Cloud aggregation remote computing power scheduling system

Abstract

The invention relates to the technical field of computing power scheduling, in particular to a cloud aggregation remote computing power scheduling system, which comprises a node access confirmation module, a path stability elimination module, a resource node matching module, a task scheduling construction module and a state difference synchronization module. According to the invention, the access validity of the remote node is verified to prevent the access of the abnormal node, the registration result set is generated based on the access information to ensure the controllability of the node identity, and the unstable path is eliminated by integrating the path hop count and the packet loss rate index, so that the path stability and the access success rate are improved; a multi-dimensional matching and screening process is executed based on node capability parameters and task demand parameters, the accuracy of a task scheduling result is ensured, dynamic assignment of a task to a bearable node is realized by performing joint evaluation and judgment on the current scheduling pool capacity, the channel occupancy rate and the response time of the node, the node load balancing effect is improved, and the task scheduling efficiency is improved. Accurate scheduling, efficient operation and dynamic controllable management of distributed remote computing power are realized.

Description

TECHNICAL FIELD

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

[0002] The technical field of computing power scheduling involves systems and methods for dynamically managing and intelligently allocating computing resources. This technical field mainly focuses on the collaborative operation of multi-node and multi-type heterogeneous computing resources, aiming to improve the execution efficiency of computing tasks, resource utilization, and flexibility of task scheduling. Key technologies involved in computing power scheduling include task queue management, resource state perception, scheduling strategy modeling, load balancing, elastic scaling mechanisms, and fault tolerance processing. This technology is commonly applied in high-performance computing clusters, cloud computing platforms, edge computing systems, data center operations, and distributed AI training platforms to solve resource conflicts and performance bottlenecks during multi-task concurrent execution.

[0003] Among them, the cloud computing power remote computing power scheduling system is a system for remote calling of computing resources in different places, used to realize the unified scheduling, dynamic allocation and collaborative operation of computing power resources distributed in different regions or data centers. Through remote access control, cross-network resource abstraction, task dispatching engine and state synchronization mechanism, the system realizes transparent calling and management of remote CPU / GPU resources, improves the coverage and execution flexibility of computing power scheduling, Traditional scheduling systems only realize remote computing power calling through remote access control and task dispatching engine. In actual execution, no effective discrimination mechanism is established for the authenticity and credibility of the access nodes, which may lead to the mixing of potential abnormal nodes into the scheduling system, affecting the safety of computing power resource scheduling. In the selection of node paths, stability factors such as path hop count and data transmission packet loss rate are not fully considered, which may cause access interruption or task failure in network fluctuation scenarios. There is a lack of fine-grained matching mechanism between tasks and nodes, which may lead to insufficient task resource adaptation or node load bias. No dynamic response evaluation process is established for the current load and response capability of the nodes, which may cause task scheduling delay or system response delay. At the same time, no difference perception and updating mechanism is established for the running state synchronization, which may cause state lag or scheduling error. SUMMARY

[0004] The purpose of the present application is to solve the problems existing in the prior art, and to provide a cloud computing power remote computing power scheduling system.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme: a cloud computing power remote computing power scheduling system, the system comprises: The node access confirmation module obtains remote computing power node information, if the link verification value is consistent with the corresponding node identifier in the whitelist, and the cluster attribution identifier is not the isolation cluster code, the node registration identifier is written into the computing power scheduling registration table, and a remote node registration result set is generated; The path stability elimination module determines whether the path hop value exceeds the upper limit threshold of the hop value and whether the packet loss rate exceeds the critical value of the packet loss rate according to the node registration identifier in the remote node registration result set, and eliminates the corresponding path when both conditions are met, and generates an available node access path set. The resource node matching module extracts the node capability parameters according to the node registration identifier in the available node access path set, extracts the task demand parameters of the task to be dispatched, calculates and selects the node with the highest matching score of the task and the node, and generates a task target node mapping set. The task scheduling construction module calls the node registration identifier in the task target node mapping set, if the average response time is lower than the set response time threshold, and the scheduling channel occupancy rate is lower than the available limit, the current task is mapped to the target node scheduling pool, and a dispatched task scheduling record is generated.

[0006] As a further scheme of the application, the remote node registration result set includes access node identifier number, scheduling authority level code, activation state identifier and computing power attribution category, the available node access path set includes path stability identifier, remaining link bandwidth, transmission success rate and path reliability level, the task target node mapping set includes node preferred index, task mapping level value, node resource allocation relationship and node selection priority sequence, and the dispatched task scheduling record is specifically task dispatched node number, scheduling execution period value, scheduling channel mapping record and cache writing marker set.

[0007] As a further scheme of the application, the node access confirmation module includes: The node information acquisition submodule obtains remote computing power node information, the node information includes node registration identifier, cluster attribution identifier, access link verification value, calls node management interface to combine and encode the three information with acquisition time stamp, performs node identity uniqueness judgment based on the encoding result, records the node registration identifier meeting the uniqueness condition to the node temporary cache area, and generates effective node identification information; The link validity discrimination submodule extracts the access link verification value from the node temporary cache area according to the node registration identifier associated with the effective node identification information, and compares the access link verification value with each channel identifier in the access channel whitelist, when the verification value is consistent with any channel identifier in the whitelist, and the cluster attribution identifier of the node is not equal to the isolation cluster code, the corresponding node identifier is written into the link validity list, and a whitelist hit comparison result is generated. The node registration writing submodule calls link valid list content corresponding to the whitelist hit comparison result, extracts a node registration identifier, and combines an access IP address of a current period of the node, a connection stable duration, and a data interaction success rate to determine whether the registration identifier is in an unregistered state in a node management table, and if the condition is met, writes the registration identifier into a registration record to generate a remote node registration result set.

[0008] As a further scheme of the application, the path stability elimination module comprises: The link parameter collection submodule obtains a node registration identifier in the remote node registration result set, extracts path hop counts of each node to a dispatching hub, a number of sent data packets in three consecutive rounds of transmission, and corresponding lost data packet counts, records the node identifier and the three link parameters in association, establishes a structured link data set, and generates a link quality monitoring data set; The path abnormality discrimination submodule discriminates nodes for which both a path hop count and a packet loss rate exceed a set upper threshold of the hop count and a set critical threshold of the packet loss rate according to the hop count and the packet loss rate of each link in the link quality monitoring data set, and marks the nodes as abnormal link paths to generate a path elimination determination result; The available path screening submodule screens node registration identifiers that are not marked as abnormal links in the path elimination determination result, excludes the elimination identifiers from original link records, organizes remaining node path information, aggregates available access path numbers and corresponding path parameters according to a node dimension, and generates an available node access path set.

[0009] As a further scheme of the application, the resource node matching module comprises: The parameter extraction submodule obtains a node registration identifier in the available node access path set, collects node capability parameters of the corresponding node, including a maximum number of concurrent threads of a GPU, a number of available CPU cores, and a remaining capacity of a node local buffer, simultaneously extracts task demand parameters of a task to be dispatched, including a computation intensity index, a thread concurrency requirement, and a task cache demand amount, establishes a parameter association mapping according to a node and task binding relationship, and generates a node task parameter association data set; The level mapping submodule obtains a matching score of a task in a current node according to the node task parameter association data set by using the node capability parameters and the task demand parameters, ranks the matching scores of the nodes in descending order, eliminates nodes lower than a minimum acceptable score, and generates a matchable node score sequence; The scoring screening submodule extracts a node registration identifier with the highest score according to each node matching score in the matching score sequence, constructs a matching successful node list, establishes a one-to-one binding relationship data structure between a task and a node according to the list, and generates a task target node mapping set.

[0010] As a further scheme of the present application, the formula for obtaining the matching score of the current node by the operation is specifically: ; Among them, is the matching score, is a normalized value of the number of available CPU cores, is a normalized value of the calculation intensity index, is a normalized value of the maximum number of concurrent threads of the GPU, is a normalized value of the thread concurrency requirement, is a normalized value of the remaining capacity of the local buffer, is a normalized value of the task cache demand, is a task priority level parameter.

[0011] As a further scheme of the present application, the task scheduling construction module comprises: The scheduling parameter extraction submodule calls the node registration identifier in the task target node mapping set, obtains the three indexes of the scheduling task pool capacity, the scheduling channel occupancy rate and the average response time corresponding to each node, binds the node registration identifier and the three indexes for storage, forms a mapping structure, and generates a node scheduling state information group; The channel availability judgment submodule judges whether the average response time is less than a set response time threshold and whether the scheduling channel occupancy rate is less than an available occupancy limit threshold according to the two indexes of the average response time and the scheduling channel occupancy rate of each node in the node scheduling state information group, records the node registration identifier that meets the two conditions, and generates an available node screening result set; The task mapping binding submodule calls the node registration identifier in the available node screening result set, extracts the task number to be scheduled and establishes a task binding comparison table with the node registration identifier, writes the mapping information into a task dispatching scheduling pool, and generates a dispatching task scheduling record.

[0012] As a further scheme of the present application, the system further comprises: The state difference synchronization module calculates a state synchronization deviation degree value according to the target node registration identifier in the dispatching task scheduling record, compares the state synchronization deviation degree value with a set state synchronization threshold value, and if the state synchronization deviation degree value is greater than the threshold value, updates the current state record to a scheduling control state table, and generates a synchronization node running state set; The set of running states of the synchronization node specifically refers to a node state update label, a task processing progress code, a node state drift intensity value and a synchronization activation timestamp.

[0013] As a further scheme of the application, the state difference synchronization module comprises: The state data extraction submodule acquires the target node registration identification in the dispatched task scheduling record, extracts the CPU usage rate, the GPU active thread number and the scheduling queue length in the current scheduling period, simultaneously extracts the corresponding three indexes of the same node in the last scheduling period, and constructs a node period state comparison table to generate a period state comparison data group; The deviation value calculation submodule calculates the state synchronization deviation value of the current node according to the state indexes of the same node in the two periods in the period state comparison data group, compares the deviation value with a state synchronization trigger threshold value, records the node identification greater than the threshold value, and generates a state deviation determination result set; The synchronization condition judgment submodule extracts the current index value of the node meeting the state synchronization condition from the original period state comparison data group according to the node identification in the state deviation determination result set, writes the node identification and the corresponding state into a scheduling control state table, establishes a synchronization record data, and generates a set of running states of the synchronization node.

[0014] As a further scheme of the application, the formula for calculating the state synchronization deviation value of the current node is specifically: ; Among them, represents the state synchronization deviation value of the node, represents the CPU usage rate in the current period, represents the CPU usage rate in the last period, represents the period average of the CPU usage rate, represents the GPU active thread number in the current period, represents the GPU active thread number in the last period, represents the period average of the GPU active thread number, represents the scheduling queue length in the current period, represents the scheduling queue length in the last period, represents the period average of the scheduling queue length.

[0015] Compared with the prior art, the application has the following advantages and positive effects: In the application, the abnormal node access is prevented by verifying the access legality of the remote node, the controllability of the node identity is ensured by generating the registration result set based on the access information, the unstable path is removed by comprehensively considering the path hop count and the packet loss rate index, the path stability and the access success rate are improved, the multi-dimensional matching and screening process is performed based on the node capability parameter and the task demand parameter, the accuracy of the task scheduling result is ensured, the dynamic distribution of the task to the bearable node is realized by jointly evaluating and judging the current scheduling pool capacity of the node, the channel occupancy rate and the response time, the node load balancing effect is improved, the periodic quantitative difference detection of the node state is performed and the state table update is triggered, the operation state of each node is kept synchronous and mastered by the scheduling hub, and the precise scheduling, efficient operation and dynamic controllable management of the distributed remote computing power are realized. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0017] Figure 1 The system flowchart of the present application is shown in the figure. Figure 2 The system framework schematic diagram of the present application is shown in the figure. Figure 3 The flowchart of the node access confirmation module of the present application is shown in the figure. Figure 4 The flowchart of the path stability elimination module of the present application is shown in the figure. Figure 5 The flowchart of the resource node matching module of the present application is shown in the figure. Figure 6 The flowchart of the task scheduling construction module of the present application is shown in the figure. Figure 7 The flowchart of the state difference synchronization module of the present application is shown in the figure. DETAILED DESCRIPTION

[0018] The technical solutions in the present application will be described below in combination with the drawings.

[0019] In the embodiments of the present application, the words such as "example", "for example" are used to represent as an example, illustration or description. Any embodiment or design scheme described as "example" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. In fact, the word "example" is intended to present the concept in a specific way. In addition, in the embodiments of the present application, the meaning expressed by "and / or" can be both, or can be one of the two.

[0020] In the embodiments of the present application, "image" and "picture" can be used interchangeably, and it should be pointed out that their meanings are consistent when the distinction is not emphasized.

[0021] In the embodiments of the present application, sometimes the subscript such as W1 can be written in the form of non-subscript such as W1, and the meanings are consistent when the distinction is not emphasized.

[0022] In order to make the technical problems, technical solutions and advantages of the present application clearer, specific embodiments will be described in detail below with reference to the drawings.

[0023] Please refer to Figure 1 , the cloud aggregation algorithm remote computing power scheduling system, the system includes node access confirmation module, path stability elimination module, resource node matching module, task scheduling construction module and state difference synchronization module; The node access confirmation module obtains remote computing power node information, the node information includes node registration identifier, cluster attribution identifier, access link verification value, and the access link verification value is matched with the access channel whitelist to judge, if the link verification value is consistent with the corresponding node identifier in the whitelist, and the cluster attribution identifier is not the isolation cluster code, then the node registration identifier is written in the computing power scheduling registration table, and the remote node registration result set is generated; The path stability elimination module obtains the path hop count value of the node to the scheduling hub, the data packet loss count and the total number of transmissions in the continuous three rounds of transmission process according to the node registration identifier in the remote node registration result set, judges whether the path hop count value exceeds the upper limit threshold of the hop count, and judges whether the ratio between the data packet loss count and the total number exceeds the packet loss rate critical value, when both conditions are met, the corresponding path is eliminated, the remaining path is arranged, and the available node access path set is generated; The resource node matching module extracts the node capability parameters according to the node registration identifier in the available node access path set, extracts the task demand parameters of the task to be dispatched at the same time, calculates the matching score of the task and the node, selects the node registration identifier with the highest matching score to construct the scheduling target list, and generates the task target node mapping set; The task scheduling construction module calls the node registration identifier in the task target node mapping set, extracts the scheduling task pool capacity, scheduling channel occupancy rate and average response time of each node, compares the average response time with the set response time threshold value, if the response time is lower than the threshold value and the scheduling channel occupancy rate is lower than the available limit, then the current task is mapped to the target node scheduling pool, and the dispatched task scheduling record is generated. The state difference synchronization module obtains CPU usage, GPU active thread number, and scheduling queue length in the current scheduling period according to the target node registration identifier in the dispatch task scheduling record, and performs difference calculation on the three indicators of the same node in the last period respectively, calculates the state synchronization deviation degree value, compares it with the set state synchronization threshold value, and if it is greater than the threshold value, the current state record is updated to the scheduling control state table, and a synchronization node running state set is generated.

[0024] The remote node registration result set includes access node identifier number, scheduling authority level code, activation state identifier, and computing power attribution category, the available node access path set includes path stability identifier, remaining link bandwidth, transmission success ratio, and path reliability level, the task target node mapping set includes node preferred index, task mapping level value, node resource allocation relationship, and node selection priority sequence, the dispatch task scheduling record is specifically task dispatch node number, scheduling execution period value, scheduling channel mapping record, and cache write marker set, and the synchronization node running state set is specifically node state update label, task processing progress code, node state drift intensity value, and synchronization activation timestamp.

[0025] Please refer to Figure 2 and Figure 3 , the node access confirmation module includes: The node information acquisition submodule acquires remote computing power node information, and the node information includes node registration identifier, cluster attribution identifier, and access link verification value, calls a node management interface to combine and encode the three pieces of information with the acquisition timestamp, performs node identity uniqueness judgment based on the encoding result, records the node registration identifier that meets the uniqueness condition to a node temporary buffer area, and generates valid node identification information; The node information acquisition submodule acquires remote computing power node information, and the module receives an active connection request from a remote computing power node through a registration listening port, each request carries a node registration identifier, a cluster attribution identifier, and an access link verification value, for example, a remote node named "node A" carries its own "cluster B" cluster identifier and the verification value "123456" corresponding to its access channel, after the connection request arrives, a preset node management interface is called to combine the three pieces of information with the current system timestamp "2025-08-26 10:00:00" to form a unique code, the code form is "node A-cluster B-123456-2025-08-26 10:00:00", by querying a node historical registration database, whether the combined code exists in the database is compared, when the query result is empty, it is determined that the code is the unique identity of the new network node, and "node A" is recorded as a valid node identifier. ​​​​The node registration identifier is recorded in the node's temporary cache area, and the contents of the temporary cache area are used as input for subsequent steps to generate valid node identification information.

[0026] The link validity determination submodule extracts the access link verification value from the node temporary cache area based on the node registration identifier associated with the valid node identification information and compares it with each channel identifier in the access channel whitelist. When the verification value is consistent with any channel identifier in the whitelist and the node's cluster affiliation identifier is not equal to the isolation cluster code, the corresponding node identifier is written into the link validity list to generate a whitelist hit comparison result. The link validity determination submodule reads the valid node identification information generated in the first step from the node temporary cache based on the node registration identifier associated with the valid node identification information. For example, it extracts "" from the temporary cache. The node registration identifier is used as an index to retrieve the associated access link verification value from the cache. For example, if the verification value is retrieved... The verification value is compared one by one with a preset access channel whitelist. The whitelist contains a set of channel identifiers allowed by the system. For example, the whitelist content is " "", "", When the verification value is " "Compare with the first channel identifier in the whitelist" If they match, the system continues to acquire data. "Cluster ownership identifier" ", and then combine it with the preset isolated cluster code " "Compare, when" "Not equal to" When the node is deemed to meet the access requirements, it will be set to " The node identifier is written into the list of valid links, and this list is used as the dataset for subsequent processing to generate whitelist match comparison results.

[0027] The node registration writing submodule calls the whitelist hit comparison result corresponding to the valid list of links, extracts the node registration identifier and combines it with the node's current access IP address, connection stability duration and data interaction success rate to determine whether the registration identifier is in an unregistered state in the node management table. If the conditions are met, it is written into the registration record to generate a remote node registration result set. The node registration and writing submodule calls the whitelist to match the valid link list content corresponding to the comparison result. The system obtains the valid link list, for example, by extracting "" from the list. The system then obtains the access IP address for the current time period from the node's registration identifier, for example, obtaining the node's registration identifier. ", the system records the stable time of the node's continuous connection, for example, the continuous connection time reaches 3600 seconds, the system counts the proportion of successful interaction by continuously sending and receiving data packets, for example, the data interaction success rate is 99.8%, the system queries the node management table to check whether there is a registration record that completely matches the registration identifier of " ". When the query result shows that there is no record in the node management table, it is determined that it is not registered, and a new record containing the registration identifier of " ", the access IP address of " ", the stable connection duration of 3600 seconds, and the data interaction success rate of 99.8% is written into the node management table, which constitutes the remote node registration result set.

[0028] Please refer to Figure 2 and Figure 4 , the channel stability elimination module includes: The link parameter acquisition submodule obtains the node registration identifier in the remote node registration result set, extracts the path hop count of each node to the dispatching hub, the number of sent data packets in three consecutive rounds of transmission and the corresponding lost data packet count, and records the node identifier and the three link parameters in association, establishes a structured link data set, and generates a link quality monitoring data set; The link parameter acquisition submodule obtains the node registration identifier in the remote node registration result set, the system obtains the node registration identifier written in the first step from the remote node registration result set, for example, the identifier of " " is obtained, the system measures the path hop count from " " to the dispatching hub through a network detection tool, for example, the measurement value is 5 hops, in the continuous three rounds of data transmission test, the system records the number of sent data packets in each round, for example, 1000, 1000, 1000, and the corresponding lost data packet count, for example, 2, 1, 0, the system binds the identifier of " " with its path hop count, the number of sent data packets and the lost data packet count, establishes a structured data record, and sorts it into a link data set, which constitutes a link quality monitoring data set.

[0029] Table 1 Link quality monitoring data set ; As shown in Table 1, the link quality monitoring data set lists the link parameters of three nodes, including path hop count and three rounds of packet transmission data.

[0030] The path anomaly identification submodule uses the hop count and packet loss rate of each link in the link quality monitoring dataset to call the set upper limit threshold for hop count and the critical threshold for packet loss rate. It filters out nodes that meet both conditions simultaneously, such as whether the number of hops exceeds the upper limit threshold for hop count and whether the packet loss rate exceeds the critical threshold for packet loss rate, and marks them as abnormal link paths, generating path removal judgment results. The path anomaly detection submodule obtains the link quality monitoring dataset based on the hop count and packet loss rate for each link in the aforementioned link quality monitoring dataset. For example, it extracts "..." from the dataset. The system calculates the packet loss rate based on the number of sent data packets and the count of lost data packets, using the link record of "". The packet loss rate is calculated by adding the number of lost data packets in three rounds, for example, " Then add the number of data packets sent in the three rounds, for example, " Divide the former by the latter, that is, " The packet loss rate is 0.1%. The system call's hop count upper limit threshold is compared with the packet loss rate critical threshold. The hop count upper limit threshold is set with reference to the 95th percentile of the path hop count of historically normally operating nodes. For example, if the path hop count of 100 historically normally operating nodes is statistically analyzed, the 95th percentile is 8, therefore the hop count upper limit threshold is set to 8. The packet loss rate critical threshold is set with reference to the average packet loss rate of historically normally operating nodes multiplied by a fluctuation adjustment coefficient. For example, if the historical average packet loss rate is 0.5%, and the fluctuation adjustment coefficient is 1 or 5, then the critical threshold is... The system judges " The system checks whether the path hop count of '5' exceeds the upper limit threshold of 8 hops and whether its packet loss rate of 0.1% exceeds the critical packet loss rate threshold of 0.75%. Since neither condition is met, the system does not assign the path hop count to '5'. "Mark it as an abnormal link path, record the judgment result, and generate a path removal judgment result."

[0031] The available path filtering submodule calls the node registration identifiers that are not marked as abnormal links in the path elimination judgment results, filters out the elimination identifiers from the original link records, organizes the remaining node path information, summarizes the available access path numbers and corresponding path parameters by node dimension, and generates a set of available node access paths. The available path filtering submodule can call the registration identifiers of nodes not marked as abnormal links in the path removal judgment results. The system retrieves nodes not marked as abnormal links from the path removal judgment results, for example, retrieving " The system uses the node registration identifier "" to exclude all nodes marked as abnormal links from the original link records, for example, filtering out nodes marked as abnormal links from the original records. Organize the remaining node path information, and then... The path parameters and node identifiers are summarized to establish a mapping relationship between available access path numbers and path parameters, generating a set of available node access paths.

[0032] Please see Figure 2 and Figure 5 The resource node matching module includes: The parameter extraction submodule obtains the node registration identifiers in the set of available node access paths, collects the node capability parameters of the corresponding nodes, including the maximum number of concurrent GPU threads, the number of available CPU cores, and the remaining capacity of the node's local buffer. At the same time, it extracts the task requirement parameters of the tasks to be assigned, including the computational density index, thread concurrency requirements, and task cache requirements. It establishes a parameter association mapping according to the node-task binding relationship and generates a node-task parameter association dataset. The parameter extraction submodule obtains the node registration identifier from the available node access path set, and the system obtains "" from the available node access path set. The system uses the node registration identifier to query the node and collect its capability parameters via an interface. For example, it collects data showing that the maximum concurrent GPU threads are 128, the number of available CPU cores is 16, and the remaining capacity of the node's local buffer is 50GB. Simultaneously, the system retrieves tasks to be assigned from the task dispatch queue. The task requires specific parameters, such as a computational density of 0.8, a thread concurrency requirement of 64, and a task cache requirement of 10GB. The system will then... "and" "Binding establishes parameter association mapping, stores all collected nodes and task parameters together, and generates a node task parameter association dataset."

[0033] The grade mapping submodule associates the dataset with the node task parameters, calculates the matching score of the task at the current node using the node capability parameters and task requirement parameters, sorts the matching scores of the nodes in descending order, removes nodes with scores lower than the minimum acceptable score, and generates a sequence of matching node scores. The formula for obtaining the matching score of the task at the current node is as follows: ; in, To match scores, This is a normalized value for the number of available CPU cores. To calculate the normalized value of the density index, This is a normalized value for the maximum concurrent thread count of the GPU. This is a normalized value for thread concurrency requirements. This is the normalized value of the remaining capacity of the local buffer. The normalized value of the task cache demand is used. The task priority parameter represents the relative urgency of a task in the queue. It is usually derived from the task classification priority mapping in the scheduling strategy. All normalized values ​​are converted from the original dimensional values ​​to dimensionless values ​​through the min-max normalization method, so that the parameters are combined under the same evaluation standard. The minimum acceptable score is obtained by collecting the average thread concurrency requirements and total cache requirements in the current round of tasks, combining them with the average capacity of the currently schedulable CPU and GPU nodes, constructing an acceptable score baseline, and taking the median or quantile as the elimination threshold. The level mapping submodule associates the node task parameters with the dataset. The system obtains the node task parameter association dataset, and uses the node capability parameters and task requirement parameters to calculate the matching score of the task at the current node. The formula is as follows: ; in, To match scores, This is a normalized value for the number of available CPU cores. To calculate the normalized value of the density index, This is a normalized value for the maximum concurrent thread count of the GPU. This is a normalized value for thread concurrency requirements. This is the normalized value of the remaining capacity of the local buffer. The normalized value of the task cache demand is used. This is a task priority parameter, whose value is derived through a task category priority mapping, such as the priority of urgent tasks. 3, high priority task 2, Normal Task The advantage of the formula is that it is 1. The parameters are obtained through the logarithmic function. This process is implemented to slow down the rate at which task priority affects the final matching score as priority increases, thus preventing high-priority tasks from excessively dominating the score. and Divide and on Taking the square root reflects the efficiency of the match between the number of CPU cores and the computational density requirements. This directly reflects the matching ratio between GPU threads and task thread requirements. This reflects the matching ratio between buffer capacity and task caching requirements; the three parts are added together and then multiplied. By averaging, the system ensures that the matching degree of the three dimensions of CPU, GPU, and storage is comprehensively considered in the final score. The system first normalizes all raw parameters. For example, for the number of available CPU cores, it uses a maximum-minimum normalization method. Given three nodes with 16, 32, and 8 available CPU cores respectively, the maximum value is 32, and the minimum value is 8. the normalized value of , if the task computation intensity index (CI) is set between 0.2 and 1.0, then the normalized value of , if the maximum number of concurrent threads of the GPU (GPU_max) is set between 64 and 256, then the normalized value of , if the thread concurrency requirement (TCR) is set between 32 and 128, then the normalized value of , if the remaining capacity of the local buffer (LC) is set between 20 GB and 100 GB, then the normalized value of , if the task cache requirement (TCR) is set between 5 GB and 20 GB, then the normalized value of , if the task priority is set as high, then , the value is 2, and the above normalized values are substituted into the formula to calculate the matching score of The minimum acceptable score is obtained by: the average thread concurrency requirement and the total cache requirement of the current round of tasks are counted, for example, the average thread concurrency requirement is 60, the total cache requirement is 12 GB, and the average of the CPU and GPU capabilities of the currently schedulable nodes is combined, for example, the CPU average is 20 cores and the GPU average is 150 threads, to build an acceptable score baseline, for example, the score baseline is 0.5, and the median or quantile is taken as the rejection threshold, for example, the median 0.5, the nodes with a calculated score lower than 0.5 are rejected, and the matching scores of the remaining nodes are sorted in descending order to generate a matching node score sequence.

[0034] The scoring and screening submodule registers the node identifier with the highest score according to the matching score of each node in the matching node score sequence, constructs a matching successful node list, establishes a one-to-one binding relationship data structure between the task and the node according to the list, and generates a task target node mapping set; The scoring and screening submodule registers the node identifier with the highest score according to the matching score of each node in the matching node score sequence, constructs a matching successful node list, establishes a one-to-one binding relationship data structure between the task and the node according to the list, and generates a task target node mapping set; The scoring and screening submodule registers the node identifier with the highest score according to the matching score of each node in the matching node score sequence, constructs a matching successful node list, establishes a one-to-one binding relationship data structure between the task and the node according to the list, and generates a task target node mapping set; The scoring and screening submodule registers the node identifier with the highest score according to the matching score of each node in the matching node score sequence, constructs a matching successful node list, establishes a one-to-one binding relationship data structure between the task and the node according to the list, and generates a task target node mapping set; The scoring and screening submodule registers the node identifier with the highest score according to the matching score of each node in the matching node score sequence, constructs a matching successful node list, establishes a one-to-one binding relationship data structure between the task and the node according to the list, and generates a task target node mapping set; The scoring and screening submodule registers the node identifier with the highest score according to the matching score of each node in the matching node score sequence, constructs a matching successful node list, establishes a one-to-one binding relationship data structure between the task and the node according to the list, and generates a task target node mapping set; The scoring and screening submodule registers the node identifier with the highest score according to the matching score of each node in the matching node score sequence, constructs a matching successful node list, establishes a one-to-one binding relationship data structure between the task and the node according to the list, and generates a task target node mapping set; ​​​​​​​", the mapping information is taken as the input of the next step to generate a task target node mapping set.

[0035] Please refer to Figure 2 and Figure 6 , the task scheduling construction module comprises: The scheduling parameter extraction submodule calls the node registration identifier in the task target node mapping set, obtains the scheduling task pool capacity, the scheduling channel occupancy rate and the average response time corresponding to each node, binds the node registration identifier and the three indicators for storage to form a mapping structure, and generates a node scheduling state information group; The scheduling parameter extraction submodule calls the node registration identifier in the task target node mapping set, and the system obtains the node registration identifier from the task target node mapping set, for example, the identifier of " " is obtained. The system collects the scheduling task pool capacity of " ", for example, the capacity is 100 tasks, obtains the scheduling channel occupancy rate, for example, the occupancy rate is 45%, and obtains the average response time, for example, 20 milliseconds. The system binds the node registration identifier of " " with the above three indicators to form a mapping structure, for example, the association data of " " is "capacity: 100, occupancy rate: 45%, average response time: 20ms". The mapping structure constitutes a node scheduling state information group.

[0036] The channel availability discrimination submodule discriminates whether the average response time is less than a set response time threshold value and whether the scheduling channel occupancy rate is less than an available occupancy limit threshold value according to the average response time and the scheduling channel occupancy rate of each node in the node scheduling state information group, records the node registration identifiers that meet the two conditions, and generates an available node screening result set; The channel availability discrimination submodule discriminates whether the average response time is less than a set response time threshold value and whether the scheduling channel occupancy rate is less than an available occupancy limit threshold value according to the average response time and the scheduling channel occupancy rate of each node in the node scheduling state information group, records the node registration identifiers that meet the two conditions, and generates an available node screening result set; The channel availability discrimination submodule discriminates whether the average response time is less than a set response time threshold value and whether the scheduling channel occupancy rate is less than an available occupancy limit threshold value according to the average response time and the scheduling channel occupancy rate of each node in the node scheduling state information group, records the node registration identifiers that meet the two conditions, and generates an available node screening result set; The channel availability discrimination submodule discriminates whether the average response time is less than a set response time threshold value and whether the scheduling channel occupancy rate is less than an available occupancy limit threshold value according to the average response time and the scheduling channel occupancy rate of each node in the node scheduling state information group, records the node registration identifiers that meet the two conditions, and generates an available node screening result set; The node registration identifier is taken as an input for the next step, and a set of available node screening results is generated.

[0037] The task mapping and binding sub-module calls the node registration identifier in the set of available node screening results, extracts the task number to be scheduled, and establishes a task binding table with the node registration identifier, writes the mapping information into the task dispatching and scheduling pool, and generates a dispatching task scheduling record; The task mapping and binding sub-module calls the node registration identifier in the set of available node screening results, and the system obtains the node registration identifier of the task to be scheduled from the set of available node screening results. The task mapping and binding sub-module calls the node registration identifier in the set of available node screening results, and the system obtains the node registration identifier of the task to be scheduled from the set of available node screening results. The task mapping and binding sub-module calls the node registration identifier in the set of available node screening results, and the system obtains the node registration identifier of the task to be scheduled from the set of available node screening results. The task mapping and binding sub-module calls the node registration identifier in the set of available node screening results, and the system obtains the node registration identifier of the task to be scheduled from the set of available node screening results.

[0038] Please refer to Figure 2 and Figure 7 , the state difference synchronization module includes: The state data extraction sub-module obtains the target node registration identifier in the dispatching task scheduling record, extracts the CPU usage, GPU active thread number, and scheduling queue length in the current scheduling period, extracts the corresponding three indicators of the same node in the last scheduling period, and constructs a node period state comparison table to generate a period state comparison data group. The state difference synchronization module obtains the target node registration identifier in the dispatching task scheduling record, and the system obtains the target node registration identifier from the dispatching task scheduling record, for example, the node registration identifier of the task to be scheduled is obtained. The state difference synchronization module obtains the target node registration identifier in the dispatching task scheduling record, and the system obtains the target node registration identifier from the dispatching task scheduling record, for example, the node registration identifier of the task to be scheduled is obtained. The state difference synchronization module obtains the target node registration identifier in the dispatching task scheduling record, and the system obtains the target node registration identifier from the dispatching task scheduling record, for example, the node registration identifier of the task to be scheduled is obtained. The state difference synchronization module obtains the target node registration identifier in the dispatching task scheduling record, and the system obtains the target node registration identifier from the dispatching task scheduling record, for example, the node registration identifier of the task to be scheduled is obtained.

[0039] Table 2: Node period state comparison table ; Referring to Table 2, the table lists three key performance indicators of in the current and last scheduling periods.

[0040] The bias value calculation submodule calculates the state synchronization bias value of the current node according to the state indicators of the same node in two periods in the period state comparison data set, compares the bias value with the state synchronization trigger threshold, records the node identifier greater than the threshold, and generates a state bias determination result set; The formula for calculating the state synchronization bias value of the current node is specifically: ; Wherein, represents the state synchronization bias value of the node, represents the CPU usage rate in the current period, represents the CPU usage rate in the last period, represents the period average of the CPU usage rate, represents the GPU active thread number in the current period, represents the GPU active thread number in the last period, represents the period average of the GPU active thread number, represents the scheduling queue length in the current period, represents the scheduling queue length in the last period, represents the period average of the scheduling queue length, and the state synchronization trigger threshold is a dynamic reference value generated by the system in the initialization stage according to the resource type, the scheduling period granularity and the historical state fluctuation amplitude of the node. Specifically, the state bias values of the same type of nodes in multiple periods are averaged and multiplied by a fluctuation adjustment coefficient to obtain the state bias value. The adjustment coefficient is set according to the resource type and the task intensity, and is usually valued between 1, 2 and 1, 6, which is used to control the trade-off relationship between the synchronization frequency and the resource consumption. The bias value calculation submodule calculates the state indicators of the same node in two periods in the period state comparison data set, the system obtains the period state comparison data set, extracts the current period indicators of the state bias value of the node and the last period indicators , using the formula: ; calculates the state synchronization bias value of the node, wherein, is the state synchronization bias value of the node, and are the CPU usage rates in the current and last periods, respectively, is the period average of the CPU usage rate, and are the GPU active thread numbers in the current and last periods, respectively, is the period average of the GPU active thread number, and are the scheduling queue lengths in the current and last periods, respectively,​ To schedule the cycle mean value of the queue length, the core logic of the formula is to calculate the square root of the square sum of the relative change of each index in two cycles, that is, the root mean square, so as to comprehensively measure the overall fluctuation degree of the node state. First, the cycle mean value of each index is calculated: , , Second, each value is substituted into the formula for calculation: ; The result shows that the state synchronization deviation value of the node is 0.203. The system compares the deviation value with the state synchronization trigger threshold value. The setting process of the state synchronization trigger threshold value is as follows: the system counts the state deviation values of the same type of nodes in multiple historical cycles in the initialization stage. For example, the average deviation value of the same type of nodes in the past 100 cycles is 0.125. The system sets the fluctuation adjustment coefficient according to the resource type and the task intensity. For example, for GPU computing tasks, the adjustment coefficient is set to 1, 5, and the state synchronization trigger threshold value is Since the node deviation value 0.203 is greater than the threshold value 0.1875, the system records the node identifier "node 1" and takes the identifier as a set for subsequent processing to generate a state deviation determination result set.

[0041] The synchronization condition judgment submodule extracts the current index value of the node that meets the state synchronization condition from the original cycle state comparison data group according to the node identifier in the state deviation determination result set, writes the node identifier and the corresponding state into the scheduling control state table, establishes synchronization record data, and generates a synchronization node running state set. The synchronization condition judgment submodule extracts the current index value of the node that meets the state synchronization condition from the original cycle state comparison data group according to the node identifier in the state deviation determination result set, writes the node identifier and the corresponding state into the scheduling control state table, establishes synchronization record data, and generates a synchronization node running state set. The synchronization condition judgment submodule extracts the current index value of the node that meets the state synchronization condition from the original cycle state comparison data group according to the node identifier in the state deviation determination result set, writes the node identifier and the corresponding state into the scheduling control state table, establishes synchronization record data, and generates a synchronization node running state set. The synchronization condition judgment submodule extracts the current index value of the node that meets the state synchronization condition from the original cycle state comparison data group according to the node identifier in the state deviation determination result set, writes the node identifier and the corresponding state into the scheduling control state table, establishes synchronization record data, and generates a synchronization node running state set. The synchronization condition judgment submodule extracts the current index value of the node that meets the state synchronization condition from the original cycle state comparison data group according to the node identifier in the state deviation determination result set, writes the node identifier and the corresponding state into the scheduling control state table, establishes synchronization record data, and generates a synchronization node running state set.

[0042] ​It should be understood that the term "and / or" in this document is merely used to describe associated objects, and it is possible that there are three relationships, for example, A and / or B can mean that A exists alone, A and B exist together, B exists alone, and A, B can be singular or plural. In addition, the character " / " in this document generally represents an "or" relationship between the front and rear associated objects, but it can also represent an "and / or" relationship, which can be understood according to the context before and after.

[0043] In this application, "at least one" means one or more, and "multiple" means two or more. "At least one of the following" or the like means any combination of the items, including single item or any combination of multiple items. For example, at least one of a, b, or c can mean a, b, c, a-b, a-c, b-c, or a-b-c, where a, b, and c can be singular or plural.

[0044] It should be understood that the size of the sequence number of the above-mentioned processes in various embodiments of the application does not mean the order of execution, and the execution order of the processes should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the application.

[0045] Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the application.

[0046] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working process of the above-mentioned devices, apparatuses and units can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0047] In several embodiments provided by the application, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the above-described device embodiments are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed objects can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0048] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, i.e., may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0049] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit.

[0050] If the functions are realized in the form of software functional units 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 or the part of the present application that essentially contributes to the prior art or the 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 a plurality of 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 method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0051] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A cloud computing remote computing power scheduling system, characterized in that, The system comprises: The node access confirmation module obtains remote computing power node information, and if the link verification value is consistent with the corresponding node identifier in the whitelist, and the cluster attribution identifier is not the isolation cluster code, the node registration identifier is written into the computing power scheduling registration table to generate a remote node registration result set; The path stability elimination module determines whether the path hop value exceeds the upper limit threshold of the hop value and whether the packet loss rate exceeds the critical value of the packet loss rate according to the node registration identifier in the remote node registration result set, and eliminates the corresponding path when both conditions are met to generate an available node access path set; The resource node matching module extracts the node capability parameters according to the node registration identifier in the available node access path set, extracts the task demand parameters of the task to be dispatched at the same time, calculates and selects the node with the highest matching score of the task and the node to generate a task target node mapping set; The task scheduling construction module calls the node registration identifier in the task target node mapping set, and if the average response time is lower than the set response time threshold, and the scheduling channel occupancy rate is lower than the available limit, the current task is mapped to the target node scheduling pool to generate a dispatched task scheduling record. 2.The cloud computing remote computing scheduling system of claim 1, wherein, The remote node registration result set comprises an access node identifier number, a scheduling authority level code, an activation state identifier and a computing power attribution category, the available node access path set comprises a path stability identifier, a remaining link bandwidth, a transmission success ratio and a path reliability level, the task target node mapping set comprises a node preferred index, a task mapping level value, a node resource allocation relationship and a node selection priority sequence, and the dispatched task scheduling record specifically comprises a task dispatched node number, a scheduling execution period value, a scheduling channel mapping record and a cache writing marker set. 3.The cloud computing remote computing scheduling system of claim 2, wherein, The node access confirmation module comprises: The node information acquisition submodule obtains remote computing power node information, and the node information comprises a node registration identifier, a cluster attribution identifier and an access link verification value, the three items of information are combined and encoded with an acquisition timestamp by calling a node management interface, the node identity uniqueness is judged based on the encoding result, the node registration identifier meeting the uniqueness condition is recorded to a node temporary cache area to generate valid node identification information; The link validity discrimination submodule extracts the access link verification value from the node temporary cache area and compares it with each channel identifier in the access channel whitelist according to the node registration identifier associated with the valid node identification information, and when the verification value is consistent with any channel identifier in the whitelist and the cluster attribution identifier of the node is not equal to the isolation cluster code, the corresponding node identifier is written into the link validity list to generate a whitelist hit comparison result; The node registration writing submodule calls the link validity list content corresponding to the whitelist hit comparison result, extracts the node registration identifier and combines the access IP address of the node in the current period, the connection stability duration and the data interaction success rate to determine whether the registration identifier is in the unregistered state in the node management table, and if the condition is met, the registration identifier is written into the registration record to generate a remote node registration result set.

4. The cloud computing remote computing scheduling system of claim 3, wherein, The path stability elimination module comprises: The link parameter collection submodule obtains the node registration identifier in the remote node registration result set, extracts the path hop count of each node to the scheduling hub, the number of transmitted data packets in the last three rounds of transmission and the corresponding lost data packet count, associates the node identifier with the three link parameters, establishes a structured link data set, and generates a link quality monitoring data set; The path anomaly discrimination submodule discriminates the node registration identifier in the link quality monitoring data set according to the hop count value and the packet loss rate corresponding to each link, calls the set hop count upper threshold and the packet loss rate critical threshold, filters the node for which the path hop count exceeds the hop count upper threshold and the packet loss rate exceeds the packet loss rate critical threshold, and marks the node as an abnormal link path to generate a path elimination judgment result; The available path screening submodule calls the node registration identifier in the path elimination judgment result that is not marked as an abnormal link, screens out the elimination identifier from the original link record, sorts the remaining node path information, aggregates the available access path number and the corresponding path parameter according to the node dimension, and generates an available node access path set. 5.The cloud computing remote computing scheduling system of claim 4, wherein, The resource node matching module comprises: The parameter extraction submodule obtains the node registration identifier in the available node access path set, collects the node capability parameters of the corresponding node, including the maximum number of concurrent threads of the GPU, the available number of CPU cores, and the remaining capacity of the node local buffer, simultaneously extracts the task demand parameters of the task to be dispatched, including the calculation intensity index, the thread concurrency requirement, and the task cache demand, establishes a parameter association mapping according to the node and task binding relationship, and generates a node task parameter association data set; The level mapping submodule uses the node capability parameters and the task demand parameters to calculate the matching score of the task in the current node according to the node task parameter association data set, sorts the matching scores of the nodes in descending order, eliminates the nodes with a matching score lower than the minimum acceptable score, and generates a matchable node score sequence; The score screening submodule extracts the node registration identifier with the highest score from the matchable node score sequence according to the matching score of each node in the matchable node score sequence, constructs a list of matching successful nodes, establishes a one-to-one binding relationship data structure between the task and the node according to the list, and generates a task target node mapping set.

6. The cloud computing remote computing scheduling system of claim 5, wherein, The formula for calculating the matching score of the task in the current node is: ； wherein, is a matching score, is a normalized value of the number of available CPU cores, is a normalized value of the computation intensity index, is a normalized value of the maximum number of concurrent threads of the GPU, is a normalized value of the thread concurrency requirement, is a normalized value of the remaining capacity of the local buffer, is a normalized value of the task cache requirement, is a task priority level parameter.

7. The cloud computing remote computing scheduling system of claim 6, wherein, The task scheduling construction module comprises: The scheduling parameter extraction submodule calls the node registration identifier in the task target node mapping set, obtains the scheduling task pool capacity, the scheduling channel occupancy rate and the average response time corresponding to each node, binds the node registration identifier with the three indexes, forms a mapping structure, generates a node scheduling state information group, and records the node registration identifier that meets the two conditions. The channel availability discrimination submodule discriminates whether the average response time is less than the set response time threshold and whether the scheduling channel occupancy rate is less than the available occupancy limit threshold according to the average response time and the scheduling channel occupancy rate of each node in the node scheduling state information group, records the node registration identifier that meets the two conditions, and generates an available node screening result set. The task mapping binding submodule calls the node registration identifier in the available node screening result set, extracts the to-be-scheduled task number and establishes a task binding comparison table with the node registration identifier, writes the mapping information into the task dispatching scheduling pool, and generates a dispatched task scheduling record. 8.The cloud computing remote computing scheduling system of claim 7, wherein, The system further comprises: The state difference synchronization module calculates a state synchronization deviation degree value according to the target node registration identifier in the dispatched task scheduling record, compares the state synchronization deviation degree value with a set state synchronization threshold value, and if the state synchronization deviation degree value is greater than the threshold value, updates the current state record to the scheduling control state table, and generates a synchronization node running state set. The synchronization node running state set specifically refers to a node state update label, a task processing progress code, a node state drift intensity value and a synchronization activation timestamp. 9.The cloud computing remote computing scheduling system of claim 8, wherein, The state difference synchronization module comprises: A state data extraction submodule obtains the target node registration identifier in the dispatched task scheduling record, extracts three indexes of CPU usage rate, GPU active thread number and scheduling queue length in the current scheduling period, extracts the corresponding three indexes of the same node in the last scheduling period, constructs a node period state comparison table, and generates a period state comparison data group; A deviation value calculation submodule calculates a state synchronization deviation value of the current node according to the state indexes of the same node in the two periods in the period state comparison data group, compares the deviation value with a state synchronization triggering threshold value, records the node identifier greater than the threshold value, and generates a state deviation determination result set; A synchronization condition judgment submodule extracts the current index value of the node meeting the state synchronization condition from the original period state comparison data group according to the node identifier in the state deviation determination result set, writes the node identifier and the corresponding state into the scheduling control state table, establishes a synchronization record data, and generates a synchronization node running state set.

10. The cloud computing remote computing scheduling system of claim 9, wherein, The formula for calculating the state synchronization deviation value of the current node is specifically: ； wherein, represents a state synchronization deviation value of a node, represents a CPU usage rate in a current period, represents a CPU usage rate in a previous period, represents a period average of a CPU usage rate, represents a number of active GPU threads in a current period, represents a number of active GPU threads in a previous period, represents a period average of a number of active GPU threads, represents a scheduling queue length in a current period, represents a scheduling queue length in a previous period, represents a period average of a scheduling queue length.

Images