Heterogeneous computing power scheduling method and system

Abstract

The embodiment of the invention provides a heterogeneous computing power scheduling method and system, and relates to the field of computing power application. The system comprises a scheduling control center and a plurality of computing nodes. And the scheduling control center searches a resource library based on the task tag of the to-be-processed task so as to find a matched target resource tag. The resource label is generated based on the hardware attribute information of the computing node, the task label represents the hardware requirement of the to-be-processed task for the computing resource, and the computing node suitable for scheduling the to-be-processed task can be accurately found by comparing the task label with the resource label. And finding a target computing node by using the hash value of the unique identification computing node, and scheduling the to-be-processed task by using the target computing node. By adopting a resource unique identification mechanism based on a hash value and a scheduling strategy based on matching of a task label and a resource label, the accuracy and stability of heterogeneous computing power resource scheduling can be effectively improved, and the problem of task scheduling failure caused by resource identification conflicts or label configuration errors is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computing power application, in particular to a heterogeneous computing power scheduling method and system. BACKGROUND

[0002] With the continuous growth of artificial intelligence and high-performance computing demand, heterogeneous computing power platforms have become an important infrastructure to support diversified computing tasks. Currently, heterogeneous computing power resources are mainly distributed in data centers and operator edge grid nodes, covering central processing units (CPU), graphics processing units (GPU), neural network processing units (NPU), field programmable gate arrays (FPGA) and other types of computing devices.

[0003] In the prior art, these heterogeneous resources are usually accessed and scheduled by a static configuration method based on a label and plug-in mechanism. However, the labels of heterogeneous resources mainly rely on manual configuration, lack of standardized verification mechanism, leading to difficulty in matching resource attributes and deviation in scheduling decisions, especially in a heterogeneous computing power cluster. At the same time, the identification of each heterogeneous resource is not unified, making it difficult to support fast retrieval and scheduling after multi-source computing power fusion. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a heterogeneous computing power scheduling method and system that can improve the accuracy and stability of heterogeneous computing power resource scheduling.

[0005] To achieve the above purpose, the technical solutions adopted by the embodiments of the present application are as follows: In a first aspect, the present application provides a heterogeneous computing power scheduling method applied to a heterogeneous computing power scheduling system, the heterogeneous computing power scheduling system comprising a scheduling control center and a plurality of computing nodes, the method comprising: The scheduling control center generates a task label according to the hardware requirements of a to-be-processed task; The scheduling control center finds a target resource label matching the task label by searching a resource library according to the task label; the resource library records the correspondence between resource labels and hash values of each computing node in the heterogeneous computing power scheduling system, the resource label is generated based on the hardware attribute information of the computing node, and the hash value is used to uniquely identify the computing node; The scheduling control center determines a target computing node based on the target hash value corresponding to the target resource label, and schedules the to-be-processed task by using the target computing node.

[0006] In an optional implementation, the dispatching control center looks up a resource library according to the task label, and finds a target resource label matching the task label, including: The dispatching control center screens at least one candidate resource label matching the task label from the resource library; If there is one candidate resource label, the dispatching control center determines the candidate resource label as the target resource label corresponding to the task label; If there are multiple candidate resource labels, the dispatching control center weights and scores each candidate resource label based on resource adaptability, resource availability, resource reliability, and resource use efficiency, and determines the candidate resource label with the highest score as the target resource label corresponding to the task label.

[0007] In an optional implementation, the method further includes: When the computing node newly joins the heterogeneous computing power dispatching system or resource information of the computing node is changed, the computing node acquires hardware attribute information, network configuration information, and running information; The computing node acquires a first field set from the hardware attribute information, the network configuration information, and the running information, and generates a hash value of the computing node based on the first field set; The computing node acquires a second field set from the hardware attribute information according to a label field rule corresponding to a resource type, and generates a resource label of the computing node based on the second field set; The computing node sends the hash value and the resource label of the node to the dispatching control center, so that the dispatching control center adds the hash value and the resource label of the computing node to the resource library.

[0008] In an optional implementation, the computing node acquires a first field set from the hardware attribute information, the network configuration information, and the running information, and generates a hash value of the computing node based on the first field set, including: The computing node acquires a first field set from the hardware attribute information, the network configuration information, and the running information in a sequence from high to low of field weights and according to a minimum field number principle; the field weight is positively correlated with stability of a field; The computing node performs hash calculation based on the first field set by using a preset hash algorithm to obtain a temporary hash value, and judges whether the temporary hash value has a conflict; If the temporary hash value has no conflict, the computing node determines the temporary hash value as the hash value of the computing node; If the temporary hash value has a conflict, the computing node selects a field with a high weight from the remaining fields of the hardware attribute information, the network configuration information, and the running information, adds the field to the first field set, and generates a corresponding temporary hash value based on the updated first field set until the temporary hash value has no conflict, obtaining the hash value of the computing node; the remaining field is different from the fields in the first field set.

[0009] In an optional implementation, the determining whether the temporary hash value has a conflict includes: The computing node checks whether the temporary hash value exists in a local hash cache; the local hash cache is periodically updated based on the resource library. If the temporary hash value exists in the local hash cache, the computing node determines that the temporary hash value has a conflict. If the temporary hash value does not exist in the local hash cache, the computing node queries the scheduling control center whether the temporary hash value has a conflict; the scheduling control center determines whether the temporary hash value has a conflict based on whether the temporary hash value exists in the resource library.

[0010] In an optional implementation, the generating the resource tag of the computing node based on the second field set includes: The computing node performs standardization conversion on each second field in the second field set, obtaining a standard field corresponding to each second field. The computing node generates the resource tag of the computing node according to the standard fields.

[0011] In an optional implementation, the scheduling control center determines a target computing node based on a target hash value corresponding to the target resource tag, and schedules the to-be-processed task by using the target computing node, including: The scheduling control center binds the to-be-processed task and the target hash value; The scheduling control center controls a target computing node corresponding to the target hash value to deploy a model; The target computing node executes the to-be-processed task based on the model.

[0012] In an optional implementation, the scheduling control center controls a target computing node corresponding to the target hash value to deploy a model, including: The scheduling control center finds a computing power adaptation plug-in corresponding to the target computing node according to a resource attribute of the target computing node corresponding to the target hash value; The scheduling control center loads the computing power adaptation plug-in corresponding to the target computing node locally, and controls the target computing node to load the corresponding computing power adaptation plug-in; The target computing node deploys a model by using the computing power adaptation plug-in.

[0013] In an optional implementation, the method further includes: The target computing node sends model deployment information, a scheduling track of the to-be-processed task, and an execution result of the to-be-processed task to the scheduling control center, so that the scheduling control center generates a scheduling link log.

[0014] In a second aspect, the present application provides a heterogeneous computing power scheduling system, which includes a scheduling control center and a plurality of computing nodes: The scheduling control center is configured to generate a task label according to a hardware requirement of a to-be-processed task; The scheduling control center is configured to find a target resource label matched with the task label by searching a resource library, wherein the resource library records a correspondence between resource labels and hash values of the computing nodes in the heterogeneous computing power scheduling system, the resource labels are generated based on hardware attribute information of the computing nodes, and the hash values are used to uniquely identify the computing nodes; The scheduling control center is configured to determine a target computing node based on a target hash value corresponding to the target resource label, and schedule the to-be-processed task by using the target computing node.

[0015] Compared with the prior art, the heterogeneous computing power scheduling method and system provided by the embodiments of the present application include a scheduling control center and a plurality of computing nodes. The scheduling control center generates a task label of a to-be-processed task, and finds a target resource label matched with the task label by searching a resource library. The resource label is automatically generated based on hardware attribute information of the computing nodes, and the task label represents a hardware requirement of the to-be-processed task on computing resources. By comparing the task label with the resource label in the resource library, the computing node suitable for scheduling the to-be-processed task can be accurately found. Further, the target computing node is found by using a hash value uniquely identifying the computing node, and the to-be-processed task is scheduled by using the target computing node.

[0016] By using the resource unique identification mechanism based on the hash value and the scheduling strategy based on matching of the task label and the resource label, the embodiments of the present application can effectively improve the accuracy and stability of heterogeneous computing power resource scheduling, and reduce the problem of task scheduling failure caused by resource identification conflict or label configuration error.

[0017] In order to make the above objectives, characteristics and advantages of the present application more apparent, more comprehensible, the following preferred embodiments are specifically described in detail below, together with the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments, and it should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be considered as a limitation to the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0019] Figure 1 A flowchart of the heterogeneous computing power scheduling method provided by the embodiments of the present application is shown.

[0020] Figure 2 Another flowchart of the heterogeneous computing power scheduling method provided by the embodiments of the present application is shown.

[0021] Figure 3 A block diagram of the heterogeneous computing power scheduling system provided by the embodiments of the present application is shown.

[0022] Figure 4 Another block diagram of the heterogeneous computing power scheduling system provided by the embodiments of the present application is shown. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application, and obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0024] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of the present application.

[0025] It should be noted that the relational terms such as first and second and the like are used only to differentiate one entity or action from another, and do not necessarily require or imply that these entities or actions occur in any particular sequence or order. Also, the terms "comprising," "including," or any other variation thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, an element defined by an expression "comprising a... " does not exclude the existence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0026] With the rapid growth of high-performance computing demands such as artificial intelligence, big data, and edge computing, heterogeneous computing platforms have become an important infrastructure to support complex computing tasks. Currently, the access and scheduling of heterogeneous computing resources usually rely on static configuration management methods based on labels and plug-in mechanisms.

[0027] Container orchestration systems represented by Kubernetes achieve resource classification and scheduling decisions by adding labels to nodes. However, this label system is usually configured manually and lacks a unified standard and automated consistency checking mechanism, resulting in potential conflicts, duplication, or omission of label information. In addition, various accelerator resources usually need to be managed through independent plug-ins, and plug-ins from different manufacturers differ in resource description format, interface definition, and metadata structure, further exacerbating the fragmentation problem of resource identification and management.

[0028] The inventors have found that the prior art solutions have the following key technical problems in practical application: First, resource identification is not unified: the current unique identification of resources usually relies on the combination of IP addresses, hostnames, labels, and other fields, lacking a stable, unique registration mechanism and global namespace, making it difficult to meet the rapid retrieval and efficient scheduling needs in the multi-source heterogeneous computing fusion scenario.

[0029] Second, resource attribute matching is difficult: in the task scheduling process, resource matching mainly relies on the consistency of label information. Due to the lack of standardized management and automatic checking mechanism for labels, in large-scale heterogeneous computing cluster environments, label misplacement and semantic ambiguity may occur, affecting the accuracy and execution efficiency of scheduling decisions.

[0030] Third, task scheduling lacks automatic adaptability: the label mechanism lacks dynamic rules and constraint logic and cannot automatically adapt to multi-dimensional hardware attribute information such as model running environment and chip driver version.

[0031] Based on this, the embodiment of the present application provides a heterogeneous computing power scheduling method and system, the heterogeneous computing power scheduling system comprises a scheduling control center and a plurality of computing nodes. The scheduling control center generates a task label of a to-be-processed task, and finds a target resource label matched with the task label based on the task label in a resource library. The resource label is automatically generated based on hardware attribute information of the computing node, and the task label represents hardware requirements of the to-be-processed task on computing resources. By comparing the task label with the resource label in the resource library, the computing node suitable for scheduling the to-be-processed task can be accurately found. Further, the target computing node is found by using a hash value uniquely identifying the computing node, and the to-be-processed task is scheduled by using the target computing node. The embodiment of the present application can effectively improve the accuracy and stability of heterogeneous computing power resource scheduling by using the resource unique identification mechanism based on the hash value and the scheduling strategy based on the matching of the task label and the resource label, and can reduce the problem of task scheduling failure caused by resource identification conflict or label configuration error.

[0032] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0033] Please refer to Figure 1 , Figure 1 A flowchart of a heterogeneous computing power scheduling method provided by the embodiment of the present application is shown. The method is applied to a heterogeneous computing power scheduling system, and the heterogeneous computing power scheduling system comprises a scheduling control center and a plurality of computing nodes. The method comprises the following steps: Step S100, the scheduling control center generates a task label according to hardware requirements of a to-be-processed task.

[0034] In the embodiment of the present application, the scheduling control center receives a task request of a to-be-processed task, and the task request comprises hardware requirement information and model requirement information. The hardware requirement information is used to represent hardware requirements of the to-be-processed task on the computing node, and the model requirement information is used to represent model requirements of the to-be-processed task on the computing node.

[0035] Specifically, the scheduling control center generates a corresponding task label based on hardware requirements (i.e. hardware requirement information) proposed by the to-be-processed task, and the task label is expressed in a standardized key-value pair, which is used to represent specific requirements of the to-be-processed task on underlying computing resources.

[0036] Suppose the user submits a task containing the following information {model: bert; requirement: "chip: GPU; cuda: "> = 11.4"; pytorch: "> = 2.0"}. The scheduling control center extracts the hardware requirements of the task to be processed from the task request, and converts the YAML or JSON format configuration into a Key-Value format task label, such as {"chip": "GPU","cuda": ">=11.4","pytorch": ">=2.0"}.

[0037] It should be noted that the task label format is consistent with the resource label, so as to establish a mapping and reverse index structure. The generation of the task label does not rely on manual configuration, but extracts the constraints of the running environment (i.e. hardware requirements) in the task request, and standardizes the processing according to the unified label structure.

[0038] Step S110, the scheduling control center looks up the resource library according to the task label, and finds the target resource label matched with the task label; the resource library records the correspondence between the resource labels and hash values of each computing node in the heterogeneous computing power scheduling system, the resource label is generated based on the hardware attribute information of the computing node, and the hash value is used to uniquely identify the computing node.

[0039] In the embodiment of the application, the resource library is a data storage structure maintained by the scheduling control center, which records the correspondence between the resource labels and hash values of all computing nodes in the heterogeneous computing power scheduling system. Among them, the computing node is uniquely identified and tracked by the hash value, and the resource label is automatically generated based on the hardware attribute information of each computing node, which covers key dimensions such as chip type, driver version and instruction set architecture. For example, a computing node equipped with NVIDIA A100 GPU and running CUDA 11.8 driver will have a resource label of {"chip": "GPU", "driver": "CUDA11.4","arch": "x86"}.

[0040] Step S120, the scheduling control center determines the target computing node based on the target hash value corresponding to the target resource label, and schedules the task to be processed by using the target computing node.

[0041] In the embodiment of the application, since the unique mapping relationship between the hash value and the computing node has been established in the resource library, the scheduling control center can directly retrieve the node IP or node identifier of the target computing node corresponding to the target hash value, and assign the task to be processed to the target computing node for execution.

[0042] It should be noted that in the heterogeneous computing power application scenario, each user task may be split into multiple to-be-processed tasks according to needs, and a matching target computing node needs to be found for each to-be-processed task. For example, a smart medical image analysis system needs to use a GPU and an NPU for image segmentation. The GPU is used for 3D recognition, and the NPU is used for computed tomography (CT) segmentation.

[0043] To sum up, the heterogeneous computing power scheduling method provided by the embodiment of the application, the scheduling control center generates a task tag of a to-be-processed task, and finds a target resource tag matched with the task tag based on the task tag. The resource tag is automatically generated based on the hardware attribute information of the computing node, and the task tag represents the hardware demand of the to-be-processed task for the computing resource. By comparing the task tag with the resource tag in the resource library, the computing node suitable for scheduling the to-be-processed task can be accurately found. Further, the target computing node is found by using the hash value uniquely identifying the computing node, and the to-be-processed task is scheduled by using the target computing node. The embodiment of the application can effectively improve the accuracy and stability of the heterogeneous computing power resource scheduling by using the resource unique identification mechanism based on the hash value and the scheduling strategy based on the matching of the task tag and the resource tag, and reduce the problem of task scheduling failure caused by resource identification conflict or label configuration error.

[0044] Optionally, the following provides a possible implementation manner for how to find a computing node matched with the task demand. Figure 1 The sub-step of step S110 can include: Step S111, the scheduling control center screens at least one candidate resource tag matched with the task tag from the resource library.

[0045] In the embodiment of the application, the task tag is used to represent a set of standardized key-value pair information of the hardware environment required for task running, and the candidate resource tag refers to a set of all resource node tags meeting the constraint conditions of the task tag. Specifically, the scheduling control center quickly locates the resource tags meeting the constraint conditions of the task by constructing an inverted index structure, and obtains the candidate resource tags matched with the task tag. The task tag and the resource tag are uniform in format, the task tag is generated by the scheduling control center, and the resource tag is generated by the corresponding computing node.

[0046] It should be noted that, considering the compatibility of different versions and models, a compatibility matrix is introduced through environmental semantic association to ensure the accuracy of matching task labels with resource labels. That is, the hardware requirements needed for task scheduling are obtained according to the task label and the compatibility matrix, and the hardware information covered in the hardware requirements may be greater than the task requirements recorded in the task label. For example, the task label records the driver version CUDA greater than or equal to 11.4. The compatibility matrix records that 11.8 is compatible with 11.4, so the resource label recording "driver": "CUDA11.8" will be selected as the candidate resource label of the task label.

[0047] In step S112, if there is a candidate resource label, the scheduling control center determines the candidate resource label as the target resource label corresponding to the task label.

[0048] In the embodiment of the present application, the scheduling control center judges whether the number of the currently screened candidate resource labels is unique. If there is only one candidate resource label, the scheduling control center directly determines it as the target resource label corresponding to the task to be processed. This embodies the design logic that no additional evaluation is needed and immediate decision can be made in the case of unique resource matching result, which helps to improve scheduling efficiency and reduce computational overhead.

[0049] In step S113, if there are multiple candidate resource labels, the scheduling control center weights and scores each candidate resource label based on resource adaptability, resource availability, resource reliability and resource use efficiency, and determines the candidate resource label with the highest score as the target resource label corresponding to the task label.

[0050] In the embodiment of the present application, if there are multiple candidate resource labels, i.e. the system detects that multiple resource nodes can meet the basic running requirements of the task, the scheduling control center needs to comprehensively evaluate the resources corresponding to these candidate resource labels.

[0051] Specifically, the scheduling control center weights and scores each candidate resource based on four key dimensions: resource adaptability (such as whether it completely matches the task label), resource availability (such as the load state of the computing node, the preloading state of the model), resource reliability (such as the historical deployment success probability, the failure rate) and resource use efficiency (such as green energy efficiency, whether the computing node and the user end are in the same network area). Among them, the weight of each dimension can be flexibly configured according to the actual application scene, for example, in the scene with high requirement for task stability, the score weight of resource reliability can be appropriately increased; while in the case of resource shortage, the resource use efficiency can be focused on.

[0052] The scheduling control center selects the candidate resource label with the highest score as the target resource label according to the weighted score result, ensures that the optimal computing node can be selected in the case of multi-resource matching, and improves the intelligent procedure and adaptability of the heterogeneous computing power scheduling system.

[0053] It can be seen that the embodiment of the application can effectively cope with the multi-solution problem in the heterogeneous computing power scheduling scene by introducing the candidate resource screening mechanism and the multi-dimensional weighted scoring mechanism, can realize intelligent priority of the target resource in the case of multiple matching resources, and thus improves the rationality, scientificity and rationality of the heterogeneous computing power scheduling, and enhances the dynamic response capability and resource utilization rate of the heterogeneous computing power scheduling system.

[0054] Optionally, for how to construct and update the resource library, a possible implementation manner is provided as follows. Please refer to Figure 2 The method further includes the following steps: Step S200, when a computing node newly joins the heterogeneous computing power scheduling system or the resource information of the computing node is changed, the computing node acquires hardware attribute information, network configuration information and running information.

[0055] In the embodiment of the application, each computing node detects the resource information periodically or event-driven, and if the field affecting the calculation of the hash value is changed, the computing node whose resource information is changed recalculates the hash value. Or, when a new computing node is added to the heterogeneous computing power scheduling system, the hash value of the newly added computing node needs to be calculated. When the computing node needs to calculate the hash value, the computing node triggers the resource information acquisition process.

[0056] As a possible implementation manner, the computing node can acquire the resource information including the hardware attribute information, the network configuration information and the running information through the following means. Specifically, for example, the hardware attribute information can be acquired through a local system command interface, for example, BIOS / mainboard information is acquired by using a dmidecode interface, and CPU, random access memory (RAM), GPU and other hardware lists are acquired by using a lshw interface.

[0057] The information corresponding to the hardware list can be obtained through the software development kit (SDK) interface of the driver. The network configuration information and the hardware fingerprint (i.e., hardware attribute information), such as a media access control (MAC) address, can also be obtained through a system file path. After the detection is completed, the information obtained by the detection is encapsulated as a structured field, and the availability of the structured field is evaluated, i.e., whether each structured field is empty, whether there is ambiguity, and whether it is unique. The structured field that is not empty, does not have ambiguity, and is unique is determined as the resource information of the computing node, which is divided into hardware attribute information, network configuration information, and running information according to the resource type.

[0058] In step S210, the computing node obtains a first field set from the hardware attribute information, the network configuration information, and the running information, and generates a hash value of the computing node based on the first field set.

[0059] In an embodiment of the present application, the heterogeneous computing power scheduling system has a built-in field priority library, which is divided into strong stability fields (such as MAC address, chip serial number), medium stability fields (such as driver version, instruction set architecture), and weak stability fields (such as running frequency, memory capacity, etc.). According to the hash calculation list supported by the computing node and the field priority library, the first field set is selected from the hardware attribute information, the network configuration information, and the running information, and a hash value for uniquely identifying the computing node is generated based on the first field set. Table 1 is used as an example to exemplarily illustrate the field priority library.

[0060] Table 1

[0061] In step S220, the computing node obtains a second field set from the hardware attribute information according to the label field rule corresponding to the resource type, and generates a resource label of the computing node based on the second field set.

[0062] In an embodiment of the present application, the hardware attribute information of the computing node belongs to a static attribute field, such as chip model, driver version, CPU architecture, memory capacity, CANN version, etc. The resource type is defined as the computing power category to which the computing node belongs, such as CPU, GPU, NPU, etc. The label field rule refers to a field selection template preset for different resource types, such as chip model + driver version + CPU structure for a computing node of the GPU type. A plurality of second fields are selected from the hardware attribute information of the computing node according to the corresponding label field rule, and a second field set is obtained. Finally, a resource label of the computing node is generated based on the second field set.

[0063] Step S230, the computing node sends the hash value and the resource label of the node to the scheduling control center, so that the scheduling control center adds the hash value and the resource label of the computing node to the resource library.

[0064] In the embodiment of the application, the computing node sends the generated hash value, resource label and resource information to the scheduling control center, and the scheduling control center adds the hash value, resource label and resource information of the computing node as a record to the resource library after receiving and verifying the reported information, to obtain the mapping relationship of the hash value, resource label and resource information. The registration mechanism of the resource library ensures that the scheduling control center can quickly match the computing resources based on the resource label during task scheduling, and lock the target node through the hash value, to realize accurate and efficient heterogeneous computing power scheduling.

[0065] It should be noted that when the computing node initiates an update to the scheduling control center, the scheduling control center replaces the old record corresponding to the computing node with the newly received hash value, resource label and resource information, and updates the resource state to be changed or to be verified. In addition, the old hash value before the change needs to be reserved for a preset time for tracing old scheduling records.

[0066] It can be seen that, by introducing the node dynamic perception mechanism, the hash value automatic generation mechanism and the resource label standardized generation mechanism, the embodiment of the application can realize the automatic registration of resource identification (i.e. hash value) and scheduling capability (i.e. resource label) in the scenario of frequent changes of computing nodes or first access to the system, ensuring the consistency of resource identification and the real-time of scheduling data, and effectively improving the adaptive ability, resource management efficiency and scheduling accuracy of the heterogeneous scheduling system.

[0067] Optionally, for how to generate a non-conflicting hash value, a possible implementation is provided as follows. Figure 2 The sub-step of step S210 can include: Step S211, the computing node acquires a first field set from the hardware attribute information, network configuration information and running information according to the order of field weight from high to low and the principle of minimum field number; the field weight is positively correlated with the stability of the field.

[0068] In the embodiment of the application, the field weight is allocated to the strong stability field, the medium stability field and the weak stability field in the field priority library according to the stability, for example, the field weight of the MAC address is set to 10, the field weight of the chip serial number is set to 9, and the field weight of the CPU frequency is set to 3.

[0069] It should be noted that each resource type is pre-configured with a hash template, and the hash template records a full set of fields that can be selected by the corresponding resource type when calculating the hash value. The computing node matches the hash template according to the resource type, and acquires a first field set according to the matched hash template, according to the order of field weight from high to low and the principle of minimum field number, that is, according to the order of field weight from high to low, multiple first fields recorded in the hash template that meet the principle of minimum field number are selected from the hardware attribute information, network configuration information and running information, and a first field set is obtained.

[0070] For example, the first field set of the GPU can be the chip model, the MAC address and the driver version, the first field set of the NPU can be the chip model, the serial number and the Compute Architecture for Neural Networks (CANN) version, and the first field set of the x86 CPU can be the motherboard serial number, the kernel version and the architecture ID.

[0071] Step S212, the computing node performs hash calculation based on the first field set by using a preset hash algorithm, obtains a temporary hash value, and judges whether the temporary hash value exists conflict.

[0072] In the embodiment of the application, the preset hash algorithm can be a hash function of an encryption level such as SHA256 and BLAKE3, the computing node splices and performs irreversible hash operation on the field values in the first field set by using the preset hash algorithm, generates a temporary hash value with a fixed length, and further judges whether the temporary hash value exists conflict in the heterogeneous computing power scheduling system. Wherein, the conflict is defined as: the hash value already exists in the registration record of other computing nodes, which may cause resource identifier duplication and task scheduling error.

[0073] Wherein, the uniqueness of the hash value is determined by the physical uniqueness of the field selection and the time randomness. The time stamp (such as the start time stamp, the first registration time stamp) determines the time stamp as an unpredictable field added to the first field set, so as to enhance the uniqueness and non-repetition of the hash input. For example, two computing nodes of the GPU type with the same resource information, if the hash values are generated at the same time, because the time stamps are different, the generated hash values can still be different.

[0074] Step S213, if the temporary hash value does not exist conflict, the computing node determines the temporary hash value as the hash value of the computing node.

[0075] In the embodiment of the present application, if the result of the judgment is that the temporary hash value does not exist conflict, the computing node directly determines the temporary hash value as the formal hash value of the computing node, and reports to the scheduling control center to complete the registration process. That is, the scheduling control center adds the mapping relationship between the hash value corresponding to the computing node and the resource label to the resource library based on the hash value.

[0076] In step S214, if the temporary hash value exists conflict, the computing node selects a field with high weight from the remaining fields of the hardware attribute information, the network configuration information and the running information, adds the field to the first field set, and generates a corresponding temporary hash value based on the updated first field set until the temporary hash value does not exist conflict, and obtains the hash value of the computing node; the remaining fields do not coincide with the fields in the first field set.

[0077] In the embodiment of the present application, if it is detected that the temporary hash value exists conflict, the computing node needs to select a field with high weight from the remaining fields and add it to the current first field set, thereby forming a new first field set. The updated first field set is subjected to hash calculation and conflict detection again until a hash value without conflict is generated. For example, the first field set is expanded from 3 fields to 4 fields, and the start time stamp is added to generate a unique hash value. If there is repetition, the temporary ID or the physical device slot of the computing node can be added as a strong entropy supplement field.

[0078] It should be noted that when the heterogeneous computing power scheduling system detects that there is a hash conflict in the system, the computing node with the hash conflict will trigger the re-computation of the hash value, and update the resource library based on the newly computed non-conflict hash value, so as to ensure the consistency and traceability of the hash value representing the resource state.

[0079] It can be seen that, by introducing the field weight ordering, the minimum field number constraint and the dynamic field expansion mechanism, the embodiment of the present application can guarantee the uniqueness of the hash value (i.e. the maximum distinguishability) while minimizing the number of fields and the calculation overhead, thereby improving the accuracy and reliability of resource registration in the heterogeneous computing power scheduling system.

[0080] Optionally, as to how to judge whether the hash value exists conflict, the following provides a possible implementation manner. The sub-step of judging whether the temporary hash value exists conflict in step S212 can include: Firstly, the computing node checks whether the temporary hash value exists in the local hash cache; the local hash cache is periodically updated based on the resource library.

[0081] In the embodiment of the present application, the computing node locally checks whether the current generated temporary hash value already exists in a data structure named "local hash cache". The local hash cache is defined as a lightweight and periodically updated set of hash values, the content of which is derived from the resource library maintained by the scheduling control center and used to provide fast local conflict detection capability. Specifically, the computing node periodically synchronously updates the local hash cache from the resource library to ensure that the hash value information contained in the local hash cache is basically consistent with the global state, but there may be a certain time delay. However, the time delay does not affect the uniqueness of the hash value.

[0082] Secondly, if the temporary hash value exists in the local hash cache, the computing node determines that the temporary hash value exists conflict.

[0083] In the embodiment of the present application, if the same record as the current temporary hash value exists in the local hash cache, the computing node determines that the hash value exists conflict, and accordingly triggers the field expansion process, that is, more field combinations are selected and hash calculation is performed again until a new hash value without conflict is generated. The determination logic is based on the premise that even if the local cache does not completely reflect the latest state of the resource library, as long as there is a duplicate in the local cache, it is very likely that the hash value has been registered and used by other nodes in the system, so it is necessary to avoid repeated generation.

[0084] Finally, if the temporary hash value does not exist in the local hash cache, the computing node queries the scheduling control center whether the temporary hash value exists conflict; the scheduling control center determines whether the temporary hash value exists conflict based on whether the temporary hash value exists in the resource library.

[0085] In the embodiment of the present application, if the temporary hash value is not found in the local hash cache, the computing node will send a query request to the scheduling control center to further confirm whether the hash value has been registered in the resource library. After receiving the query request, the scheduling control center determines whether the temporary hash value already exists based on the existing data in the resource library, and feeds back the determination result to the computing node.

[0086] It can be seen that, by introducing the two-level conflict detection mechanism combining the preliminary verification of the local hash cache and the final confirmation of the scheduling control center, the embodiment of the present application can effectively improve the conflict identification efficiency and accuracy in the hash value generation process, not only reduces the network overhead caused by frequent access to the scheduling control center, but also guarantees the uniqueness and consistency of the resource identifier in the global range of the system, thereby enhancing the stability and scalability of the heterogeneous computing power scheduling system.

[0087] Optionally, as to how to generate the resource label, a possible implementation manner is provided as follows. Figure 2The sub-step of generating the resource label of the computing node based on the second field set in step S220 can include: In step S221, the computing node performs a standardized conversion on each second field in the second field set to obtain a standard field corresponding to each second field.

[0088] In the embodiment of the application, each second field is standardized and converted according to a preset field rule to obtain a standard field corresponding to the second field. The standardized conversion includes but is not limited to removing non-identification characters in the field, unifying the representation (such as unifying "CUDA11.8" to "driver=CUDA11.8"), normalizing the architecture name (such as normalizing "x86_64" to "arch=x86"), and the like, to ensure that the subsequent label generation module can correctly parse the field meaning.

[0089] In step S222, the computing node generates a resource label of the computing node according to the standard fields.

[0090] In the embodiment of the application, the preset field rule dictionary is traversed, and each standard field is converted into a corresponding key-value pair according to the requirements of the field rule dictionary, and the resource label of the computing node is generated using the key-value pairs. For example, the standard fields are "chip=GPU", "driver=CUDA11.8", and "arch=x86", and the resource label generated by the computing node is {“chip”:“GPU”,“driver”:“CUDA11.8”,“arch”:“x86”}.

[0091] It should be noted that after the resource label of the computing node is generated, the legality of the resource label needs to be judged, that is, whether the optional field is missing and whether there is a semantic conflict. If the optional field is missing or there is a semantic conflict, the resource label is determined to be illegal, and the state of the resource label is marked as "unschedulable".

[0092] It can be seen that, by generating the resource label after the standardized conversion of the second field set, the embodiment of the application can effectively eliminate the scheduling misjudgment risk caused by the diversity of field sources and the inconsistency of formats, improve the readability and matchability of the resource label in the global resource library, and further enhance the intelligent degree and scheduling accuracy of the heterogeneous computing power scheduling system.

[0093] Optionally, the following provides a possible implementation manner for how to schedule the to-be-processed task. Figure 1 The sub-step of step S120 can include: In step S121, the scheduling control center binds the to-be-processed task with the target hash value.

[0094] In the embodiment of the present application, the mapping relationship between the establishment task of the binding finger and the hash value uniquely identifying the computing node is established, so that the task scheduling path and the actual execution resource form a one-to-one corresponding connection. Specifically, the scheduling control center filters out a target resource tag meeting the task demand from the resource library according to the task tag matching mechanism, and locks the unique computing node through the target hash value corresponding to the target resource tag. The binding operation provides a basis for subsequent model deployment and execution.

[0095] Step S122, the scheduling control center controls the target computing node corresponding to the target hash value to deploy the model.

[0096] In the embodiment of the present application, the model refers to the core running unit of the computing task such as the artificial intelligence model, algorithm component or executable program package required by the task. Before the to-be-processed task is scheduled, the model of the target computing node corresponding to the target hash value needs to be deployed through the scheduling control center.

[0097] Step S123, the target computing node executes the to-be-processed task based on the model.

[0098] In the embodiment of the present application, after the model deployment is completed, the target computing node enters the task execution phase, processes the to-be-processed task according to the model input parameter, and generates the execution result.

[0099] It can be seen that the task binding mechanism and the model deployment control flow based on the hash value in the embodiment of the present application realize the efficient cooperation of task scheduling and execution in the heterogeneous computing power environment, and significantly improve the automation degree, task execution success rate and resource utilization rate of the heterogeneous computing power scheduling system.

[0100] In actual application, in order to support the scheduling of domestic and foreign chips at the same time, the embodiment of the present application designs the heterogeneous computing power adaptation layer based on plug-in, realizes the unified scheduling control of domestic chips and foreign general-purpose chips, each type of computing power corresponds to a standardized plug-in, the plug-in encapsulates the drive calling logic, deployment command conversion and model running interface, and provides a unified interface to the outside, so as to realize the consistency and transparency of task deployment.

[0101] Optionally, as for how to deploy the model, a possible implementation manner is provided as follows. The sub-step of step S122 can include: Firstly, the scheduling control center finds the computing power adaptation plug-in corresponding to the target computing node according to the resource attribute of the target computing node corresponding to the target hash value.

[0102] In the embodiment of the application, the resource attributes include but are not limited to resource types (such as GPU, NPU), driver versions (such as CUDA 11.8, CANN 6.0), and instruction set architectures (such as x86, ARM), and the resource attributes have been reported to the scheduling control center when the computing node is registered to the resource library. The scheduling control center identifies the hardware platform to which the target computing node belongs based on the resource attributes, and matches the corresponding computing power adaptation plug-in from the preset plug-in library.

[0103] Secondly, the scheduling control center loads the computing power adaptation plug-in corresponding to the target computing node locally, and controls the target computing node to load the corresponding computing power adaptation plug-in.

[0104] Finally, the target computing node deploys the model by using the computing power adaptation plug-in.

[0105] In the embodiment of the application, the scheduling control center loads the found computing power adaptation plug-in locally, and sends a loading instruction to the target computing node to control the target computing node to synchronously load the same adaptation plug-in. The target computing node uses the computing power adaptation plug-in to complete model format conversion, preloading, tensor structure mapping based on the running environment of the target computing node, and triggers model startup. The scheduling control center communicates with the plug-in through a standard interface, without needing to perceive chip bottom layer implementation details.

[0106] It can be seen that, by the mechanism of resource attribute matching and plug-in dynamic loading, the embodiment of the application can effectively improve the task scheduling capability and model deployment efficiency of the heterogeneous computing power scheduling system when facing various hardware architectures, reduce the scheduling failure rate caused by hardware differences, and thus enhance the adaptability and expansibility of the system.

[0107] Optionally, in order to improve the traceability management of the heterogeneous computing power scheduling, a possible implementation manner is provided below. The method further includes the following steps: The target computing node sends model deployment information, a scheduling track of the to-be-processed task, and an execution result of the to-be-processed task to the scheduling control center, so that the scheduling control center generates a scheduling link log.

[0108] In the embodiment of the application, after the target computing node completes the model deployment based on the computing power adaptation plug-in, the target computing node records key parameters and state information in the deployment process in real time, including a model loading path, a runtime environment configuration, a dependent library version, and whether deployment is successful or not, to form model deployment information.

[0109] In the actual execution phase of the to-be-processed task, the target computing node continuously tracks and records the running track of the task, including the task start time, resource occupation change, abnormal event in the execution process and task termination reason, and thus a complete scheduling track is constructed. The scheduling track refers to all key path information experienced by the task from scheduling decision, deployment landing to actual execution process, and reflects the life cycle state evolution of the task in the whole system.

[0110] After the execution of the task ends, the target computing node encapsulates the final output result or error log of the task as an execution result, and sends the execution result, the aforementioned model deployment information and the scheduling track to the scheduling control center together.

[0111] The scheduling control center receives the task feedback data from each computing node, and archives and analyzes these information based on a unified data structure, to generate a scheduling link log. The scheduling link log includes but is not limited to task ID, target hash value, scheduling timeline, deployment details and execution result summary, and can be used for auditing, fault backtracking and scheduling strategy optimization.

[0112] It can be seen that, by using the full-process feedback mechanism of the model deployment information, the scheduling track and the execution result, the embodiment of the present application can realize the visual tracking and complete log recording of the task scheduling path in the heterogeneous computing power scheduling system, and significantly improve the auditability, maintainability and intelligent scheduling capability of the system.

[0113] Based on the same inventive concept, the basic principles and technical effects of the heterogeneous computing power scheduling system provided by the embodiments of the present application are the same as those of the above-mentioned embodiments. For brevity of description, the parts not mentioned in this embodiment can be referred to the corresponding contents in the above-mentioned embodiments.

[0114] Please refer to Figure 3 , Figure 3 A block schematic diagram of the heterogeneous computing power scheduling system 300 provided by the embodiment of the present application is shown. The heterogeneous computing power scheduling system 300 includes a scheduling control center 310 and a plurality of computing nodes 320, and the scheduling control center 310 is in communication connection with the computing nodes 320.

[0115] The scheduling control center 310 is configured to generate a task tag according to the hardware requirement of a to-be-processed task.

[0116] The scheduling control center 310 is configured to find a target resource tag matched with the task tag by searching a resource library. The resource library records the correspondence between the resource tags and hash values of the computing nodes 320 in the heterogeneous computing power scheduling system 300. The resource tag is generated based on the hardware attribute information of the computing nodes 320, and the hash value is used to uniquely identify the computing nodes 320.

[0117] The scheduling control center 310 is configured to determine the target computing node 320 based on the target hash value corresponding to the target resource tag, and schedule the to-be-processed task by using the target computing node 320.

[0118] In summary, the heterogeneous computing power scheduling system provided by the embodiment of the application can generate a task tag of a to-be-processed task by a scheduling control center, and find a target resource tag matched with the task tag based on the task tag, wherein the resource tag is automatically generated based on hardware attribute information of a computing node, and the task tag represents hardware requirements of the to-be-processed task on computing resources. By comparing the task tag with the resource tag in the resource library, a computing node suitable for scheduling the to-be-processed task can be accurately found. Further, a target computing node is found by using a hash value uniquely identifying the computing node, and the to-be-processed task is scheduled by using the target computing node. By using the resource unique identification mechanism based on the hash value and the scheduling strategy based on matching of the task tag and the resource tag, the embodiment of the application can effectively improve the accuracy and stability of heterogeneous computing power resource scheduling, and reduce the problem of task scheduling failure caused by resource identification conflict or tag configuration error.

[0119] As a possible implementation manner, as shown in Figure 4 The heterogeneous computing power scheduling system 300 includes a scheduling control center 310 and a plurality of computing nodes 320. The scheduling control center 310 includes a scheduling controller 311 and a resource registration center 312. The computing node 320 includes a resource collector 321 and an adaptive executor 322.

[0120] The scheduling controller 311 is configured to receive a task request, and generate a task tag according to the task request. The scheduling controller 311 is further configured to obtain a matched resource tag based on the task tag, and schedule the received task request based on the resource tag and a corresponding hash value.

[0121] The resource registration center 312 records a resource library, and manages resource information of each computing node in the heterogeneous computing power scheduling system through the resource library.

[0122] The resource collector 321 is configured to collect resource information of the computing node, and generate a hash value and a resource tag corresponding to the computing node based on the resource information.

[0123] The adaptive executor 322 is configured to interact with hardware during task deployment, such as loading a model, performing inference, task cleaning, and the like.

[0124] The heterogeneous computing power scheduling is realized by cooperative work of the scheduling control center 310, the scheduling controller 311, the resource registration center 312, the computing node 320, the resource collector 321, and the adaptive executor 322.

[0125] In several embodiments provided in the present application, it should be understood that the disclosed methods and systems can also be implemented in other manners. The above described system embodiments are merely exemplary. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation modes of the methods and systems according to the embodiments of the present application. In this regard, each block in the flowcharts or block diagrams can represent a module, a segment or a part of code which contains one or more executable instructions for implementing the specified logic function. It should also be noted that in some alternative implementation modes, the functions noted in the blocks can occur in a different order from that noted in the accompanying drawings. For example, two consecutive blocks can actually be executed in a substantially parallel manner, and they can also be executed in a reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and the combination of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system which executes the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0126] In addition, each functional module in the embodiments of the present application can be integrated together to form a separate part, or each module can exist independently, or two or more modules can be integrated to form a separate part.

[0127] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application, essentially or the part that contributes to the prior art, or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the present application.

[0128] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A heterogeneous computing power scheduling method, characterized in that: Applied to a heterogeneous computing power scheduling system, the heterogeneous computing power scheduling system includes a scheduling control center and multiple computing nodes. The method includes: The scheduling control center generates a task tag according to the hardware requirements of the task to be processed; The scheduling control center searches the resource library according to the task tag and finds the target resource tag that matches the task tag; the resource library records the correspondence between the resource tag and the hash value of each computing node in the heterogeneous computing power scheduling system, the resource tag is generated based on the hardware attribute information of the computing node, and the hash value is used to uniquely identify the computing node; The scheduling control center determines a target computing node based on a target hash value corresponding to the target resource tag, and schedules the to-be-processed task using the target computing node.

2. The heterogeneous computing power scheduling method according to claim 1, characterized in that: The scheduling control center searches the resource library according to the task tag to find the target resource tag that matches the task tag, including: The scheduling control center selects at least one candidate resource tag matching the task tag from the resource library; If there is a candidate resource tag, the scheduling control center determines the candidate resource tag as the target resource tag corresponding to the task tag; If there are multiple candidate resource tags, the scheduling control center will perform a weighted score on each candidate resource tag based on resource adaptability, resource availability, resource reliability and resource utilization efficiency, and determine the candidate resource tag with the highest score as the target resource tag corresponding to the task tag.

3. The heterogeneous computing power scheduling method according to claim 1, characterized in that: The method further comprises: When the computing node is newly added to the heterogeneous computing power scheduling system or the resource information of the computing node changes, the computing node obtains hardware attribute information, network configuration information and operation information; The computing node obtains a first field set from the hardware attribute information, the network configuration information, and the operation information, and generates a hash value of the computing node based on the first field set; The computing node obtains a second field set from the hardware attribute information according to a label field rule corresponding to the resource type, and generates a resource label of the computing node based on the second field set; The computing node sends the hash value and resource label of the node to the scheduling control center, so that the scheduling control center adds the hash value and resource label of the computing node to the resource library.

4. The heterogeneous computing power scheduling method according to claim 3, characterized in that: The computing node obtains a first field set from the hardware attribute information, the network configuration information, and the operation information, and generates a hash value of the computing node based on the first field set, including: The computing node obtains a first set of fields from the hardware attribute information, the network configuration information, and the operation information in descending order of field weight and in accordance with a principle of minimum field number; the field weight is positively correlated with field stability; The computing node performs hash calculation based on the first field set using a preset hash algorithm to obtain a temporary hash value, and determines whether there is a conflict in the temporary hash value; If there is no conflict in the temporary hash value, the computing node determines the temporary hash value as the hash value of the computing node; If there is a conflict in the temporary hash value, the computing node selects a field with a high field weight from the remaining fields of the hardware attribute information, the network configuration information and the operation information and adds it to the first field set, and regenerates the corresponding temporary hash value based on the updated first field set until there is no conflict in the temporary hash value, and obtains the hash value of the computing node; the remaining fields do not overlap with the fields in the first field set.

5. The heterogeneous computing power scheduling method according to claim 4, characterized in that: The determining whether the temporary hash value conflicts includes: The computing node checks whether the temporary hash value exists in the local hash cache; the local hash cache is periodically updated based on the resource library; If the temporary hash value exists in the local hash cache, the computing node determines that the temporary hash value has a conflict; If the temporary hash value does not exist in the local hash cache, the computing node queries the scheduling control center whether there is a conflict in the temporary hash value; the scheduling control center determines whether there is a conflict in the temporary hash value based on whether the temporary hash value exists in the resource library.

6. The heterogeneous computing power scheduling method according to claim 3, characterized in that: Generating the resource tag of the computing node based on the second field set includes: The computing node performs a standardization conversion on each second field in the second field set to obtain a standard field corresponding to each second field; The computing node generates a resource tag of the computing node according to each of the standard fields.

7. The heterogeneous computing power scheduling method according to claim 1, characterized in that: The scheduling control center determines a target computing node based on a target hash value corresponding to the target resource tag, and schedules the task to be processed using the target computing node, including: The scheduling control center binds the task to be processed with the target hash value; The scheduling control center controls the target computing node deployment model corresponding to the target hash value; The target computing node executes the task to be processed based on the model.

8. The heterogeneous computing power scheduling method according to claim 7, characterized in that: The scheduling control center controls the target computing node deployment model corresponding to the target hash value, including: The scheduling control center searches for a computing power adapter plug-in corresponding to the target computing node according to the resource attributes of the target computing node corresponding to the target hash value; The scheduling control center locally loads the computing power adaptation plug-in corresponding to the target computing node, and controls the target computing node to load the corresponding computing power adaptation plug-in; The target computing node utilizes the computing power adaptation plug-in deployment model.

9. The heterogeneous computing power scheduling method according to claim 7 or 8, characterized in that: The method further comprises: The target computing node sends the model deployment information, the scheduling trajectory of the task to be processed, and the execution result of the task to be processed to the scheduling control center, so that the scheduling control center generates a scheduling link log.

10. A heterogeneous computing power scheduling system, characterized in that: The heterogeneous computing power scheduling system includes a scheduling control center and multiple computing nodes: The scheduling control center is used to generate task tags according to the hardware requirements of the tasks to be processed; The scheduling control center is used to search the resource library according to the task tag and find the target resource tag that matches it; the resource library records the correspondence between the resource tag and the hash value of each computing node in the heterogeneous computing power scheduling system, the resource tag is generated based on the hardware attribute information of the computing node, and the hash value is used to uniquely identify the computing node; The scheduling control center is used to determine a target computing node based on a target hash value corresponding to the target resource tag, and use the target computing node to schedule the task to be processed.

Images