed in a hybrid scenario of high-performance computing, supercomputing, large-scale cluster computing and edge computing, static mapping strategies cannot dynamically respond to resource contention and computing unit capability mismatch on critical paths in the data flow graph, resulting in excessive latency and low resource utilization.

Method used

A heterogeneous computing server system based on a reconfigurable data flow architecture is adopted. The task parsing module extracts the topological features of the data flow graph and generates an initial mapping scheme. When the critical path latency exceeds the standard, the reconfiguration group determination module identifies the continuous exceeding operators as reconfiguration groups. The cost calculation module calculates the combined cost of migration and local reconfiguration. The reconfiguration decision module dynamically selects the optimal reconfiguration strategy, including migration or local reconfiguration.

Benefits of technology

It enables rapid adaptive reconfiguration decisions in latency-sensitive scenarios, balancing response speed and global optimality, improving system resource utilization and computational efficiency, and adapting to the real-time scheduling requirements of high-performance computing and large-scale cluster computing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114657A_ABST
    Figure CN123114657A_ABST
Patent Text Reader

Abstract

This invention relates to the field of digital data processing technology, and particularly to a heterogeneous computing server system based on a reconfigurable dataflow architecture, comprising: a task parsing module for parsing computing tasks into a dataflow graph and extracting the topological features of the dataflow graph; an initial mapping module for generating and executing an initial mapping scheme; a latency monitoring module for determining the heterogeneous computing unit where each operator on the critical path is currently located based on the initial mapping scheme; a reconfiguration group determination module for determining multiple consecutive out-of-spec operators with data dependencies into a reconfiguration group; a cost calculation module for determining a first comprehensive cost and a second comprehensive cost; and a reconfiguration decision module for determining a reconfiguration strategy, which can dynamically select the optimal reconfiguration strategy when the latency of the critical path operator exceeds the standard, taking into account both response speed and global optimality.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of digital data processing technology, and in particular to a heterogeneous computing server system based on a reconfigurable data flow architecture. Background Technology

[0002] With the rapid development of applications such as the Internet of Things, artificial intelligence, and big data analytics, computing tasks are becoming increasingly diverse and dynamic. Different computing tasks have significantly different requirements for computing resources. For example, some tasks are suitable for execution on general-purpose processors, some are suitable for execution on dedicated accelerators such as GPUs and NPUs, and some tasks, due to their obvious data flow characteristics, are suitable for execution on reconfigurable dataflow computing units.

[0003] Existing heterogeneous computing systems typically employ static mapping strategies, allocating operators to different computing units according to preset rules before task execution and not adjusting them during execution. However, when operators on the critical path of the data flow graph experience excessive latency due to resource contention or mismatched computing unit capabilities, static mapping schemes cannot respond dynamically, leading to a decrease in overall system throughput. To address this issue, some solutions attempt to introduce runtime task migration mechanisms by pre-deploying multiple identical computing replicas and switching to a backup replica when a single replica fails. However, this only applies to redundancy switching between homogeneous units and cannot leverage the capability differences between heterogeneous computing units for performance optimization. Furthermore, in scenarios involving the hybrid deployment of high-performance computing, supercomputing, large-scale cluster computing, and edge computing, the dynamic fluctuations in data flow load are more pronounced, further amplifying the latency exceedances and low resource utilization issues of traditional static mapping mechanisms. Summary of the Invention

[0004] This application provides a heterogeneous computing server system based on a reconfigurable dataflow architecture, which can dynamically select the optimal reconfiguration strategy by comparing the combined costs of migration operations and local reconfiguration operations when the latency of critical path operators exceeds the limit, thus balancing response speed and global optimality.

[0005] To address the above problems, the embodiments of this application provide the following technical solutions: In a first aspect, this application provides a heterogeneous computing server system based on a reconfigurable dataflow architecture, comprising: The task parsing module receives computation tasks, parses them into a data flow graph, and extracts the topological features of the data flow graph. These features include the operation type of each operator in the data flow graph, the input and output data sizes of each operator, the data dependencies between operators, and the execution order of operators on the critical path. The critical path is the path in the data flow graph with the largest sum of delays from the starting operator to the ending operator. The initial mapping module generates and executes an initial mapping scheme based on the computational features of each operator in the data flow graph and the capability features of each heterogeneous computing unit in the heterogeneous computing resource pool. The delay monitoring module determines the heterogeneous computing unit where each operator on the critical path is currently located, based on the initial mapping scheme, when the operator's delay exceeds the limit on the critical path. The reconstruction group determination module identifies multiple consecutive operators with data dependencies that exceed the limit as a reconstruction group. The cost calculation module determines the first comprehensive cost of migrating the reconstruction group to other heterogeneous computing units for execution, and the second comprehensive cost of performing local reconstruction of the corresponding region in the reconstructable data flow computing unit. The reconstruction decision module determines the reconstruction strategy based on the first and second comprehensive costs.

[0006] This invention extracts the topological features of the data flow graph through a task parsing module, generates and executes an initial mapping scheme through an initial mapping module, and identifies continuously exceeding delay operators as a reconstruction group when the critical path latency exceeds the limit. The cost calculation module calculates the first comprehensive cost of the migration operation and the second comprehensive cost of the local reconstruction, respectively. The reconstruction decision module determines the optimal reconstruction strategy based on the comparison of the two. It can adaptively select migration or local reconstruction according to the actual latency at runtime, taking into account both response speed and global optimality. It can adapt to the real-time scheduling requirements of massive parallel data streams under high-performance computing, supercomputing, and large-scale cluster computing.

[0007] Furthermore, the refactoring decision module includes a migration execution submodule, which is used to migrate the refactoring group to other heterogeneous computing units for execution if the first synthesis cost is less than the second synthesis cost.

[0008] Based on the above technical means, when the first comprehensive cost is less than the second comprehensive cost, the migration strategy is selected first, and the reconstructed group is migrated to other heterogeneous computing units for execution. The migration operation does not require hardware reconfiguration of the target computing unit, the decision speed is fast, the execution overhead is small, and it is suitable for latency-sensitive scenarios.

[0009] Furthermore, the system also includes: a pre-generation module, used to perform predictive analysis on the critical paths in the data flow graph based on topological characteristics before executing the computation task, and pre-generate multiple candidate reconstruction schemes; the reconstruction decision module also includes: a query submodule, used to query whether there is a candidate reconstruction scheme matching the reconstruction group among the pre-generated candidate reconstruction schemes when the second comprehensive cost is less than or equal to the first comprehensive cost; a cache loading submodule, used to load the matching candidate reconstruction scheme to configure the target region in the reconstructable data flow computation unit when a matching candidate reconstruction scheme exists, and redeploy the operators in the reconstruction group to the target region and resume execution after the configuration is completed; and a real-time generation submodule, used to generate a reconstruction scheme in real time according to the data dependency relationship of each operator in the reconstruction group when no matching candidate reconstruction scheme exists, configure the target region and execute it.

[0010] Based on the above technical means, a pre-generation module pre-generates multiple candidate reconstruction schemes based on predictive analysis before task execution. When the local reconstruction cost is better, the candidate scheme in the cache is queried first. If a match is found, the configuration is loaded directly, avoiding the large computational overhead of generating reconstruction schemes in real time and significantly shortening the reconstruction response time. If no match is found, a scheme is generated in real time to ensure the completeness of the scheme.

[0011] Furthermore, the pre-generation module includes: an estimation submodule, used to determine the estimated execution latency of each operator on various heterogeneous computing units in the heterogeneous computing resource pool based on the operation type, input data scale, and output data scale of each operator on the critical path; a combination partitioning submodule, used to identify operators on the critical path whose estimated execution latency exceeds a preset threshold as candidate operators based on the estimated execution latency, and to partition the candidate operators into one or more candidate operator combinations according to the continuity of data dependencies; and a scheme generation submodule, used to generate corresponding candidate reconstruction schemes based on each candidate operator combination, each candidate reconstruction scheme containing region partitioning information and operator mapping information of the candidate operator combination on the reconfigurable data flow computing unit.

[0012] Based on the above technical means, by estimating the execution latency of each operator on various heterogeneous computing units, consecutive operators with excessive latency are combined into candidate operator combinations and corresponding candidate reconstruction schemes are generated. This makes the pre-generated schemes have clear optimization objectives, targeting only consecutive operator segments that may become bottlenecks on the critical path, thus avoiding the storage and computational overhead caused by blindly pre-generating the entire graph.

[0013] Furthermore, the cost calculation module includes: a migration time determination submodule, used to determine the migration time required to migrate the refactoring group from its current heterogeneous computing unit to other types of heterogeneous computing units in the heterogeneous computing resource pool; a migration delay determination submodule, used to determine the post-migration execution delay of the refactoring group on the migrated heterogeneous computing unit; a migration impact determination submodule, used to determine the impact cost of the migration operation on the execution of operators on non-critical paths; and a first comprehensive cost calculation submodule, used to weighted sum the migration time, post-migration execution delay, and impact cost to obtain the first comprehensive cost.

[0014] Based on the above technical means, when calculating the first comprehensive cost, not only the migration time is considered, but also the execution delay after migration and the impact cost of migration on non-critical path operators are comprehensively considered. This avoids making locally optimal but globally suboptimal decisions based on a single dimension (such as lower execution delay after migration), and ensures that the migration strategy improves the overall performance of the system.

[0015] Furthermore, the heterogeneous computing resource pool includes at least reconfigurable dataflow computing units, general-purpose processor units, and dedicated accelerator units; the dedicated accelerator units include at least one of GPU accelerator cards, NPU accelerator cards, and FPGA accelerator cards; the reconfigurable dataflow computing units are coarse-grained reconfigurable architecture arrays, containing multiple reconfigurable processing units and programmable data routing units, which are interconnected through programmable data routing units. Each reconfigurable processing unit triggers the start and end of computing tasks based on received valid data signals and control messages from upstream and downstream processing units. The reconfigurable dataflow computing units support partial reconfiguration of some reconfigurable processing units and interconnection relationships at runtime.

[0016] Based on the above technical means, the heterogeneous computing resource pool integrates reconfigurable dataflow computing units, general-purpose processor units, and various dedicated accelerator units. The reconfigurable dataflow computing units adopt a coarse-grained reconfigurable architecture, support runtime local reconfiguration, and provide the system with flexible hardware reconfigurability, making the local reconfiguration strategy physically feasible and efficient.

[0017] Furthermore, the cost calculation module also includes: a reconstruction time determination submodule, used to determine the reconstruction time required for local reconstruction of the reconstructible region corresponding to the reconstruction group in the reconstructible data flow calculation unit; a reconstruction delay determination submodule, used to determine the post-reconstruction execution delay of the reconstruction group on the reconstructed reconstructible region; a reconstruction impact determination submodule, used to determine the impact cost caused by the reconstruction operation on the execution of operators on non-critical paths; the impact cost is determined based on the data dependency width and the degree of competition for shared resources between the reconstruction group and the operators on non-critical paths; and a second comprehensive cost calculation submodule, used to weighted summation of the reconstruction time, post-reconstruction execution delay, and impact cost to determine the second comprehensive cost.

[0018] Based on the above technical means, when calculating the second comprehensive cost, the reconstruction time, the execution delay after reconstruction, and the cost of the impact of reconstruction on non-critical path operators (determined based on data dependency width and the degree of competition for shared resources) are weighted and summed to comprehensively evaluate the comprehensive cost of local reconstruction operations, so as to avoid the overall performance from declining due to excessive overhead of the reconstruction operation itself or excessive impact on non-critical paths.

[0019] Furthermore, the system also includes an update module, which updates the available resource status of each heterogeneous computing unit and the cache status of candidate reconfiguration schemes based on the execution results of the reconfiguration strategy.

[0020] Based on the above technical means, after the reconstruction strategy is executed, the available resource status of each heterogeneous computing unit and the cache status of candidate reconstruction schemes are updated, providing accurate basic data for subsequent reconstruction decisions and forming a closed-loop mechanism of "execution-feedback-update", enabling the system to continuously optimize the quality of decisions.

[0021] Furthermore, the update module includes: a hit update submodule, used to update the hit count and most recent hit time of a candidate reconstruction scheme when the reconstruction strategy is to load a matching candidate reconstruction scheme and execute it successfully; an add submodule, used to add the real-time generated reconstruction scheme to the cache as a new candidate reconstruction scheme when the reconstruction strategy is to generate a reconstruction scheme in real time and execute it successfully; and an eviction submodule, used to evict a candidate reconstruction scheme from the cache when the most recent hit time of any candidate reconstruction scheme in the cache is more than a preset period away from the current time and the hit count is less than a preset threshold.

[0022] Based on the aforementioned technical methods, a three-level cache management mechanism of hit update, addition, and eviction is implemented. Frequently hit candidate reconstruction schemes are given higher retention priority, while schemes with long-term misses and low hit counts are evicted. This achieves efficient utilization of cache space and avoids query efficiency degradation caused by cache bloat. Simultaneously, it adapts to the lightweight reconstruction scheduling requirements of distributed heterogeneous hardware nodes in edge computing.

[0023] Furthermore, the query submodule includes: an information acquisition unit, used to acquire the operator type, data dependency relationship, and input data scale of each operator in the reconstruction group; a preliminary selection unit, used to compare the operator type of each operator in the reconstruction group with the operator type corresponding to the operator mapping information in each candidate reconstruction scheme, and determine the candidate reconstruction scheme with completely consistent operator types as the preliminary selection scheme; a secondary selection unit, used to compare the data dependency relationship between each operator in the reconstruction group with the data dependency relationship corresponding to the operator mapping information in each preliminary selection scheme, and determine the preliminary selection scheme with consistent data dependency relationship as the secondary selection scheme; and a final selection unit, used to compare the input data scale of the reconstruction group with the reconstructable region capacity corresponding to the region division information in each secondary selection scheme, and select the secondary selection scheme with a reconstructable region capacity not less than the input data scale as the candidate reconstruction scheme matching the reconstruction group.

[0024] Based on the above technical means, a three-level progressive screening strategy of operator type matching, data dependency matching, and reconfigurable region capacity matching is adopted to gradually narrow down the candidate range from coarse granularity to fine granularity, ensuring that the matched candidate reconstruction schemes meet the requirements in terms of functional correctness and physical feasibility, and avoiding reconstruction failure due to inaccurate matching.

[0025] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written documents of this application.

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

[0027] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the architecture of a heterogeneous computing server system based on a reconfigurable data flow architecture, provided for an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of the internal structure of the reconstruction decision module in an embodiment of the present invention.

[0029] Figure 3 This is a schematic diagram of the internal structure of the pre-generated module in an embodiment of the present invention.

[0030] Figure 4 This is a schematic diagram of the internal structure of the cost calculation module in an embodiment of the present invention.

[0031] Figure 5This is a schematic diagram of another heterogeneous computing server system based on a reconfigurable data flow architecture, provided as an embodiment of the present invention.

[0032] Figure 6 This is a schematic diagram of the internal structure of the query submodule in an embodiment of the present invention. Detailed Implementation

[0033] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0034] To enable those skilled in the art to better understand the technical solutions of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0035] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0036] In the description of the embodiments of this application, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0037] This application provides a heterogeneous computing server system based on a reconfigurable data flow architecture. The system can monitor the latency of critical path operators during runtime and dynamically select the optimal strategy by comparing the combined costs of migration and local reconstruction when the latency exceeds the limit.

[0038] like Figure 1 As shown in the figure, this embodiment provides a heterogeneous computing server system based on a reconfigurable data flow architecture, including a task parsing module, an initial mapping module, a latency monitoring module, a reconfiguration group determination module, a cost calculation module, and a reconfiguration decision module.

[0039] The task parsing module receives computation tasks submitted from outside.

[0040] The aforementioned computational tasks can be computational loads with data flow characteristics, such as deep learning inference tasks, image processing pipelines, and data flow analysis tasks.

[0041] The task parsing module parses the computation task into a data flow graph. The data flow graph is represented as a directed acyclic graph (DAG), where each node represents an operator, and each directed edge represents the data dependency between operators. The task parsing module also extracts the topological features of the data flow graph, including the operation type of each operator (e.g., convolution, matrix multiplication, activation function, pooling, etc.), the input and output data sizes of each operator, the data dependencies between operators, and the execution order of operators on the critical path.

[0042] The input and output data scales of the above operators are the size of the data to be processed and the size of the data generated by each operator, expressed in bytes or tensor dimensions.

[0043] The data dependencies between the operators mentioned above are the input-output connections between operators in the data flow graph, that is, the directed relationship in which the output data of one operator serves as the input data of another operator.

[0044] The critical path refers to the path in the data flow graph with the largest sum of delays from the starting operator to the ending operator, which determines the theoretical shortest completion time of the entire computation task.

[0045] Specifically, the critical path is determined by traversing all paths from the start operator to the end operator in the data flow graph, summing the estimated execution delays for all operators on each path, and taking the path with the largest sum of delays as the critical path.

[0046] The estimated execution latency values ​​mentioned above can be obtained from a pre-established mapping table indexed by operation type, data size, and computing unit type. This mapping table is obtained by offline calibration and average statistics based on the actual execution latency data recorded by each computing unit during the execution of historical computing tasks.

[0047] The initial mapping module generates and executes an initial mapping scheme based on the computational characteristics of each operator in the data flow graph and the capability characteristics of each heterogeneous computing unit in the heterogeneous computing resource pool.

[0048] Computational characteristics include the type of operation, computational density, memory access mode, and data parallelism of the operators; capability characteristics include the peak computing power, memory bandwidth, supported operation types, and currently available resources of each computing unit.

[0049] Specifically, the initial mapping module selects computing units that support the operation type of each operator, and then matches the processing capabilities of the computing units according to the computing density and memory access mode. It prioritizes mapping computationally intensive operators with high data parallelism to GPUs or NPUs, maps operators with control logic intensive to general-purpose processors, and maps continuous operators with pipelined data dependencies to reconfigurable dataflow computing units.

[0050] The initial mapping module takes into account the above characteristics and assigns each operator to the most suitable heterogeneous computing unit to generate an initial mapping scheme.

[0051] The latency monitoring module monitors the execution latency of each operator on the critical path in real time. When the latency of an operator on the critical path exceeds the limit, that is, when the actual execution latency of a certain operator exceeds the latency threshold set for that operator, the latency monitoring module determines the heterogeneous computing unit where each operator on the critical path is currently located based on the initial mapping scheme, providing location information for subsequent reconfiguration decisions.

[0052] The aforementioned latency threshold is determined based on the historical execution latency statistics of the operator on the corresponding heterogeneous computing unit in the initial mapping scheme. For example, the 95th percentile of the historical execution latency of the operator on the corresponding computing unit is taken as the latency threshold.

[0053] After detecting excessive latency, the refactoring module determines the range of operators that need to be refactored.

[0054] Specifically, the reconstruction group determination module identifies multiple consecutive out-of-scale operators with data dependencies as a reconstruction group.

[0055] For example, if operators A, B, and C on the critical path have data dependencies and all exceed latency limits, then these three operators are identified as a reconstruction group. The data dependencies between consecutively exceeding operators mean that latency accumulates along the data dependency chain; therefore, reconstructing consecutively exceeding operators as a whole is more efficient than processing each operator individually.

[0056] The cost calculation module calculates the combined cost of the two reconstruction paths respectively.

[0057] The first comprehensive cost corresponds to the migration strategy of migrating the reconstructed group to other heterogeneous computing units, and the second comprehensive cost corresponds to the reconstructing strategy of performing local reconstructing of the corresponding region in the reconstructable data flow computing unit.

[0058] The cost calculation module comprehensively considers three dimensions: time cost, execution delay, and external impact, to ensure the global optimality of the decision.

[0059] The reconstruction decision module determines the final reconstruction strategy based on the comparison between the first and second comprehensive costs. When the first comprehensive cost is less than the second comprehensive cost, a migration strategy is selected; when the second comprehensive cost is less than or equal to the first comprehensive cost, a local reconstruction strategy is selected.

[0060] like Figure 2 As shown, the refactoring decision module includes a migration execution submodule, which is used to migrate the refactoring group to other heterogeneous computing units for execution when the first synthesis cost is less than the second synthesis cost.

[0061] The above migration operation keeps the original computational logic of the operator unchanged, only changing its execution location. The execution overhead is mainly the time for data movement and context switching.

[0062] Specifically, data transfer refers to transferring intermediate state data (such as the current completed calculation progress and internal cache data of operators) and input data to be processed from the local memory of the current computing unit to the memory of the target computing unit.

[0063] Context switching refers to saving the execution context of an operator on the current computing unit and then re-establishing the operator execution environment on the target computing unit. This includes loading the computing kernel corresponding to the operator, allocating memory space on the target computing unit, and establishing data paths with upstream and downstream operators.

[0064] Among them, the upstream operator refers to the operator that provides input data to the operators within the reconstruction group, and the downstream operator refers to the operator that receives output data from the operators within the reconstruction group.

[0065] In one possible implementation, during the migration process, the system needs to pause the execution of operators within the reconstruction group while keeping the upstream operators running and temporarily storing their output data in a buffer. Once the migration is complete, the buffered data is sent to the reconstruction group operator on the target computing unit. The downstream operators then wait for the reconstruction group to resume execution at the new location before continuing to receive output data.

[0066] For example, when a refactoring group originally runs on a general-purpose processor and is migrated to a GPU accelerator card, the migration time mainly includes the time to transfer the intermediate state data of the operator from the CPU memory to the GPU memory via the PCIe bus, and the initialization time to start the corresponding CUDA kernel on the GPU.

[0067] The system also includes a pre-generation module, which performs predictive analysis on critical paths in the data flow graph based on topological features before executing computational tasks, and pre-generates multiple candidate reconstruction schemes.

[0068] The aforementioned predictive analysis is completed before the task is officially executed and does not consume critical path time during runtime. Predictive analysis refers to the pre-generation module estimating the execution latency of each operator on various heterogeneous computing units based on the operation type, input data size, and output data size of each operator on the critical path, identifying operator segments that may become performance bottlenecks, and pre-calculating the optimal layout scheme on reconfigurable dataflow computing units for these operator segments.

[0069] Optionally, each candidate refactoring scheme includes at least the following information: the operator type and operator identifier of each operator in the candidate operator combination, the data dependency topology between operators, the target region partitioning on the refactorable dataflow computation unit (i.e., which refactorable processing units and programmable data routing units are allocated), the mapping relationship between each operator and the refactorable processing unit, and the estimated execution latency after refactoring. After the candidate refactoring scheme is generated, it is stored in a cache for easy retrieval and loading at runtime.

[0070] The reconstruction decision module also includes a query submodule, a cache loading submodule, and a real-time generation submodule, which are used to complete local reconstruction when the second comprehensive cost is less than or equal to the first comprehensive cost.

[0071] After determining that a local reconstruction strategy should be adopted, the query submodule queries whether there are any candidate reconstruction schemes that match the reconstruction group among the pre-generated candidate reconstruction schemes.

[0072] Candidate reconstruction schemes are pre-calculated and cached by the pre-generation module based on the topological characteristics of the critical path before task execution.

[0073] The query submodule employs a three-level progressive matching strategy: the first level matches operator types, meaning the set of operator types included in the operator mapping information of the candidate solution must be completely consistent with the operator types of each operator in the reconstruction group; the second level matches data dependencies, meaning the topology of data dependencies between operators in the candidate solution must be consistent with the topology of data dependencies between operators within the reconstruction group; and the third level matches resource capacity, meaning the number of available reconfigurable processing units and storage capacity of the target region in the candidate solution must meet the input data scale and computational resource requirements of the reconstruction group. The specific implementation of the three-level matching will be detailed in Example 8.

[0074] When a matching candidate reconstruction scheme exists, the cache loading submodule directly loads the scheme to configure the target region in the reconstructable data flow calculation unit, and after the configuration is completed, redeploys the operators in the reconstruction group to the target region and resumes execution.

[0075] The above configuration refers to loading the configuration bitstream recorded in the candidate reconstruction scheme into the target area of ​​the reconfigurable data stream computing unit. The configuration bitstream includes the operation function definition of each reconfigurable processing unit and the routing path definition of each programmable data routing unit within the target area.

[0076] Once configured, the data flow paths between reconfigurable processing units within the target area will perfectly match the data dependencies of each operator within the reconfiguration group.

[0077] The aforementioned redeployment refers to transmitting the current intermediate state data and pending input data of each operator within the reconstructed group from the current execution location to the target region, and binding each operator with the corresponding reconstructable processing unit.

[0078] Resumption of execution refers to the automatic start of computation in the data flow-driven mode by the reconfigurable processing unit within the target area after receiving a valid signal from the upstream data.

[0079] It should be understood that since the candidate reconstruction schemes have been pre-generated, the above loading and configuration process only needs to perform the transmission and writing operations of the configuration bit stream, without the need for real-time layout calculation and routing planning. Therefore, it can be completed in a very short time, significantly reducing reconstruction latency.

[0080] It should be understood that the time overhead of loading and configuration is typically in the hundreds of microseconds to milliseconds, while the time overhead of generating a refactoring solution in real time is typically in the seconds. Therefore, cache hits can reduce refactoring latency by at least two orders of magnitude.

[0081] In one possible implementation, the real-time generation submodule generates a reconstruction scheme in real time based on the data dependencies of each operator within the reconstruction group when no matching candidate reconstruction scheme exists.

[0082] It should be understood that real-time generation needs to take into account the data dependencies of operators, the hardware resource constraints of the target region, and the interconnection relationships with other operators. After generation, the target region is configured and executed.

[0083] In one possible implementation, the real-time generation process includes the following steps: I. Perform topological sorting of the data dependencies of each operator within the reconstructed group to determine the execution order of each operator on the reconstructable data flow computation unit.

[0084] II. Based on the operation type and input data scale of each operator, allocate processing units that meet the computing power requirements from the idle reconfigurable processing units of the reconfigurable data stream computing unit.

[0085] In this context, meeting the computing power requirements means that the set of operation types supported by the allocated reconfigurable processing unit includes the operation type of the operator, and the local storage capacity of the reconfigurable processing unit is not less than the sum of the input data size and the output data size of the operator, so as to ensure that the required computing functions and storage space for the operator to execute are met.

[0086] III. Based on the data dependencies between operators, plan the routing paths of programmable data routing units among the allocated reconfigurable processing units to ensure that data can flow between operators in the correct dependency order.

[0087] IV. Compile the above allocation results and routing planning results into a configuration bitstream and load it into the target area for configuration.

[0088] During real-time generation, it is also necessary to consider the interconnection between the reconstruction group and other running reconfigurable processing units outside the target area to avoid data path interruption or resource conflicts caused by the reconstruction operation to the operators being executed on non-critical paths.

[0089] Since real-time generation involves multiple computationally intensive operations such as layout calculation, route planning, constraint solving, and bitstream compilation, its time overhead is significantly higher than that of the cache loading method. However, it can still ensure that local reconstruction can be completed in scenarios without pre-generated scheme coverage, thereby improving the robustness of the system.

[0090] like Figure 3 As shown, the pre-generation module includes an estimation sub-module, a combination partitioning sub-module, and a scheme generation sub-module.

[0091] The estimation submodule determines the estimated execution latency of each operator on various heterogeneous computing units in the heterogeneous computing resource pool based on the operation type, input data size, and output data size of each operator on the critical path.

[0092] Among them, the estimated execution latency refers to the expected time for an operator to complete a single operation on a certain type of heterogeneous computing unit. It is a basic quantitative indicator for determining whether an operator becomes a performance bottleneck and generating candidate reconstruction schemes.

[0093] Specifically, the estimation submodule obtains the estimated execution delay by looking up a table. That is, a mapping table is pre-established from "operation type + data size + computing unit type" to "estimated execution delay". This mapping table is statistically analyzed and calibrated based on the historical execution data of each computing unit.

[0094] One possible implementation is to run standard test loads with different operation types and data sizes on each computing unit during the system offline phase, record the actual execution time for each execution, take the average of the execution times of multiple executions under the same combination of "operation type + data size + computing unit type" as the estimated execution latency of that combination, and periodically update the mapping table incrementally based on the newly added execution data to maintain the accuracy of the estimation.

[0095] Optionally, when there is no estimated execution delay for a certain combination in the mapping table, the estimation submodule can estimate it by interpolation, that is, by linear interpolation based on the estimated execution delay of the closest recorded combination (the same operation type on the same computing unit but adjacent data size), so as to ensure that a reasonable estimate can be obtained under any circumstances.

[0096] The combined partitioning submodule is based on the estimated execution latency. Operators on the critical path whose estimated execution latency exceeds a preset threshold are identified as candidate operators. The preset threshold is a latency boundary value used to determine whether an operator constitutes a performance bottleneck of the critical path. Operators exceeding this threshold are considered to have excessive execution latency on the current computing unit and are worth refactoring and optimization.

[0097] Specifically, the aforementioned preset threshold can be set according to the system's tolerance for latency. One possible implementation is to set the preset threshold as a preset proportion of the sum of the estimated execution latencies of all operators on the critical path. That is, when the estimated execution latency of a certain operator exceeds 10% of the total latency of the critical path, the operator is determined to be a candidate operator.

[0098] Another possible implementation is to set the preset threshold as an absolute time value, that is, operators with an estimated execution delay of more than 5 milliseconds are judged as candidate operators, where the selection of the absolute time value can be determined based on the system's maximum tolerance limit for single operator delay.

[0099] It should be understood that the selection of the above-mentioned preset ratio and the setting of the absolute time value need to be weighed in combination with the sensitivity to latency in the actual application scenario. In scenarios with higher latency sensitivity, the preset ratio or absolute time value should be set smaller in order to trigger the reconstruction earlier.

[0100] After determining the candidate operators, the combination partitioning submodule divides the candidate operators into one or more candidate operator combinations according to the continuity of data dependencies. The continuity of data dependencies means that there is a direct directed edge connection between two candidate operators in the data flow graph, that is, the output data of one candidate operator is used as the input data of another candidate operator, and there are no non-candidate operators between the two candidate operators, thus ensuring that the operators within the candidate operator combination constitute a continuous data processing pipeline segment.

[0101] For example, if the critical path contains operators A, B, C, D, and E in sequence, with estimated execution delays of 3 milliseconds, 8 milliseconds, 9 milliseconds, 4 milliseconds, and 10 milliseconds respectively, and the preset threshold is set to 5 milliseconds, then the candidate operators with estimated execution delays exceeding 5 milliseconds are B, C, and E. Among them, operators B and C have a direct data dependency relationship and there are no non-candidate operators between them, so they are classified as candidate operator combination {B, C}. Operator E and operator C are separated by non-candidate operator D (delay of 4 milliseconds, not exceeding the threshold), and the data dependency relationship is not continuous, so operator E is classified as candidate operator combination {E} separately.

[0102] It should be understood that grouping consecutive candidate operators into the same set rather than processing them separately is because there are data dependencies between consecutive operators. Reconstructing one operator while keeping the adjacent operators in their original positions would require data to be frequently transferred back and forth between the reconstructed region and the original execution location, which would increase additional communication overhead. Reconstructing consecutive operators as a whole can form a complete local data flow closed loop within the reconstructable data flow computing unit, maximizing the benefits of reconstruction.

[0103] The scheme generation submodule generates corresponding candidate reconstruction schemes based on each candidate operator combination. Each candidate reconstruction scheme includes the region division information and operator mapping information of the candidate operator combination on the reconfigurable data flow computing unit.

[0104] Specifically, the region partitioning information is used to specify which reconfigurable processing units in the reconfigurable data stream computing unit are assigned to the candidate operator combination. This includes the number of reconfigurable processing units assigned, their location identifiers, and the routing path planning between each reconfigurable processing unit. The location identifier is used to uniquely determine the coordinate position of each assigned reconfigurable processing unit in the two-dimensional array of reconfigurable data stream computing units. The operator mapping information is used to specify which assigned reconfigurable processing unit each operator in the candidate operator combination is mapped to for execution, i.e., establishing a one-to-one correspondence between operators and reconfigurable processing units. The mapping decision comprehensively considers the matching degree between the operator's operation type and the operation functions supported by the reconfigurable processing unit, the adaptability between the operator's input data size and the local storage capacity of the reconfigurable processing unit, and the topological consistency between the data dependencies between operators and the routing path between reconfigurable processing units. One possible implementation is that the scheme generation submodule adopts a heuristic layout algorithm, sorts the topology according to the data dependencies of each operator in the candidate operator combination, and assigns the optimal reconfigurable processing unit to each operator in sequence. It prioritizes selecting the idle reconfigurable processing unit that is closest to the previous operator in terms of routing distance and has matching operation function, thereby minimizing the data transmission delay within the reconfigurable data flow calculation unit.

[0105] The beneficial effects of this embodiment are as follows: By estimating the latency of each operator on the critical path through the estimation submodule, potential performance bottleneck operators can be identified before task execution, avoiding passive responses due to latency exceeding the limit only at runtime. The preparation time for the reconstruction scheme is brought forward from runtime to before task execution, without occupying the computation time of the critical path. By combining and partitioning the candidate operators with continuous data dependencies into the same combination through the combination and partitioning submodule, it is ensured that the operators form a local data flow closed loop within the reconstructible data flow computation unit after reconstruction, avoiding the additional communication overhead caused by cross-regional data transmission. By pre-generating a complete reconstruction scheme containing region partitioning and operator mapping through the scheme generation submodule, it can be directly loaded at runtime without real-time calculation of layout and routing, significantly reducing reconstruction latency.

[0106] like Figure 4 As shown, the cost calculation module includes a migration time determination submodule, a migration delay determination submodule, a migration impact determination submodule, and a first comprehensive cost calculation submodule.

[0107] The migration time determination submodule determines the migration time required to migrate the refactoring group from its current heterogeneous computing unit to other types of heterogeneous computing units in the heterogeneous computing resource pool. The migration time refers to the time interval between the migration decision being made and the refactoring group being able to start execution on the target computing unit, which includes two components: data transfer time and environment preparation time.

[0108] Specifically, data transmission time refers to the time required to transfer the intermediate state data of each operator in the reconstruction group (i.e., the computation progress information that the operator has completed, the internal cache data of the operator, and the input data that the operator has not yet processed) from the local memory of the current computing unit to the memory of the target computing unit. This time is equal to the total amount of data to be transmitted divided by the effective transmission bandwidth between the current computing unit and the target computing unit.

[0109] The total amount of data to be transmitted is the sum of the intermediate state data and unprocessed input data of all operators in the reconstruction group. The effective transmission bandwidth depends on the bus type (such as PCIe, NVLink, CCIX, etc.) connecting the two computing units and the current bus load.

[0110] The aforementioned environment preparation time refers to the time required to initialize the operator execution environment on the target computing unit. This includes loading the computing kernel program corresponding to the operator, allocating the memory space required for the operator to run on the target computing unit, establishing data paths with upstream and downstream operators, and setting the initial running parameters of the operator. This time is related to the initialization speed of the target computing unit and the complexity of the operator execution environment.

[0111] It should be understood that the larger the data volume and the lower the transmission bandwidth, the longer the data transmission time; the slower the initialization speed of the target computing unit and the more complex the operator execution environment, the longer the environment preparation time.

[0112] The migration delay determination submodule determines the post-migration execution delay of the refactoring group on the migrated heterogeneous computing unit. The post-migration execution delay refers to the expected time required for the refactoring group to complete a full execution on the target computing unit, and is used to evaluate whether the execution efficiency of the operators has been improved after migration.

[0113] Specifically, the execution latency after migration can be estimated based on the target computing unit's processing capacity for the corresponding operation type and the size of the input data.

[0114] One possible approach is to use historical execution data of similar operation types and data sizes on the target computing unit as the estimated value of the execution latency after migration, taking the statistical median of the historical execution latency.

[0115] For example, if the reconstruction group contains convolution operators and the target computing unit is a GPU accelerator card, the execution latency after migration can be estimated based on the measured performance data of convolution operations with the same kernel size and similar input feature map size running on the GPU. The estimation also needs to take into account the competitive impact of other tasks running on the GPU on computing resources.

[0116] It should be understood that the execution latency after migration is inversely related to the computing power of the target computing unit; the stronger the computing power of the target computing unit, the lower the execution latency after migration.

[0117] The migration impact determination submodule determines the impact cost of migration operations on operator execution on non-critical paths.

[0118] The impact cost refers to the additional latency caused by the migration operation during data transmission to operators executing on non-critical paths due to the occupation of shared system resources. It is used to quantify the indirect negative impact of the migration operation on the overall system performance.

[0119] Specifically, during data transfer, the migration operation requires the use of the system bus bandwidth connecting the current computing unit and the target computing unit. This bus bandwidth is a shared system resource. When operators on non-critical paths also need to transfer data through the same bus, bandwidth contention arises between the migration operation and the non-critical path operators, leading to increased data transfer latency for the non-critical path operators. In addition, the initialization operation of the target computing unit during the migration operation may also consume the computing resources of the target computing unit, affecting the execution efficiency of other non-critical path operators currently being executed on the target computing unit.

[0120] One possible implementation is that the migration impact determination submodule monitors the bus bandwidth and computing resources occupied by the migration operation, and calculates the sum of the latency increases of each operator on the non-critical path as the impact cost by combining the degree of dependence of each operator on the shared resources.

[0121] The first comprehensive cost calculation submodule calculates the first comprehensive cost by weighting and summing the migration time, post-migration execution delay, and impact cost.

[0122] The first comprehensive cost is a quantitative indicator of the total cost of the migration strategy, which is used to compare with the second comprehensive cost (the total cost of the local reconstruction strategy) to determine the optimal reconstruction strategy.

[0123] Specifically, the formula for calculating the first comprehensive cost is: First comprehensive cost = W1 × migration time + W2 × post-migration execution delay + W3 × impact cost, where W1, W2, and W3 are the weighting coefficients of migration time, post-migration execution delay, and impact cost, respectively, and the sum of the three weighting coefficients is 1.

[0124] W1 reflects the system's sensitivity to the time consumption of the migration operation itself. Migration requires moving the operator state, intermediate data, and execution context of the reconstructed group from the current heterogeneous computing unit to the target unit. This process interrupts the execution of the reconstructed group. A larger W1 indicates that the system is more averse to the immediate interruption latency caused by migration, which is suitable for real-time scenarios with stringent response speed requirements, such as autonomous driving and industrial control. W2 reflects the system's emphasis on the efficiency of operator operation after migration. The core purpose of migration is to transfer operators from the current underperforming computing unit to a more suitable target unit (such as migrating a serial bottleneck operator from the CPU to the GPU). A larger W2 indicates that the system is more focused on the long-term throughput gains after migration, which is suitable for continuous operation scenarios such as batch processing and scientific computing. W3 reflects the system's sensitivity to the disruption of global stability caused by the migration operation. Migrating to other heterogeneous computing units may trigger resource contention in the target unit, thereby affecting the execution of operators on non-critical paths. A larger W3 indicates that the system pays more attention to overall system stability, which is suitable for mixed load concurrency scenarios.

[0125] Optionally, the weighting coefficients can be dynamically adjusted based on the importance placed on each indicator in the actual application scenario. For example, in scenarios highly sensitive to latency, W2 (the weight of execution latency after migration) can be set higher, making the decision more inclined to choose the strategy with the lowest execution latency. In scenarios with high requirements for overall system throughput, W3 can be set higher, making the decision more inclined to choose the strategy with the least impact on non-critical path operators. It should be understood that the setting of the above weighting coefficients needs to be combined with the system optimization objectives of the actual application scenario, and the optimal weight combination should be determined through offline tuning or online adaptive adjustment.

[0126] It should be noted that in this embodiment, the impact cost is a unified mapping of the comprehensive impact of the migration operation on the execution of operators on non-critical paths into a time equivalent.

[0127] Specifically, based on the increased waiting time caused by resource contention for operators on non-critical paths due to migration operations, the overload delay caused by data cache invalidation, and the context switching overhead caused by task interruption, the above items are converted into time quantities in milliseconds or microseconds. Thus, migration time, post-migration execution delay, and impact cost have a unified time dimension and can be linearly weighted and summed in the same expression.

[0128] The beneficial effect of this embodiment is that by weighting and fusing the three dimensions of migration time, post-migration execution delay and impact cost to obtain the first comprehensive cost, it avoids making a migration decision that is locally optimal but globally suboptimal based on only a single dimension. It ensures that the comprehensive cost evaluation of the migration strategy can fully reflect the comprehensive effect of the migration operation on the system's time overhead, execution efficiency improvement and indirect impact, thereby taking into account both the execution acceleration of the operator itself and the overall throughput guarantee of the system in the migration decision.

[0129] In one embodiment, the heterogeneous computing resource pool includes at least a reconfigurable dataflow computing unit, a general-purpose processor unit, and a dedicated accelerator unit.

[0130] The heterogeneous computing resource pool is a collection of all hardware computing units in the system that can be used to perform computing tasks. Different types of computing units differ in computing power, applicable operation types, and memory access characteristics, and together they provide diverse computing resources for the system.

[0131] It should be understood that a general-purpose processor unit can be one or more multi-core CPUs, which are suitable for executing operators with complex control logic and dense branches. Complex control logic means that the execution of the operator contains a large number of condition judgments, loop control and state transition operations. Such operators are difficult to execute efficiently on dataflow-driven accelerators because dataflow architecture is naturally suitable for deterministic pipelined computation and has weak support for branch prediction and dynamic scheduling.

[0132] Specifically, the general-purpose processor unit efficiently processes control-intensive operators through mechanisms such as superscalar pipelines, branch prediction, and out-of-order execution. At the same time, its large-capacity cache and flexible memory access modes are also suitable for processing computational tasks with irregular data access patterns.

[0133] The dedicated accelerator unit includes at least one of a GPU accelerator card, an NPU accelerator card, and an FPGA accelerator card.

[0134] The GPU accelerator card contains thousands of lightweight computing cores and achieves large-scale parallel computing through a single-instruction multithreaded architecture. It is suitable for operators with high data parallelism, such as convolution operations and matrix multiplication in image processing. The NPU accelerator card has been specially optimized for matrix multiplication and accumulation operations in neural network inference. It has a built-in dedicated tensor computing unit and efficient data reuse path, and is suitable for scenarios with high requirements for latency and energy efficiency in deep learning inference tasks. The FPGA accelerator card (Field Programmable Gate Array accelerator card) consists of programmable logic units and programmable interconnect resources. It supports customized configuration of computing logic at the hardware level and is suitable for customized computing scenarios that require extremely low latency, deterministic execution, and custom data path optimization.

[0135] It should be understood that the types of the aforementioned dedicated accelerator units can be expanded according to actual deployment needs. The system supports the access of new types of accelerator units, which can be achieved by simply adding a description of the capability characteristics of the new type of accelerator unit to the operator-computation unit matching rules of the initial mapping module.

[0136] The reconfigurable data stream computing unit is a coarse-grained reconfigurable architecture array, which includes multiple reconfigurable processing units and programmable data routing units. The coarse-grained reconfigurable architecture refers to reconfigurable computing with the data width at the word level or even coarser, which is different from the fine-grained reconfigurable architecture with the bit level as the basic unit. It achieves a better balance between configuration speed and computing efficiency.

[0137] Specifically, the reconfigurable processing unit is the basic unit in the array that performs actual computational operations. Each reconfigurable processing unit contains an arithmetic logic operation unit, a local register file, and a local data storage space. It supports a variety of common operation types (such as addition, multiplication, comparison, shift, and logical operations). Its operation function can be switched by loading different configuration information.

[0138] The programmable data routing unit is an interconnecting unit in the array responsible for transmitting data between reconfigurable processing units. It supports programming configuration of data routing paths and can dynamically establish data paths between reconfigurable processing units based on data dependencies between operators, so that data arrives at the correct target processing unit at the correct time.

[0139] Each reconfigurable processing unit is interconnected through a programmable data routing unit to form a flexibly configurable two-dimensional grid-like data flow network, in which any two reconfigurable processing units can communicate through a data link composed of one or more programmable data routing units.

[0140] Each reconfigurable processing unit triggers the start and end of the computing task based on the received valid data signal and the control messages from upstream and downstream processing units. That is, it adopts a data flow driven rather than control flow driven execution model. Data flow driven means that the start condition of the computing unit is not controlled by the unified clock or instruction of the central controller, but is triggered by the arrival event of input data. The subsequent processing unit will automatically start computing only after the preceding processing unit generates output data and sends a valid data signal to the subsequent processing unit.

[0141] Specifically, the aforementioned data validity signal is a flag signal generated by the reconfigurable processing unit hardware, indicating that the output data of that processing unit is ready and valid, and can be used by downstream processing units. The control messages between upstream and downstream processing units refer to handshake signals transmitted between reconfigurable processing units via dedicated control signal lines, used to coordinate the data transmission rhythm and backpressure control between different stages of the pipeline. It should be understood that adopting a dataflow-driven execution model eliminates the need for a central controller to uniformly schedule the reconfigurable dataflow computing units. Each processing unit works autonomously and collaboratively in a decentralized manner, thereby eliminating the control overhead and scheduling bottlenecks caused by central scheduling, while naturally supporting parallel pipeline execution.

[0142] The reconfigurable dataflow computing unit supports partial reconfiguration of some reconfigurable processing units and interconnections at runtime. Partial reconfiguration means that while the reconfigurable dataflow computing array as a whole remains running, only one or more selected reconfigurable processing units and their related programmable data routing units in the array are configured and updated, without affecting the normal operation of other unselected reconfigurable processing units in the array.

[0143] Specifically, local reconfiguration is achieved in the following way: the reconfigurable data flow computing unit is physically divided into multiple independently configurable configuration areas. Each configuration area has an independent configuration interface and configuration controller. When a configuration area needs to be reconfigured, the system only sends a new configuration bit stream to the configuration controller of that configuration area. The reconfigurable processing unit and programmable data routing unit in that configuration area update their functions after receiving the configuration bit stream, while the processing units in other configuration areas continue to execute their currently assigned computing tasks normally.

[0144] It should be understood that the local reconfiguration capability is the core feature that distinguishes the reconfigurable dataflow computing unit from the traditional static configuration accelerator. It enables the reconfiguration operation to affect only a local area in the array, while other areas can continue to provide computing services for operators on non-critical paths in the system. This minimizes the impact of reconfiguration overhead on the overall system throughput and achieves dynamic runtime optimization while maintaining the overall availability of the system.

[0145] The beneficial effects of this embodiment are as follows: by integrating general-purpose processor units, various dedicated accelerator units, and reconfigurable dataflow computing units, the system forms a heterogeneous computing resource pool covering three types of computing modes: control-intensive, compute-intensive, and dataflow-intensive, providing a diverse range of operator deployment options for the initial mapping module; the reconfigurable dataflow computing unit adopts a dataflow-driven execution model, eliminating central scheduling overhead and naturally supporting pipelined parallelism. At the same time, its runtime local reconfiguration capability enables the system to reconfigure local areas without downtime, taking into account both runtime dynamic optimization flexibility and overall system availability.

[0146] like Figure 4 As shown, the cost calculation module also includes a reconstruction time determination submodule, a reconstruction delay determination submodule, a reconstruction impact determination submodule, and a second comprehensive cost calculation submodule.

[0147] The reconstruction time determination submodule determines the reconstruction time required to perform local reconstruction of the reconstructible region corresponding to the reconstruction group in the reconstructible data stream calculation unit.

[0148] The reconstruction time refers to the time interval between the local reconstruction decision being made and the target reconfigurable region being configured and ready to start executing the reconstruction group operator. It includes two components: configuration bitstream loading time and interconnection relationship reconstruction time.

[0149] Specifically, the configuration bitstream loading time refers to the time required to transfer the new configuration bitstream from the configuration memory to the configuration interface of the target reconfigurable region and write it into each configuration register. The configuration bitstream is a binary data sequence that describes the operational function definition of each reconfigurable processing unit and the routing path definition of each programmable data routing unit within the target reconfigurable region. Its size depends on the number of reconfigurable processing units and programmable data routing units within the target reconfigurable region. The configuration interface bandwidth is the maximum transmission rate of the configuration bitstream on the transmission channel.

[0150] Interconnection reconstruction time refers to the time required for the programmable data routing units within the target reconfigurable region to re-establish routing connections between the reconfigurable processing units and between the target region and other regions in the reconfigurable data stream computing array after the configuration bitstream loading is completed. This time is related to the complexity of the routing path and the configuration speed of the programmable data routing units.

[0151] It should be understood that the reconstruction time is directly proportional to the size of the configuration bitstream and inversely proportional to the bandwidth of the configuration interface. The smaller the configuration bitstream and the higher the configuration interface bandwidth, the shorter the reconstruction time.

[0152] The data dependency width refers to the number of directed edges in the data flow graph that have data dependencies between operators within the reconstruction group and operators on non-critical paths. Each directed edge corresponds to a transmission channel of a tensor, that is, a directed edge represents a one-way data transmission relationship of a specific tensor from the output of an operator within the reconstruction group to an operator on a non-critical path or from the output of an operator on a non-critical path to an operator within the reconstruction group. The larger the data dependency width, the wider the impact of the reconstruction operation on operators on non-critical paths.

[0153] The Reconstruction Delay Determination Submodule determines the post-reconstruction execution delay of the reconstruction group on the reconstructable region after reconstruction. The post-reconstruction execution delay refers to the expected time required for the reconstruction group to complete a full execution on the target region of the reconstructable data flow computing unit after completing the local reconstruction.

[0154] Specifically, after reconstruction, the data flow paths between reconfigurable processing units within the reconfigurable data flow computing unit are specifically customized for the data dependencies of each operator within the reconfiguration group. The data flow path between operators is the shortest and there are no redundant intermediate routing nodes. Therefore, the execution latency is usually lower than the execution latency of the reconfiguration group at the original execution location (general-purpose processor or dedicated accelerator).

[0155] One possible implementation is that the reconstruction delay determination submodule calculates the estimated value of the execution delay after reconstruction by accumulating the data dependency relationship path of each operator in the reconstruction group level by level, based on the region division information and operator mapping information in the reconstruction scheme, combined with the operation delay of each reconfigurable processing unit in the reconfigurable data flow computing unit and the routing delay of the programmable data routing unit.

[0156] It should be understood that the accuracy of the estimated execution delay after refactoring directly affects the accuracy of the second comprehensive cost, and thus the reliability of the refactoring decision.

[0157] The Reconstruction Impact Determination Submodule determines the impact cost of the reconstruction operation on the execution of operators on non-critical paths. The impact cost refers to the additional delay caused by the reconstruction operation during the configuration bitstream loading and interconnection reconstruction, which consumes the configuration resources of the reconstructable data stream computing unit and affects the operators being executed on non-critical paths.

[0158] Specifically, the impact cost is determined based on the data dependency width and the degree of competition for shared resources between the operators in the refactoring group and the operators on the non-critical path. The data dependency width refers to the number of data links in the data flow graph that have data dependencies between the operators in the refactoring group and the operators on the non-critical path. That is, how many input data of the operators on the non-critical path come from the operators in the refactoring group or how many output data flow to the operators in the refactoring group. The larger the data dependency width, the wider the impact of the refactoring operation on the operators on the non-critical path.

[0159] The degree of competition for shared resources refers to whether operators on non-critical paths compete with the reconfiguration group for the same programmable data routing unit or reconfigurable processing unit. When the programmable data routing unit that needs to be reconfigured for the reconfiguration operation happens to be located on the data path of the operator on the non-critical path, the data flow transmission of the operator on the non-critical path will be interrupted until the route reconstruction is completed. The higher the degree of competition, the greater the impact on the operator on the non-critical path.

[0160] It should be understood that the greater the breadth of data dependencies and the higher the degree of competition for shared resources, the greater the impact and cost of refactoring on operators on non-critical paths.

[0161] The second comprehensive cost calculation submodule calculates the weighted sum of reconstruction time, post-reconstruction execution delay, and impact cost to determine the second comprehensive cost. The second comprehensive cost is a quantitative indicator of the total cost of the local reconstruction strategy, which is used to compare with the first comprehensive cost (the total cost of the migration strategy) to determine the optimal reconstruction strategy.

[0162] Specifically, the formula for calculating the second comprehensive cost is: Second comprehensive cost = W4 × Reconstruction time + W5 × Post-reconstruction execution delay + W6 × Impact cost, where W4, W5, and W6 are the weighting coefficients for reconstruction time, post-reconstruction execution delay, and impact cost, respectively, and the sum of the three weighting coefficients is 1.

[0163] W4 reflects the system's sensitivity to the time consumption of the reconstruction operation itself. Reconstruction requires loading a bitstream to reconfigure the reconfigurable array, which usually takes longer than migration. Therefore, the design value of W4 is generally lower than W1 in the first synthesis cost to avoid the system almost never choosing the reconstruction scheme due to excessive aversion to reconstruction time. W5 reflects the system's emphasis on the efficiency of operator operation after reconstruction. Reconstruction customizes a dedicated hardware data path for operators, and the performance improvement is usually better than simple migration to other general heterogeneous units. Therefore, W5 can be relatively increased to reflect the pursuit of long-term performance benefits. W6 reflects the system's sensitivity to the global stability of the reconstruction operation. Reconstruction only affects the local area in the reconfigurable array corresponding to the reconstruction group. The area where the operator is located on the non-critical path is not affected. Its impact range is smaller than that of the migration operation (migration may cause resource contention and data movement between heterogeneous units). Therefore, the design value of W6 is usually lower than W3 in the first synthesis cost.

[0164] Optionally, the aforementioned weighting coefficients can be set independently of the weighting coefficients (W1, W2, W3) of the first comprehensive cost. This is because the mechanisms by which reconstruction and migration operations affect the system are different. For example, the reconstruction time of a reconstruction operation is usually longer than the migration time of a migration operation (because it involves hardware reconfiguration), but the execution latency after reconstruction is usually lower than that after migration (because the data flow path is more customized). Therefore, in practical applications, if the system is more concerned about the execution efficiency after reconstruction, W5 can be set higher; if the system is more concerned about the interruption time of online services during the reconstruction process, W4 can be set higher.

[0165] It should be understood that the weighting coefficients of the first comprehensive cost (W1, W2, W3) and the weighting coefficients of the second comprehensive cost (W4, W5, W6) can be configured independently to accommodate the essential differences in cost structure between the migration strategy and the reconstruction strategy.

[0166] It should be noted that the first and second comprehensive costs are handled using a unified dimensional approach. Specifically, migration time and post-migration execution delay are both measured in time units. The impact cost includes resource contention waiting time, data reload delay, and task context switching overhead caused by the migration operation on operators executing on non-critical paths. All of these items are converted into a unified time equivalent quantity through a pre-defined mapping function. Similarly, reconstruction time and post-reconstruction execution delay are both measured in time units. The impact cost includes additional waiting delay caused by the reconstruction operation on operators executing on non-critical paths, communication delay increment caused by data routing reconnection, and increased access delay caused by shared resource contention. These items are also converted into a unified time equivalent quantity through a mapping function. Thus, in each cost calculation formula, all variables participate in a linear weighted summation under the same time dimension. The weight coefficients W1, W2, W3 and W4, W5, W6 are all dimensionless constants, and their sum is 1, respectively representing the system's relative preference for the three indicators of immediate interruption time, long-term performance gains, and global stability impact under the corresponding scheme.

[0167] The beneficial effects of this embodiment are as follows: by weighting and fusing the three dimensions of reconstruction time, post-reconstruction execution delay, and impact cost to obtain the second comprehensive cost, the comprehensive cost of local reconstruction operations is comprehensively evaluated. In particular, by quantifying the impact cost of reconstruction on non-critical path operators through the two dimensions of data dependency width and shared resource competition, the reconstruction decision not only considers the direct overhead of reconstruction itself, but also the indirect impact on the overall system operation status, thus avoiding the "robbing Peter to pay Paul" problem caused by the excessive impact of reconstruction operations on non-critical path operators.

[0168] like Figure 5As shown, the system also includes an update module, which updates the available resource status of each heterogeneous computing unit and the cache status of candidate reconfiguration schemes based on the execution results of the reconfiguration strategy.

[0169] The available resource status of each heterogeneous computing unit refers to the amount of computing resources (such as remaining computing power, remaining memory bandwidth, and remaining storage space) that are currently available for allocating new operators in each computing unit. The cache status of the candidate reconstruction scheme refers to the number of hits, the most recent hit time, and the validity status of the candidate reconstruction scheme generated by the pre-generation module in the cache.

[0170] The update module includes hit update submodule, add submodule, and deprecation submodule.

[0171] If the refactoring strategy is to load a matching candidate refactoring scheme and execute it successfully, the hit update submodule updates the hit count and the most recent hit time of the candidate refactoring scheme. The hit count refers to the total number of times the candidate refactoring scheme has been successfully loaded and executed since its generation, and the most recent hit time refers to the timestamp of the most recent successful loading and execution of the candidate refactoring scheme.

[0172] Specifically, the hit update submodule increments the hit count of the candidate reconstruction scheme by one based on the current value and updates the most recent hit time to the current system timestamp. The hit count and the most recent hit time together constitute the popularity evaluation index of the candidate reconstruction scheme, which is an important basis for the cache elimination submodule to make cache elimination decisions.

[0173] It should be understood that a higher number of hits and a more recent hit time closer to the current time indicate that the candidate reconstruction scheme has been frequently used recently and has high retention value.

[0174] If the refactoring strategy is to generate a refactoring solution in real time and the refactoring is executed successfully, the submodule adds the real-time generated refactoring solution to the cache as a new candidate refactoring solution. The real-time generated refactoring solution refers to the refactoring solution dynamically calculated and generated by the real-time generation submodule at runtime based on the specific situation of the refactoring group, rather than the solution pre-generated by the pre-generation module before the task is executed.

[0175] Specifically, the initial hit count for newly added candidate reconstruction schemes is set to 1 (indicating that the scheme has been successfully used once), the most recent hit time is set to the current system timestamp, and the scheme content includes the region division information and operator mapping information determined during the real-time generation process. By adding the real-time generated scheme to the cache, the system can directly hit the scheme from the cache without performing real-time generation again when it encounters a reconstruction requirement with the same combination of operator types and the same data dependency topology as the reconstruction group. This distributes the one-time computational overhead of real-time generation across multiple reuses, improving the overall efficiency of the system.

[0176] It should be understood that before adding a new solution, the submodule needs to check whether there is already a candidate refactoring solution in the cache that is exactly the same as the solution in terms of operator type combination and data dependency topology. If it already exists, it will not be added again, and only the hit count and the most recent hit time of the existing solution will be updated.

[0177] The eviction submodule periodically checks the candidate reconstruction schemes in the cache. When the most recent hit time of any candidate reconstruction scheme in the cache is more than a preset period away from the current time and the number of hits is less than a preset threshold, the candidate reconstruction scheme is evicted from the cache. The preset period is the time threshold at which the system determines whether a candidate reconstruction scheme is "expired", and the preset threshold is the minimum number of hits required by the system to determine whether a candidate reconstruction scheme is "unpopular".

[0178] Specifically, one possible implementation is that the elimination submodule traverses all candidate reconstruction schemes in the cache every preset check interval (e.g., every 1 hour) and performs two judgments on each scheme: the first judgment is the timeliness judgment, that is, checking whether the most recent hit time of the scheme is more than a preset period (e.g., 24 hours) from the current system time. If it is, it means that the scheme has not been used recently and may no longer be suitable for the current computing task characteristics; the second judgment is the popularity judgment, that is, checking whether the number of hits of the scheme is less than a preset threshold (e.g., 3 times). If it is less than that, it means that the scheme has rarely been used since it was generated and has low reuse value.

[0179] The eviction submodule removes the eviction scheme from the cache and releases the cache storage space it occupies only when a candidate reconstruction scheme meets both the conditions of "timeliness expired" and "insufficient popularity". This dual eviction strategy of "timeliness + popularity" ensures that the cache retains high-value schemes that are frequently used recently, avoids the cache space being occupied by outdated schemes that have not been used for a long time, and prevents the accidental eviction of new schemes that have been generated recently but have not yet accumulated enough hits.

[0180] Optionally, the eviction submodule can also prioritize eviction of the scheme with the earliest recent hit time and the lowest hit count when the cache space usage exceeds the preset limit, so as to ensure efficient use of cache space.

[0181] The beneficial effects of this embodiment are as follows: by implementing a three-level cache management mechanism of hit update, addition and eviction, the full lifecycle management of candidate reconstruction scheme cache is realized, which enables high-value schemes that frequently hit to be retained for a long time, new schemes generated in real time to be added to the cache for subsequent reuse, and old schemes that have not been used for a long time and have a low hit count to be automatically eliminated. This maximizes the cache hit rate within the limited cache space, reduces the frequency of real-time generation of reconstruction schemes, and reduces the delay of reconstruction decision at runtime.

[0182] like Figure 6As shown, the query submodule includes an information acquisition unit, a preliminary selection unit, a secondary selection unit, and a final selection unit.

[0183] The information acquisition unit acquires the operator type, data dependency relationship and input data scale of each operator in the reconstruction group. The operator type refers to the operation type identifier of the operator (such as convolution, pooling, fully connected, matrix multiplication, activation function, etc.), the data dependency relationship refers to the input-output connection topology between each operator in the reconstruction group, and the input data scale refers to the total amount of input data received by the reconstruction group as a whole.

[0184] It should be understood that the above three pieces of information are the basic data for subsequent three-level matching queries, corresponding to the three matching dimensions of the three-level matching. The operator type is used for the first-level initial selection matching, the data dependency relationship is used for the second-level secondary selection matching, and the input data scale is used for the third-level final selection matching.

[0185] The initial selection unit compares the operator types of each operator in the reconstruction group with the operator types corresponding to the operator mapping information in each candidate reconstruction scheme. The candidate reconstruction scheme with completely consistent operator types is determined as the initial selection scheme. Completely consistent operator types mean that the set of operator types contained in the operator mapping information of the candidate reconstruction scheme is exactly the same as the set of operator types of each operator in the reconstruction group. That is, the operator types in the two sets correspond one-to-one with no omissions and no redundancies.

[0186] For example, if the reconstruction group contains three operators {convolution, pooling, fully connected}, only the candidate reconstruction schemes whose operator mapping information also happens to contain these three operator types (and the number is the same, i.e., one each of convolution, pooling, and fully connected) will pass the initial selection. If the candidate scheme contains {convolution, convolution, pooling} (one more convolution and one less fully connected) or {convolution, pooling, fully connected, activation function} (one more activation function), then neither will pass the initial selection.

[0187] It should be understood that operator type matching is the first level of screening and the coarsest level of screening. Its purpose is to quickly eliminate large quantum candidate schemes with mismatched operator types. This is because the computational function configuration of each reconfigurable processing unit in the reconfigurable data flow computation unit is bound to the operator type. Schemes with inconsistent operator types cannot have their reconfigurable processing unit configurations reused by the reconfiguration group. Even if they are forcibly loaded, it will lead to errors in computational function.

[0188] The selection unit compares the data dependencies between operators in the reconstruction group with the data dependencies corresponding to the operator mapping information in each preliminary scheme. The preliminary scheme with consistent data dependencies is determined as the selection scheme. Consistent data dependencies mean that the topology of data dependencies between operators in the reconstruction group is isomorphic to the topology of data dependencies between operators in the preliminary scheme, that is, the directed edge connection relationship between nodes in the two topologies is completely consistent.

[0189] Specifically, the selection unit compares the data dependencies between operators in the reconstruction group as a directed graph G1 (nodes are operators, edges are data dependencies), and the data dependencies between operators in the initial selection scheme are represented as a directed graph G2 (nodes are operators, edges are data dependencies). Then, it is determined whether G1 and G2 are isomorphic, that is, whether there is a one-to-one mapping from nodes in G1 to nodes in G2, such that the directed edge relationship between any two nodes in G1 also holds between the corresponding nodes in G2.

[0190] It should be understood that data dependencies determine the data routing paths between operators in the reconfigurable data flow computation unit. Even if the operator types are the same, the data routing paths between reconfigurable processing units cannot be reused if the dependencies are inconsistent. If forced loading is attempted, it will cause the data to flow in the wrong order between operators, resulting in reconstruction failure or execution results that do not meet expectations.

[0191] The final selection unit compares the input data size of the reconstruction group with the reconfigurable region capacity corresponding to the region division information in each selection scheme, and selects the selection scheme with a reconfigurable region capacity not less than the input data size as the candidate reconstruction scheme that matches the reconstruction group. The reconfigurable region capacity refers to the total available storage capacity of the target reconfigurable region in the candidate reconstruction scheme, which is the sum of the local storage space of all reconfigurable processing units in the reconfigurable region. The input data size refers to the total amount of input data that the first operator in the reconstruction group needs to receive.

[0192] Specifically, the reconfigurable region capacity needs to meet two conditions: First, the reconfigurable region capacity must be no less than the input data size of the reconfigurable group to ensure that all the input data of the reconfigurable group can be loaded into the local storage of the target reconfigurable region, avoiding the need for data to be repeatedly read from external memory due to insufficient storage space, thereby losing the advantage of local data reuse of the data flow computing unit; Second, the number of reconfigurable processing units in the reconfigurable region must be able to allocate at least one reconfigurable processing unit to each operator in the reconfigurable group.

[0193] If multiple alternative schemes satisfy both of the above conditions, the final selection unit chooses the scheme with the reconfigurable region capacity closest to the input data size as the final matching scheme. This maximizes the resource utilization of the reconfigurable data stream computing units, avoiding resource waste caused by allocating too many reconfigurable processing units to the reconstruction group, and allowing the remaining reconfigurable processing units to continue serving other computing tasks. It should be understood that if no alternative scheme has a reconfigurable region capacity that meets the input data size requirement, the final selection unit determines that there is no matching candidate reconstruction scheme. In this case, the query submodule returns a "missed" result, and the system will then call the real-time generation submodule for real-time generation.

[0194] The beneficial effects of this embodiment are as follows: by employing a three-level progressive screening strategy of operator type matching, data dependency matching, and reconfigurable region capacity matching, the range of candidate solutions is gradually narrowed from coarse-grained to fine-grained. Each level of screening is based on explicit physical constraints, ensuring that the finally matched candidate reconstruction solutions meet the requirements in terms of functional correctness, data flow correctness, and physical feasibility. This avoids reconstruction failure or resource waste caused by inaccurate matching. At the same time, the computational complexity of the three-level progressive screening is much lower than that of comparing all candidate solutions one by one across all dimensions, achieving high query performance while ensuring matching accuracy.

[0195] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above content is merely a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0196] It should be noted that the structure of the above embodiments does not constitute a limitation on the system. Those skilled in the art will understand that the structure of the above embodiments is only related to some of the solutions in this application. The embodiments of this application can be extended or modified based on the structure of the above embodiments, and this application is not limited thereto.

[0197] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0198] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0199] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from the spirit and scope of this invention, and this invention is also intended to include these modifications and variations.

Claims

1. A heterogeneous computing server system based on a reconfigurable dataflow architecture, characterized in that, include: The task parsing module is used to receive computing tasks, parse the computing tasks into data flow graphs, and extract the topological features of the data flow graphs. The topological features include the operation type of each operator in the data flow graph, the input data size and output data size of each operator, the data dependencies between operators, and the execution order of each operator on the critical path; the critical path is the path with the largest sum of delays from the starting operator to the ending operator in the data flow graph. The initial mapping module is used to generate and execute an initial mapping scheme based on the computational characteristics of each operator in the data flow graph and the capability characteristics of each heterogeneous computing unit in the heterogeneous computing resource pool. The latency monitoring module is used to determine the heterogeneous computing unit where each operator on the critical path is currently located based on the initial mapping scheme when the operator latency on the critical path exceeds the limit. The reconstruction group determination module is used to identify multiple consecutive out-of-scale operators with data dependencies as a reconstruction group; The cost calculation module is used to determine the first comprehensive cost of migrating the reconstruction group to other heterogeneous computing units, and to determine the second comprehensive cost of performing local reconstruction of the corresponding region in the reconfigurable data flow computing unit. The reconstruction decision module is used to determine the reconstruction strategy based on the first comprehensive cost and the second comprehensive cost.

2. The heterogeneous computing server system based on a reconfigurable data flow architecture according to claim 1, characterized in that, The reconstruction decision module includes: The migration execution submodule is used to migrate the refactoring group to other heterogeneous computing units for execution when the first synthesis cost is less than the second synthesis cost.

3. The heterogeneous computing server system based on a reconfigurable data flow architecture according to claim 1, characterized in that, Also includes: The pre-generation module is used to perform predictive analysis on the critical paths in the data flow graph based on the topological features before executing the computation task, and to pre-generate multiple candidate reconstruction schemes. The reconstruction decision module also includes: The query submodule is used to query whether there is a candidate reconstruction scheme that matches the reconstruction group among the pre-generated candidate reconstruction schemes when the second comprehensive cost is less than or equal to the first comprehensive cost. The cache loading submodule is used to load the matching candidate reconstruction scheme when a matching candidate reconstruction scheme exists, configure the target region in the reconfigurable data flow calculation unit, and redeploy the operators in the reconstruction group to the target region and resume execution after the configuration is completed. The real-time generation submodule is used to generate a reconstruction scheme in real time based on the data dependency relationship of each operator in the reconstruction group when no matching candidate reconstruction scheme exists, and to configure and execute the target region.

4. The heterogeneous computing server system based on a reconfigurable data flow architecture according to claim 3, characterized in that, The pre-generation module includes: The estimation submodule is used to determine the estimated execution latency of each operator on various heterogeneous computing units in the heterogeneous computing resource pool based on the operation type, input data scale and output data scale of each operator on the critical path. The combined partitioning submodule is used to identify operators whose estimated execution delay on the critical path exceeds a preset threshold as candidate operators based on the estimated execution delay, and to partition the candidate operators into one or more candidate operator combinations according to the continuity of the data dependency relationship. The scheme generation submodule is used to generate corresponding candidate reconstruction schemes based on each candidate operator combination. Each candidate reconstruction scheme includes the region division information and operator mapping information of the candidate operator combination on the reconfigurable data stream computing unit.

5. The heterogeneous computing server system based on a reconfigurable data flow architecture according to claim 1, characterized in that, The cost calculation module includes: The migration time determination submodule is used to determine the migration time required to migrate the reconfiguration group from its current heterogeneous computing unit to other types of heterogeneous computing units in the heterogeneous computing resource pool. The migration delay determination submodule is used to determine the post-migration execution delay of the reconfiguration group on the migrated heterogeneous computing units; The migration impact determination submodule is used to determine the impact cost of the migration operation on the execution of operators on non-critical paths; The first comprehensive cost calculation submodule is used to weight and sum the migration time, the post-migration execution delay, and the impact cost to obtain the first comprehensive cost.

6. The heterogeneous computing server system based on a reconfigurable data flow architecture according to claim 1, characterized in that, The heterogeneous computing resource pool includes at least the reconfigurable dataflow computing unit, the general-purpose processor unit, and the dedicated accelerator unit; The dedicated accelerator unit includes at least one of a GPU accelerator card, an NPU accelerator card, and an FPGA accelerator card; The reconfigurable data stream computing unit is a coarse-grained reconfigurable architecture array, which includes multiple reconfigurable processing units and programmable data routing units. The reconfigurable processing units are interconnected through the programmable data routing units. Each reconfigurable processing unit triggers the start and end of the computing task based on the received data validity signal and the control messages of the upstream and downstream processing units. The reconfigurable data stream computing unit supports partial reconfiguration of some reconfigurable processing units and interconnection relationships during runtime.

7. The heterogeneous computing server system based on a reconfigurable data flow architecture according to claim 1, characterized in that, The cost calculation module also includes: The reconstruction time determination submodule is used to determine the reconstruction time required to perform local reconstruction of the reconstructible region corresponding to the reconstruction group in the reconstructible data stream computing unit. The reconstruction delay determination submodule is used to determine the post-reconstruction execution delay performed by the reconstruction group on the reconstructed reconstructable region; The refactoring impact determination submodule is used to determine the impact cost of the refactoring operation on the execution of operators on non-critical paths; the impact cost is determined based on the data dependency width and the degree of competition for shared resources between the refactoring group and the operators on the non-critical paths. The second comprehensive cost calculation submodule is used to calculate the second comprehensive cost by weighted summing of the reconstruction time, the post-reconstruction execution delay, and the impact cost.

8. The heterogeneous computing server system based on a reconfigurable data flow architecture according to claim 1, characterized in that, Also includes: The update module is used to update the available resource status of each heterogeneous computing unit and the cache status of candidate reconstruction schemes based on the execution result of the reconstruction strategy.

9. The heterogeneous computing server system based on a reconfigurable data flow architecture according to claim 8, characterized in that, The update module includes: The hit update submodule is used to update the hit count and most recent hit time of the candidate reconstruction scheme when the reconstruction strategy is to load a matching candidate reconstruction scheme and execute it successfully. Add a submodule to add the real-time generated reconstruction scheme to the cache as a new candidate reconstruction scheme when the reconstruction strategy is to generate a reconstruction scheme in real time and execute it successfully. The elimination submodule is used to eliminate the candidate reconstruction scheme from the cache when the most recent hit time of any candidate reconstruction scheme in the cache exceeds a preset period and the number of hits is less than a preset threshold.

10. The heterogeneous computing server system based on a reconfigurable data flow architecture according to claim 3, characterized in that, The query submodule includes: An information acquisition unit is used to acquire the operator type, data dependency relationship and input data scale of each operator in the reconstruction group; The initial selection unit is used to compare the operator type of each operator in the reconstruction group with the operator type corresponding to the operator mapping information in each candidate reconstruction scheme, and to determine the candidate reconstruction scheme with completely consistent operator types as the initial selection scheme. The selection unit is used to compare the data dependency relationship between each operator in the reconstruction group with the data dependency relationship corresponding to the operator mapping information in each initial selection scheme, and to determine the initial selection scheme with consistent data dependency relationship as the selection scheme. The final selection unit is used to compare the input data size of the reconstruction group with the reconfigurable region capacity corresponding to the region division information in each selection scheme, and select the selection scheme with a reconfigurable region capacity not less than the input data size as the candidate reconstruction scheme that matches the reconstruction group.