Method for controlling multi-user concurrent access to a server

Abstract

The present application relates to the technical field of access control, and particularly relates to a control method for concurrent access of multiple users of a server, comprising the following steps: real-time analysis of a task type and resource requirement characteristics of a user request, generation of a triple including a calculation mode identifier, a data topology fingerprint and a time constraint factor; spatiotemporal coupling matching of the triple with a server node, calculation of a spatiotemporal weight score, generation of a binding execution unit based on the spatiotemporal weight score; quantumized fragmentation scheduling of the binding execution unit, insertion of a reverse dependence barrier between adjacent atomic operation sequences, generation of execution trajectory data during execution; adjustment of calculation mode matching through an execution trajectory feedback mechanism, and calculation mode identifier remapping of unexecuted tasks using path efficiency of completed tasks. The present application effectively reduces calculation resource adaptation deviation of tasks, reduces remote data access overhead, and improves task throughput and execution stability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of access control, in particular to a control method for concurrent access of multiple users of a server. BACKGROUND

[0002] In modern server cluster environments, multi-user concurrent access control is a core problem of task scheduling and computing resource management. With the increasing growth of high-performance computing, cloud computing and artificial intelligence training tasks, server resource allocation faces challenges such as computing-intensive, storage-intensive and other task demands. Traditional task scheduling methods mostly use static strategies, such as fixed resource allocation based on task type or simple scheduling rules based on load balancing. These methods have many shortcomings in computing resource adaptability, data localization access optimization and task execution stability, and are difficult to meet the needs of complex concurrent computing environments.

[0003] Inaccurate identification and matching of task computing patterns, existing task scheduling usually only matches statically according to the computing resource requirements declared in the user request, without in-depth analysis of the actual computing patterns of the task, such as vectorized computing (SIMD), GPU parallel computing, storage access patterns and other features, resulting in low resource utilization, especially in SIMD-intensive tasks, GPU computing tasks and large-scale memory operation tasks. The deviation in computing pattern recognition may lead to a decrease in the computing efficiency of task scheduling.

[0004] Secondly, the matching of data access topology and computing tasks is insufficient. Data in server clusters is usually distributed in different storage levels, such as local SSD, NVMe, cache, distributed storage (such as HDFS), etc. Traditional scheduling methods do not fully consider the data access path of the task, resulting in the task being scheduled to a node with mismatched storage topology, thereby increasing the remote data access delay, reducing the task throughput, and even causing data transmission bottlenecks, affecting the overall computing performance.

[0005] In addition, the scheduling strategy of task splitting and synchronization control is relatively rough. Existing parallel task scheduling usually uses fixed time slices or task blocks for splitting, without dynamic adjustment according to the current load state of the server, resulting in too large task granularity in high load situations, blocking computing resources; in low load situations, the task granularity is too small, increasing the scheduling overhead and reducing the throughput. At the same time, the traditional task synchronization method usually inserts barriers based on static dependency rules, without intelligent insertion according to the resource conflict probability, which may lead to excessive synchronization or resource competition, affecting the task execution efficiency. SUMMARY

[0006] The present application provides a control method for concurrent access of multiple users of a server.

[0007] A control method for multi-user concurrent access of a server, comprising the following steps:

[0008] S1: Real-time analysis of the task type and resource requirement characteristics of a user request, generating a triple including a computing mode identifier, a data topology fingerprint, and a time constraint factor;

[0009] S2: Based on the real-time resource topology state of the server cluster, spatiotemporal coupling matching of the triple and the server nodes, calculation of a spatiotemporal weight score, generation of a binding execution unit based on the spatiotemporal weight score, the binding execution unit including a combination of server nodes that meet the screening constraint conditions, the screening constraint conditions including (filtering candidate nodes in the following priority order):

[0010] Computing mode matching layer: matching of the computing mode identifier and the server characteristics;

[0011] Data topology adaptation layer: matching of the storage hierarchy access path coincidence degree;

[0012] Time window verification layer: the resource reservation time slice corresponding to the time constraint factor covers the task expected execution window;

[0013] S3: Implementation of quantumized fragmentation scheduling on the binding execution unit, splitting each task into multiple atomic operation sequences, and inserting a reverse dependency barrier between adjacent atomic operation sequences according to the dynamic load rate of the server nodes, and generating an execution trajectory data during execution;

[0014] S4: Adjustment of the computing mode matching through the execution trajectory backfeeding mechanism, and remapping of the computing mode identifier of the unexecuted task using the path efficiency of the completed task.

[0015] Optionally, the S1 specifically comprises:

[0016] S11: Identification of the computing mode identifier M based on the score taking maximum method;

[0017] M = argmax (S cpu , S gpu , S mem ), wherein M is the computing mode identifier of the task, used to represent the main computing characteristics of the task in server resource allocation, S cpu is the CPU computing intensity score, S gpu is the GPU computing intensity score, and S mem is the memory intensive score;

[0018] S12: Extraction of the storage hierarchy distribution and data access path of the task associated data, construction of the data topology fingerprint, and measurement of the data topology fingerprint through the storage hierarchy access path coincidence degree C:

[0019] wherein, is the storage location identifier of the i-th data block of the task, is the local storage identifier of the j-th server node, δ is the location matching function, and n represents the number of data blocks accessed by the task;

[0020] S13: dynamically calculating a time constraint factor, which is based on a dynamic urgency:

[0021] parsing the absolute deadline T of the task declaration deadline , and calculating the dynamic urgency U(t):

[0022] wherein, T resrve is the benchmark reserved time window (taking the average value) according to the historical execution time statistics of the same type of task, t current is the current time, i.e., the server system time at the time of task parsing, and α(t) is a dynamic adjustment coefficient based on the current resource competition state.

[0023] Optionally, in the computing mode matching layer:

[0024] if the computing mode identifier is GPU-intensive, the server nodes with CUDA core number ≥ task demand and PCIE bandwidth ≥ 16 GB / s are screened;

[0025] if it is CPU-intensive, the server nodes supporting AVX-512 instruction set and L3 cache capacity ≥ 1.2 times of the task data amount are screened;

[0026] if it is memory-intensive, the server nodes with NUMA node memory bandwidth utilization rate < 60% are screened;

[0027] In the data topology adaptation layer, based on the storage level access path coincidence degree C, only the server nodes with C ≥ 80% are retained to reduce the remote data access overhead.

[0028] Optionally, in the time window verification layer, the task expected execution window T required is calculated based on the estimated execution time T exec est in combination with the dynamic urgency U(t), and it is checked whether there is a continuous resource reserved time slice [t start , t end ] in the node resource reservation table that meets t end -t start ≥ T required .

[0029] Optionally, the S2 comprises calculating the space-time weight score of the server nodes screened:

[0030] Wherein, Score is the space-time weight score of the current computing node, a is the data topology weight, g is the time adaptation weight, l is the load balancing weight, C is the storage level access path coincidence degree, T reserve is the benchmark reserved time window, i.e., the length of the current node's reservable time slice, T required is the task expected execution window, Q pending is the current node's current pending task queue length, Q max is the maximum task queue capacity of the current node;

[0031] The node with the highest space-time weight score is selected to generate a binding execution unit.

[0032] Optionally, the binding execution unit comprises:

[0033] (a) Hardware resource lock: binding exclusive resource combination identifier, comprising:

[0034] GPU resource (GPU card serial number, such as GPU-1, GPU-3).

[0035] CPU resource (CPU physical core ID range, such as CPUcore4-7).

[0036] Storage resource (NVMe storage ID, such as NVMe-2).

[0037] (b) Space-time contract: time and space allocation protocol between task and server node, comprising:

[0038] Maximum execution window allowed by the task: i.e., the time constraint for the task to complete at the latest;

[0039] Data access path mapping: including whether to allow data migration, whether to adopt local cache acceleration strategy.

[0040] Optionally, the S3 specifically comprises:

[0041] S31, defining the quantumization splitting rule of the atomic operation sequence:

[0042] S311, each atomic operation satisfies the minimum resource exclusivity;

[0043] S312, the splitting granularity G is adaptively adjusted according to the node dynamic load rate h;

[0044] S32, constructing an insertion strategy of a reverse dependency barrier: predicting resource conflict risks between adjacent atomic operation sequences, if the time at which a current sequence operation releases a certain resource is later than the time at which a subsequent operation applies the same resource, the resource is marked as having a conflict, the probability of the conflict is calculated according to historical conflict data and current load conditions, when the conflict probability exceeds a set threshold, a reverse dependency barrier is inserted between the previous operation and the subsequent operation, and it is ensured that the subsequent operation can be executed within a protection time interval after the previous operation releases the resource.

[0045] Optionally, the execution trajectory data is generated by recording the spatiotemporal distribution information of the atomic operation sequence in task execution, including the unique identifier of the operation, the occupied server resource, the start time and the end time of the execution, for the tasks having a dependency relationship, the number of the front dependency barrier is recorded, and the post conflict is predicted.

[0046] Optionally, the S4 includes storage level access path benefit evaluation on the execution trajectory data of the completed task, and a storage path benefit value E path ;

[0047] The storage path benefit values of continuous task execution paths of the same user are monitored, and when the storage path benefit values of three continuous tasks satisfy: E path <0.6, the computing mode identifier of the task of the user is forced to be switched to the memory-intensive type.

[0048] Optionally, the storage path benefit value is calculated as:

[0049] Wherein, D local represents the local storage access times, D remote represents the cross-node access times, and τ avg represents the average access delay.

[0050] Advantages of the present application:

[0051] In the present application, the computing mode identifier resolution technology is used to identify the computing type of the task, the storage level access path coincidence degree is constructed in combination with the data topology fingerprint, the task is ensured to be distributed to the node with high data localization rate as much as possible, meanwhile, the time window of task distribution is ensured to cover the expected execution time in combination with the dynamic time constraint factor, and the historical timeout rate is adaptively adjusted, compared with the traditional static scheduling mode, the computing resource adaptation deviation of the task is effectively reduced, the remote data access overhead is reduced, and the task throughput and execution stability are improved.

[0052] The present application introduces a dynamic load adaptive quantization task splitting strategy, so that the task splitting granularity is adjusted according to the server load, and a finer granularity is adopted in a high-load environment to improve the flexibility of resource scheduling and avoid blocking of global computing resources by large tasks. In addition, a probabilistic reverse dependency barrier insertion strategy is adopted, and a synchronization barrier is inserted only when the resource conflict probability is high, thereby avoiding the problem of reduced computing throughput caused by excessive insertion of barriers in the traditional scheduling mode. Compared with the fixed time slice scheduling mode, the flexibility of task scheduling can be improved, the computing concurrency can be improved in a high-load environment, and the resource conflict rate can be reduced.

[0053] The present application performs automatic analysis of trajectory data, calculates the storage level access path benefit value, evaluates the storage access mode of the task, and when the path benefit value of a plurality of continuous tasks is low, the system can automatically adjust the computing mode identifier, so that the task is more inclined to be scheduled to a server with high bandwidth memory or a more optimal data access path, thereby reducing the overhead of cross-node data access. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only illustrate the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0055] Fig. 1 The control method flowchart of the embodiment of the present application is shown in the figure.

[0056] Fig. 2 The screening constraint condition diagram of the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0057] The present application will be described in detail below in combination with the drawings and specific embodiments. It should be noted that in order to make the embodiments more detailed, the following embodiments are the best, preferred embodiments, and other alternative ways can also be used by those skilled in the art to implement some known technologies; and the drawings are only used to more specifically describe the embodiments, and are not intended to specifically limit the present application.

[0058] As shown in the figure, the control method of server multi-user concurrent access includes the following steps: Figs. 1-2

[0059] S1: Real-time analysis of the task type and resource demand characteristics of user request, generating a three-tuple including a computing mode identifier, a data topology fingerprint, and a time constraint factor;

[0060] ​S2: Based on the real-time resource topology state of the server cluster, the triple is matched with the server node in space-time coupling, the space-time weight score is calculated, the binding execution unit is generated based on the space-time weight score, and the binding execution unit includes the combination of server nodes that meet the filtering constraint conditions, and the filtering constraint conditions include (filtering candidate nodes in the following priority order):

[0061] The calculation mode matching layer: the calculation mode identifier is matched with the server characteristics;

[0062] The data topology adaptation layer: match the storage level access path coincidence degree;

[0063] The time window verification layer: the resource reservation time slice corresponding to the time constraint factor covers the task expected execution window;

[0064] The filtering constraint condition is used to filter the candidate nodes to determine the candidate server nodes;

[0065] The calculation space-time weight score is used to get the score to select the optimal server node;

[0066] S3: Implement quantumized fragmentation scheduling on the binding execution unit, split each task into multiple atomic operation sequences, and according to the dynamic load rate of the server node, insert a reverse dependency barrier between adjacent atomic operation sequences, and generate an execution trajectory data during execution;

[0067] S4: Adjust the calculation mode matching through the execution trajectory backfeeding mechanism, and use the path efficiency of the completed task to remap the calculation mode identifier of the unexecuted task.

[0068] S1 specifically includes:

[0069] S11: Based on the maximum method, identify the calculation mode identifier M;

[0070] M = argmax(S cpu ,S gpu ,S mem ), wherein M is the calculation mode identifier of the task, used to represent the main calculation characteristics of the task in server resource allocation, S cpu is the CPU computing intensity score, S gpu is the GPU computing intensity score, and S mem is the memory intensive score;

[0071] CPU computing intensity score: w cpu is the CPU weight coefficient, taking 1.0;

[0072] GPU computing intensity score: w gpu is the GPU weight coefficient, taking 1.5;

[0073] Memory-intensive score: w mem is the memory weight coefficient, taking 1.2;

[0074] If S cpu is the maximum, then M=S cpu , which is identified as CPU-intensive;

[0075] If S gpu is the maximum, then M=S gpu , which is identified as GPU-intensive;

[0076] If S mem is the maximum, then M=S mem , which is identified as memory-intensive;

[0077] S12: Extract the storage level distribution and data access path of the task associated data, construct the data topology fingerprint, and measure the data topology fingerprint through the storage level access path coincidence degree C:

[0078] Wherein, is the storage location identifier of the i-th data block of the task, is the local storage identifier of server node j, δ is the position matching function, and n represents the number of data blocks accessed by the task, is defined as:

[0079]

[0080] S13: Dynamically calculate the time constraint factor, which is based on the dynamic urgency:

[0081] Parse the absolute deadline T deadline of the task declaration, and calculate the dynamic urgency U(t):

[0082] Wherein, T resrve is the benchmark reserved time window (taking the average value) according to the historical execution time statistics of similar tasks, t current is the current time, i.e. the server system time when the task is parsed, and α(t) is the dynamic adjustment coefficient based on the current resource competition state, the value range is 0.5≤□≤2.0.

[0083] In the calculation mode matching layer:

[0084] If the calculation mode identifier is GPU-intensive, filter server nodes with CUDA core number ≥ task demand and PCIe bandwidth ≥ 16 GB / s. The core advantage of GPU calculation is large-scale parallel computing. If the CUDA core number is insufficient, the calculation will be limited, affecting the throughput. Data transmission is required between CPU and GPU. If the PCIe bandwidth is too low, data transfer will become a bottleneck, resulting in insufficient utilization of GPU resources;

[0085] If it is CPU-intensive, filter server nodes that support AVX-512 instruction set and L3 cache capacity ≥ 1.2 times the task data volume. CPU calculation (such as matrix operation, numerical simulation, etc.) relies on SIMD (Single Instruction Multiple Data). AVX-512 is the strongest instruction set for current CPU vectorization, which can significantly improve computing efficiency. CPU calculation is very sensitive to cache hit rate. Task data should be placed in L3 cache as much as possible to reduce memory access. If the data volume far exceeds the L3 cache, the CPU needs to frequently access the main memory, resulting in a decrease in computing performance;

[0086] If it is memory-intensive, filter server nodes with NUMA node memory bandwidth utilization < 60%. Under NUMA (Non-Uniform Memory Access Architecture), cross- NUMA access will cause memory delay to increase, reducing task throughput. Therefore, prefer NUMA nodes with non-overloaded memory bandwidth to ensure efficient use of local memory bandwidth by the task;

[0087] In the data topology adaptation layer, based on the storage level access path coincidence degree C, only server nodes with C ≥ 80% are retained to reduce remote data access overhead.

[0088] In the time window verification layer, the task expected execution window T required Based on the estimated execution time T exec ext Combined with dynamic urgency U(t), it is represented as:

[0089] T required = T exec est × max(1, U(t)) × (1 + β), where β is the fault tolerance coefficient (default 0.2). Check if there is a continuous resource reservation time slice [t start , t end ] in the node resource reservation table that meets t end -t start ≥ T required .

[0090] S2 includes calculating the space-time weight score of the filtered server nodes:

[0091] Wherein, Score is the spatiotemporal weight score of the current computing node, α is the data topology weight (α = 0.6), which measures the matching degree between data and node storage to ensure that data is stored locally as much as possible, γ is the time adaptation weight (γ = 0.3), which measures whether the available time window is sufficient to prevent task interruption, λ is the load balancing weight (λ = 0.1), which measures the current server load to prevent tasks from being concentrated on a few high-load nodes, C is the storage tier access path overlap, and T... reserve The baseline reserved time window, i.e., the length of the reservable time slice for the current node, T required Q is the expected execution window for the task. pending Q represents the current length of the pending task queue for the current node. max This represents the maximum task queue capacity of the current node.

[0092] The node with the highest spatiotemporal weight score is selected to generate the bound execution unit.

[0093] The bound execution unit includes:

[0094] (a) Hardware resource lock: Binds an exclusive resource combination identifier, including:

[0095] GPU resources (GPU card serial number, such as GPU-1, GPU-3).

[0096] CPU resources (CPU physical core ID range, such as CPUcore4-7).

[0097] Storage resources (NVMe storage ID, such as NVMe-2).

[0098] (b) Spatiotemporal Contract: A time and space allocation protocol between the task and the server node, including:

[0099] Maximum execution window allowed for the task: i.e., the latest time constraint for the task to complete (based on the time window to verify the results of the layer);

[0100] Data access path mapping: including whether data migration is allowed and whether a local caching acceleration strategy is adopted (based on the data topology adaptation layer, server nodes with data access path overlap C≥80% to ensure data is localized as much as possible).

[0101] S3 specifically includes:

[0102] S31, defines the rules for quantization of atomic operation sequences:

[0103] S311, each atomic operation satisfies minimum resource exclusivity:

[0104] It occupies the integer / floating-point arithmetic unit of a single physical core;

[0105] Exclusive PCIe channel or exclusive contiguous memory block

[0106] S312, the split granularity G is adaptively adjusted according to the node dynamic load rate η:

[0107] Wherein, η is the current CPU / GPU / memory comprehensive load rate of the node;

[0108] S32, the insertion strategy of the reverse dependence barrier is constructed: the resource conflict risk between adjacent atomic operation sequences is predicted, and the time when the current sequence operation releases a certain resource is later than the time when the subsequent operation applies the same resource. Then, the resource conflict is marked, the conflict probability is calculated according to the historical conflict data and the current load condition, and when the conflict probability exceeds the set threshold, the reverse dependence barrier is inserted between the previous operation and the subsequent operation, so that the subsequent operation can be executed within the protection time interval after the previous operation releases the resource, so as to avoid task exception or execution efficiency caused by resource preemption.

[0109] S321, the resource conflict risk of adjacent atomic operation sequences is predicted:

[0110] If the time t i when the previous operation O k releases the resource R release is later than the time t acquire when the subsequent operation O j applies R k , it is marked as conflict.

[0111] The conflict probability is calculated:

[0112] S322, when P c >0.8, the barrier is inserted between O i and O j , and the following is forced to be satisfied: Where, Δt guard is the protection interval, and the default value is 10 μs.

[0113] The execution trajectory data is generated by recording the space-time distribution information of the atomic operation sequence in the task execution, including the unique identifier of the operation, the occupied server resource, the starting time and the ending time of the execution. For the task with dependence relationship, the number of front dependence barrier is recorded, and the post conflict is predicted.

[0114] Operation ID: the unique number of the current atomic operation sequence;

[0115] Resource lock: record the server resources exclusively used during task execution;

[0116] Start time: record the actual execution starting point of the atomic operation;

[0117] End time: record the actual end point of the atomic operation execution;

[0118] Dependent barrier: Preceding barrier ID: if the operation must wait for a certain barrier to be released before execution, record the ID of the barrier; Post-collision prediction: predict the operation that may occur resource conflict subsequently.

[0119] S4 includes storage level access path benefit evaluation of the execution trajectory data of the completed task, calculating the storage path benefit value E path

[0120] The storage path benefit of the continuous task execution path of the same user is monitored, and when the storage path benefit values of three continuous tasks meet: E path <0.6, the computing mode identifier of the user task is forced to switch to memory intensive, and the resource matching strategy of the task is adjusted to preferentially schedule to the server node with higher memory bandwidth and lower data access delay.

[0121] The storage path benefit value is calculated as:

[0122] Wherein, D local represents the local storage access times, D remote represents the cross-node access times, and τ avg represents the average access delay.

[0123] The present application covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present application. In order to make the public have a thorough understanding of the present application, specific details are described in the following preferred embodiments of the present application, and the present application can also be fully understood without the description of these details to those skilled in the art. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits, etc. are not described in detail.

[0124] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.​

Claims

1. A method for controlling multiple user concurrent access to a server, characterized in that, The method comprises the following steps: S1: Real-time analysis of user request task type and resource demand characteristics, generating a triple including calculation mode identifier, data topology fingerprint, time constraint factor; S2: Based on the real-time resource topology state of the server cluster, the triple is matched with the server node in space-time coupling, the space-time weight score is calculated, and the binding execution unit is generated based on the space-time weight score, the binding execution unit includes the combination of server nodes that meet the screening constraint conditions, and the screening constraint conditions include: Calculation mode matching layer: calculation mode identifier matches server characteristics; Data topology adaptation layer: match storage level access path coincidence degree; Time window verification layer: the resource reservation time slice corresponding to the time constraint factor covers the task expected execution window; S3: Implement quantumized fragmentation scheduling on the binding execution unit, split each task into multiple atomic operation sequences, and insert reverse dependency barriers between adjacent atomic operation sequences according to the dynamic load rate of the server node, and generate execution track data during execution; S4: Adjust the calculation mode matching through the execution track back-feeding mechanism, and remap the calculation mode identifier of the unexecuted task using the path efficiency of the completed task.

2. The control method of server multi-user concurrent access according to claim 1, characterized in that, The S1 specifically comprises: S11: Based on the score taking maximum method, identify the calculation mode identifier M; M = arg max(S cpu ,S gpu ,S mem ), where M is a compute mode identifier for the task that indicates the primary compute characteristics of the task in server resource allocation, S cpu is a CPU compute intensity score, S gpu is a GPU compute intensity score, and S mem is a memory intensive score. S12: Extract the storage level distribution and data access path of the task associated data, construct the data topology fingerprint, and measure the data topology fingerprint through the storage level access path coincidence degree C: wherein, is the storage location identifier of the i-th data block of the task, is the local storage identifier of the server node j, δ is the location matching function, and n represents the number of data blocks accessed by the task; S13: Dynamically calculate the time constraint factor, which is based on dynamic urgency: parsing an absolute deadline T declared for the task deadline calculating a dynamic urgency U(t): where T reserve is the baseline reserved time window according to the historical execution time statistics of the same task, t current is the current time, and a(t) is a dynamic adjustment coefficient based on the current resource contention state.

3. The method of claim 2, wherein the server multi-user concurrent access control method is characterized by, In the calculation mode matching layer: If the calculation mode identifier is GPU intensive, select the server node with CUDA core number ≥ task demand and PCIE bandwidth ≥ 16 GB / s; If it is CPU intensive, select the server node that supports AVX-512 instruction set and L3 cache capacity ≥ 1.2 times of task data volume; If it is memory intensive, select the server node with NUMA node memory bandwidth utilization rate < 60%; In the data topology adaptation layer, based on the storage level access path coincidence degree C, only the server nodes with C ≥ 80% are retained.

4. The method of claim 3, wherein, In the time window verification layer, the task expected execution window T required Based on the estimated execution time T exec est Combined with the dynamic urgency U(t), it is found out whether there is a continuous resource reservation time slice [t start ,t end ] in the node resource reservation table that satisfies t end -t start ≥T required server node.

5. The method of claim 1, wherein the server multi-user concurrent access control method is characterized by, The S2 includes calculating the space-time weight score of the server nodes selected through screening: wherein Score is the spatio-temporal weight score of the current computing node, a is the data topology weight, g is the time adaptation weight, l is the load balancing weight, C is the storage hierarchy access path overlap degree, T reserve is the baseline reserved time window, T required is the task expected execution window, Q pending is the current node's current pending task queue length, Q max is the current node's maximum task queue capacity; Select the node with the highest space-time weight score to generate the binding execution unit.

6. The method of claim 5, wherein, The binding execution unit includes: (a) Hardware resource lock: binding exclusive resource combination identifier, including: GPU resource; CPU resource; storage resource; (b) Space-time contract: time and space allocation protocol between task and server node, including: Maximum execution window allowed by the task: the time constraint of the latest completion of the task; Data access path mapping: including whether to allow data migration, whether to adopt local cache acceleration strategy.

7. The method of claim 1, wherein the server multi-user concurrent access control method is characterized by, The S3 specifically comprises: S31, define the quantumized splitting rule of atomic operation sequence: S311: Each atomic operation meets the minimum resource exclusivity; S312: The splitting granularity is adaptively adjusted according to the dynamic load rate of the node; S32, constructing an insertion strategy of the reverse dependence barrier: predicting a resource conflict risk between adjacent atomic operation sequences, if a time at which a current sequence operation releases a certain resource is later than a time at which a subsequent operation applies the same resource, the resource is marked as having a conflict, a conflict probability is calculated according to historical conflict data and a current load condition, and when the conflict probability exceeds a set threshold, a reverse dependence barrier is inserted between the previous operation and the subsequent operation, so as to ensure that the subsequent operation can be executed within a protection time interval after the previous operation releases the resource.

8. The method of claim 1, wherein the server multi-user concurrent access control method is characterized by, The execution trajectory data is generated by recording space-time distribution information of the atomic operation sequence in task execution, including a unique identifier of the operation, occupied server resources, a starting time and an ending time of execution, for tasks having a dependence relationship, a number of a preceding dependence barrier is recorded, and post-conflict is predicted.

9. The method of claim 1, wherein the server multi-user concurrent access control method is characterized by, The S4 includes a storage level access path benefit evaluation of the execution trajectory data of the completed task, and a storage path benefit value E is calculated path ; The continuous task execution path benefit of the same user is monitored, and when the storage path benefit values of 3 continuous tasks satisfy: E path <0.6, the computing mode identifier of the task of the user is forced to be switched to memory-intensive.

10. The method of claim 1, wherein, The storage path benefit value is calculated as: where D local denotes the local storage access times, D remote denotes the cross-node access times, τ avg denotes the average access latency.

Images