Multi-cloud heterogeneous computing power scheduling method and system

Abstract

The invention belongs to the technical field of cloud, and particularly relates to a multi-cloud heterogeneous computing power scheduling method and system, and the method comprises the steps: 1, describing the distribution of each node resource along with time evolution through employing a partial differential equation of the node resources; 2, comparing the macroscopic available resource vector with a measured task time-varying demand to form a feasibility constraint; 3, constructing a comprehensive optimization objective function; the objectives of the comprehensive optimization objective function comprise resource mismatch, delay and energy consumption; and 4, constructing a KKT system by using a primal-dual interior point method, and solving an overall problem of a partial differential equation containing node resources in combination with a large nonlinear optimization problem to obtain a solving result. According to the method, the dynamic evolution and spatial distribution of resources can be accurately described, the time-varying property and uncertainty of task requirements are reasonably dealt with, the risk in the resource scheduling process is effectively controlled, and global optimization is achieved in the aspects of resource mismatching, delay, energy consumption and the like.

Description

Technical Field

[0001] The present invention belongs to the field of cloud technology, and specifically relates to a multi-cloud heterogeneous computing power scheduling method and system. Background Art

[0002] With the rapid development of cloud computing technology, the demand for computing resources has gradually shown explosive growth. Cloud computing not only provides enterprises and individuals with efficient and flexible computing capabilities, but also promotes the development of emerging technologies such as big data, artificial intelligence, and the Internet of Things. In practical applications, cloud computing platforms are usually composed of multiple data centers, and each data center contains a large number of heterogeneous nodes. The types and capabilities of resources between nodes often vary significantly. These resources include but are not limited to hardware facilities such as CPU, GPU, and memory bandwidth. In order to meet complex application requirements, users need to assign tasks to these nodes for execution, and how to efficiently schedule resources and assign tasks has become a key technical problem that needs to be solved in the current cloud computing field.

[0003] At present, there are many studies and technologies on cloud computing resource scheduling, most of which are based on static resource description, heuristic scheduling algorithm and linear programming model. However, in the actual multi-cloud heterogeneous environment, these traditional methods have certain limitations, which are mainly manifested in the following aspects: First, the resource modeling method in the existing technology is too simplified and cannot truly reflect the dynamic distribution and topological characteristics of the resources inside the node. In the traditional resource scheduling system, hardware resources such as CPU, GPU and memory are usually abstracted as discrete fixed units. For example, CPU resources are described as a certain number of cores, GPU resources are represented as several computing units, and memory resources are regarded as a linear space with fixed capacity. Although this discretized resource modeling method is easy to implement, it seriously ignores the dynamic evolution characteristics of the resources inside the node. In the actual operation process, node resources have significant temporal dynamics and spatial inhomogeneity, such as the migration of CPU core load between different tasks, the uneven occupation of GPU memory due to the distribution of computing tasks, and local congestion of memory bandwidth due to differences in task access patterns. If the traditional discretization method is used, these subtle resource changes cannot be accurately captured, resulting in resource waste or decreased task execution performance in the scheduling system during resource allocation. Secondly, the existing technology is relatively simple in modeling task requirements and fails to fully consider the time-varying and uncertain nature of tasks. Most existing scheduling methods assume that the resource requirements of tasks are deterministic or static, such as the CPU, GPU, and memory resources required by tasks throughout the execution cycle are fixed. However, in actual application scenarios, the resource requirements of tasks often change over time. For example, some data analysis tasks require a large amount of CPU resources for data preprocessing in the initial stage, and may rely more on GPUs for parallel computing in subsequent stages. In addition, there is significant uncertainty in task requirements, such as fluctuations in the scale of data input, load changes, and the impact of the external system environment, which may lead to real-time changes in task resource requirements. The existing technology cannot effectively model this time-varying and uncertain nature, resulting in the inability of the scheduling system to make timely and reasonable adjustments when resource supply and demand fluctuate, thereby affecting the stability of task execution and the overall performance of the system. Summary of the invention

[0004] In view of this, the main purpose of the present invention is to provide a multi-cloud heterogeneous computing power scheduling method and system, which can accurately describe the dynamic evolution and spatial distribution of resources, reasonably respond to the time-varying and uncertainty of task requirements, effectively control the risks in the resource scheduling process, and achieve global optimization in terms of resource mismatch, delay and energy consumption.

[0005] The technical solution adopted by the present invention is as follows: A multi-cloud heterogeneous computing power scheduling method, the method comprising: Step 1: Use the partial differential equation of node resources to describe the distribution of each node resource over time; solve the partial differential equation of node resources and obtain the macroscopic available resource vector through spatial integration; Step 2: Compare the macro-available resource vector with the measured time-varying requirements of the task to form feasibility constraints; introduce uncertainty scenarios, sample scenarios for the time-varying requirements of the task, and use conditional risk value to control risks; Step 3: Construct a comprehensive optimization objective function; the objectives of the comprehensive optimization objective function include resource mismatch, delay and energy consumption; integrate all constraints into a large nonlinear optimization problem; the constraints include: equality constraints after discretization of partial differential equations of node resources, feasibility constraints and allocation constraints; Step 4: Use the primal-dual interior point method to construct the KKT system, and combine it with the large-scale nonlinear optimization problem to solve the overall problem of the partial differential equation containing node resources to obtain the solution result; when the solution result converges, discretize it according to the final solution result value to obtain the final scheduling plan.

[0006] Furthermore, let the task set be ;No. Tasks There is execution time ; The value range is 1 to ; is the total number of tasks; Representative tasks; let the multi-cloud collection be ; is the number of clouds in the multi-cloud set; For the cloud; Cloud There is a node collection ; The value range is 1 to ; For the Cloud The total number of nodes; Indicates Cloud No. nodes; Cloud The node resources in the feature space The continuous distribution, Indicates Cloud No. Nodes The hardware resource feature domain is obtained by mapping the CPU core topology, GPU array and memory channel into a continuous domain; the node resources include: CPU, GPU and MEM; CPU represents the processor occupancy rate, which ranges from 0 to 1; GPU represents the video memory occupancy rate, which ranges from 0 to 1; MEM represents the memory occupancy rate, which ranges from 0 to 1.

[0007] Furthermore, in step 1, the partial differential equation of node resources is used to describe the distribution of each node resource evolving over time through the following formula: ; ; ; in, is a resource vector field describing the nodes At feature space point Place and time Available node resources; is the coordinate point in the feature space; is a time variable; It is a positive definite symmetric diffusion tensor obtained by measuring and modeling the node hardware topology and resource scheduling mechanism; represents the spatial gradient operator, represents the divergence operator; is the dissipation rate function, which is a scalar function or a diagonal matrix function; It is an exogenous regulatory item; is the initial condition function, giving Available node resources, which are obtained through benchmark testing and system initial status detection; is the maximum execution time; Representation Node At feature space point Place and time Available processor usage at the time Representation Node At feature space point Place and time Available video memory usage at the time Representation Node At feature space point Place and time Available memory usage at time ; This is the transpose operation of a vector.

[0008] Furthermore, in step 1, the partial differential equation of the node resources is solved by the following formula and then the macroscopic available resource vector is obtained by spatial integration: ; in, is the macro available resource vector; The feature space The coordinate point integration variable in .

[0009] Furthermore, in step 1, the macro available resource vector is compared with the measured task time-varying demand through the following formula, and in step 2, the feasibility constraint is formed through the following formula: ; in, Representation Node In time The total amount of processor occupancy available for assignment to tasks; Representation Node In time The total amount of video memory available for allocation; Representation Node In time The total amount of memory available for allocation; is the resource requirement vector, representing the task In time Requirements for processor occupancy, video memory occupancy, and memory occupancy; For the task In time Required processor resource utilization; For tasks in time Required video memory occupancy; For tasks in time Required memory usage.

[0010] Furthermore, in step 1, the following formula is used, and in step 2, the following formula is used to introduce uncertainty scenarios, sample scenarios for time-varying task requirements, and use conditional risk value to control risks: ; in, Describe the worst case scenario under uncertainty -Percentile expected loss; different scenarios under uncertainty Represents possible implementation methods of task requirements; It is the cost incurred if resources are insufficient or mismatched in a given scenario; ); Therefore, when resources are insufficient to meet the scenario When the demand is will become larger; is the L2 norm; It is an auxiliary optimization variable introduced in the definition of conditional risk value, which is used to linearize the definition of conditional risk value and is defined as a quantile or critical loss level; when Exceed hour, Taking into account excess losses, the definition of conditional value at risk is ensured to be an optimizable convex problem; is the confidence level used to determine the proportion of the worst-case scenario being considered; is the total number of scenes.

[0011] Furthermore, in step 3, a comprehensive optimization objective function is constructed by the following formula: ; in, is a binary decision variable. Assign to Node hour, ; otherwise 0; It's in the scene Next, the task At the initial moment of the task The resource demand vector of represents the L1 norm; Is a node The delay factor of Is a node The average available bandwidth to the data source; For Node Specific node resources Energy consumption coefficient; Includes: CPU, GPU and MEM; Is a node In time For node resources The macro available resources vector.

[0012] Furthermore, in step 3, all constraints are integrated into a large nonlinear optimization problem through the following process: The equality constraints after discretization of the partial differential equation of the node resources are expressed by the following formula: ; in, for The discretization result is that the step size is used in space or meshing, using time steps in time To divide, represents the time step index; Is a node Approximate solution of resource vector fields at discrete grid points and discrete time levels; is the right-hand side of the discretized partial differential equation of node resources; The coefficient matrix obtained after discretization of the partial differential equation of node resources; the allocation constraint is expressed as: ; The feasibility constraint is expressed as: ; in, Yes Yes Node In time For node resources The macro-available resource vector; Because it is in the scene Next, the task On task at time The resource demand vector.

[0013] Furthermore, in step 4, the KKT system is constructed using the primal-dual interior point method through the following process to solve the overall problem of the partial differential equation containing node resources and obtain the solution result: Construct the Lagrangian function The comprehensive optimization objective function, the partial differential equation of node resources, and the large nonlinear optimization problem that integrates all constraints are integrated to obtain the following formula: ; in, is the dual pairing operator; is the barrier function, which is used to ensure The feasible region does not deviate from during the solution process of the primal-dual interior point method; ; Due to the final need It is a 0-1 integer decision, but it is too difficult to solve the 0-1 problem directly. , in the initial stage of the solution, it is allowed In the interval (0,1), as the barrier parameter decreases, It will force the solution to approach 0 or 1, so that the primal-dual interior point method approaches the binary solution on the basis of continuous relaxation; is the dual variable of the allocation constraint, which requires that each task must be assigned to exactly one node. It measures the sensitivity of relaxing the allocation constraints to the Lagrangian function near the optimal solution; It is the dual variable of the equality constraint after the partial differential equation of the node resources is discretized, and is used to measure the influence of the equality constraint after the partial differential equation of the node resources is discretized on the Lagrangian function if it is relaxed or tightened; for The dual variable of If the resource constraint of a node is tight, that is, it just satisfies the equation, will become larger, indicating that the node resource has become a bottleneck; if there are surplus node resources, is 0, indicating that the constraint is not a tight constraint problem; the KKT condition is solved iteratively by the primal-dual interior point method, and the following process is performed in each iteration: using the current Approximate Update ; Use Newton's method to solve the linearized KKT system to find the optimal , and ; When making When the value of is the largest, stop the iteration and get The final value of ;right Perform a threshold determination, and if it is greater than the set threshold, it is assigned to the node.

[0014] A multi-cloud heterogeneous computing power scheduling system, the system comprising: an available resource calculation unit, used to use the partial differential equation of node resources to describe the distribution of each node resource evolving over time; after solving the partial differential equation of node resources, a macroscopic available resource vector is obtained by spatial integration; a scenario and constraint construction unit, used to compare the macroscopic available resource vector with the measured time-varying demand of tasks to form a feasibility constraint; introduce uncertainty scenarios, sample scenarios for the time-varying demand of tasks and use conditional risk value to control risks; an optimization target construction unit, used to construct a comprehensive optimization objective function; the objectives of the comprehensive optimization objective function include resource mismatch, delay and energy consumption; all constraints are integrated into a large nonlinear optimization problem; the constraints include: equality constraints, feasibility constraints and allocation constraints after the partial differential equation of node resources is discretized; a scheduling unit, used to construct a KKT system using the primal-dual interior point method, and solve the overall problem of the partial differential equation containing node resources in combination with the large nonlinear optimization problem to obtain a solution result; when the solution result converges, discretization is performed according to the value of the final solution result to obtain a final scheduling plan.

[0015] By adopting the above technical scheme, the present invention produces the following beneficial effects: the present invention realizes a refined description of the dynamic evolution and spatial distribution of node resources through partial differential equation (PDE) modeling technology. This is a height that the existing technology cannot reach, because in the traditional method, the node resources are usually simplified into a fixed number of discrete units, which cannot reflect the complex topological structure of resources in space and the dynamic changes in the time dimension. In the present invention, the resource state of each node is modeled by a continuous vector field, which contains the availability distribution of CPU, GPU and memory resources. By introducing diffusion terms, dissipation terms and exogenous control terms, the distribution and evolution of node resources in space and time dimensions are accurately characterized. For example, the diffusion term reflects the migration and balance process of resources within the node, the dissipation term describes the natural attenuation of resources, and the exogenous control term reflects the influence of the management system on the scheduling and adjustment of resources. This continuous domain modeling method makes the dynamic state of node resources no longer a rough static description, but an accurate expression with physical meaning, providing real and quantifiable data support for subsequent resource scheduling optimization. In this way, resource allocation no longer relies solely on rough estimates, but is based on the real-time dynamics of node internal resources, achieving accurate matching of resource supply and demand. Secondly, the present invention solves the defects of overly simple and lack of adaptability of task demand modeling in the prior art by introducing time-varying demand and uncertainty scenario modeling. In an actual cloud computing environment, the resource demand of a task often changes over time and has uncertainty. The present invention converts the task demand into a series of representative scenarios through scenario sampling technology, and each scenario corresponds to a specific task demand curve. This method can not only cover a variety of possible situations of task demand, but also dynamically adjust according to the actual fluctuation of demand. In addition, in order to cope with the impact of task demand uncertainty on system performance, the present invention introduces the conditional value at risk (CVaR) as an advanced risk control indicator. CVaR not only focuses on the satisfaction of the resource demand of the task under normal circumstances, but also quantifies the risk of resource mismatch in the worst case, so as to effectively control the risk during the scheduling process. This risk control mechanism effectively makes up for the shortcomings of traditional methods that are prone to performance degradation or task failure in the face of uncertainty, so that the present invention can still provide a reliable scheduling solution in a highly uncertain environment. The present invention realizes multi-objective global optimization of resource mismatch, delay and energy consumption by constructing a comprehensive optimization objective function. In a multi-cloud heterogeneous environment, task scheduling needs to consider multiple performance indicators at the same time, including the precise matching of resource supply and demand, the response time of task execution, and the minimization of system energy consumption. Traditional methods can usually only optimize a certain indicator alone, while ignoring the balance between other indicators, resulting in poor system performance. The present invention integrates resource mismatch cost, data transmission delay cost and resource energy consumption cost into the optimization framework by constructing a comprehensive objective function.For example, the resource mismatch cost is quantified by calculating the deviation between node resources and task requirements, the delay cost is modeled in combination with the data transmission bandwidth and the task initialization time, and the energy consumption cost is quantified according to the resource occupancy rate and the node power consumption model. This multi-objective comprehensive optimization can not only achieve the best matching of resource supply and demand, but also minimize the response delay and energy consumption of the system, thereby significantly improving the overall performance of the scheduling scheme. The present invention adopts the original dual interior point method to solve large-scale nonlinear optimization problems, effectively breaking through the bottleneck of solving under high-complexity nonlinear constraints. Since the task scheduling problem involves the dynamic evolution of resources, time-varying demand and risk control, there is a high degree of nonlinear coupling between the objective function and the constraints, and traditional linear programming or heuristic algorithms are difficult to solve effectively. The present invention introduces the original dual interior point method, constructs the Lagrangian function and solves the KKT system, realizing the alternating optimization of the original variable and the dual variable. In the solution process, the task allocation variables are optimized by stepwise approximation, and the allocation variables are continuously relaxed by the barrier function, and finally the continuous solution is discretized into an executable scheduling scheme. This solution method can not only find a solution close to the global optimal solution under high-complexity constraints, but also maintain high computing efficiency in large-scale task and resource scenarios, ensuring the real-time and executable nature of the scheduling solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A method flow chart of a multi-cloud heterogeneous computing power scheduling method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] All features disclosed in this specification, or steps in all methods or processes disclosed, except mutually exclusive features and / or steps, can be combined in any manner.

[0018] Any feature disclosed in this specification (including any additional claims and abstract), unless otherwise stated, may be replaced by other alternative features that are equivalent or have similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0019] Example 1: Reference Figure 1 , a multi-cloud heterogeneous computing power scheduling method, the method comprising: Step 1: Use the partial differential equation of node resources to describe the distribution of each node resource over time; solve the partial differential equation of node resources and obtain the macroscopic available resource vector through spatial integration; The core principle of step 1 is to use partial differential equations (PDEs) to accurately describe the dynamic evolution of each node's resources in the time dimension and the distribution in the space dimension, so as to obtain accurate quantification of the node resource status, and then provide accurate data support for subsequent scheduling decisions. In a multi-cloud heterogeneous environment, different nodes have their own hardware configurations and resource characteristics, such as the distribution of CPU cores, the deployment of GPU computing units, the transmission capacity of memory channels, etc. These resources are not static, but evolve dynamically with the passage of time and changes in task load. Traditional resource description methods mostly use discrete static models, which simply summarize the node's computing resources into a fixed number of available CPUs, GPUs, and memory bandwidth parameters, ignoring the uneven distribution of resources within the node and the time evolution of resources. It is difficult to reflect the actual node resource utilization, and it is even more unable to adapt to highly complex computing power scheduling requirements. The present invention introduces partial differential equations as a mathematical modeling tool, innovatively abstracting the internal resource state of the node into a dynamic field that changes with time and is continuously distributed in space. By establishing PDE to characterize the resource state and considering the diffusion, dissipation and external control factors of resources based on physical and statistical principles, the spatiotemporal change relationship of node resources can be rigorously described in the continuous domain.

[0020] Specifically, in a multi-cloud environment, each node is in a complex resource load and management state, which makes the resource availability uncertain in time and uneven in space. For example, due to the migration of computing loads between CPU cores within a node, there may be a phenomenon of computing resources diffusing to high-load areas; GPU units may also have different resource allocations due to different interconnection bandwidths; and the scheduling of memory controllers may cause local bandwidth congestion due to the access patterns of different tasks. In this case, the dynamic distribution of node resources can be modeled by partial differential equations. PDEs characterize the evolution of resources over time and the migration law of space through continuous spatial gradients and time derivatives. Specifically, the diffusion term in PDE is used to describe the balance process of resources within the node. When the resource occupancy in a local area is too high, the system will naturally diffuse to the low occupancy area, thereby achieving dynamic balance of resources within the node. The dissipation term is used to describe the natural loss of resources due to factors such as system maintenance threads, hardware power consumption, and frequency attenuation, thereby reflecting the gradual attenuation of the actual availability of resources in the time dimension. At the same time, the exogenous control item represents the external input of the system manager to actively schedule resources at a specific time, such as dynamically increasing the CPU core frequency, activating the GPU computing unit, or adjusting the memory access priority. This part is controlled by the system management policy to achieve active intervention in node resources. Through this PDE modeling method, the spatiotemporal dynamics of node resources can be numerically solved to obtain the complete evolution process. In practical applications, due to the continuity of PDE, the internal resource state of the node can be described in a high-precision feature space, thus getting rid of the limitations of the rough description of resources in traditional methods. Then, in order to convert the resource distribution in the continuous space into a macro-indicator available for scheduling, the solution of PDE is integrated in the node feature space to obtain the resource availability vector of the node as a whole. This resource vector covers the total available CPU computing power, GPU computing power and memory bandwidth of the node at a certain point in time, reflecting the overall computing power level of the node, and has the operability to directly match the task requirements. Through this method, the present invention realizes a refined description of node resources, which not only captures the law of dynamic evolution of resources in the time dimension, but also reveals the unevenness and diffusion characteristics of resource distribution in the spatial dimension, providing accurate data support for multi-cloud heterogeneous computing power scheduling, and can effectively improve the rationality of scheduling decisions and resource utilization.

[0021] Step 2: Compare the macro-available resource vector with the measured time-varying requirements of the task to form feasibility constraints; introduce uncertainty scenarios, sample scenarios for the time-varying requirements of the task, and use conditional risk value to control risks; Step 2 compares the node macro-available resource vector with the time-varying demand of the task to establish the feasibility constraint of resource allocation. Combined with the uncertainty of task demand, the risk is quantified and controlled by introducing scenario sampling and conditional value at risk (CVaR) to ensure efficient scheduling and robustness of tasks in complex multi-cloud heterogeneous environments. In a multi-cloud environment, the execution requirements of tasks often change dynamically over time. For example, a task may have high CPU resource requirements in a certain period of time, while the demand for GPU or memory resources rises sharply in another period of time. At the same time, due to the fluctuation of the task's own operating load, input data scale or external system environment, task requirements are often uncertain. This uncertainty is very common in actual scenarios. Traditional computing power scheduling methods usually assume that task requirements are certain, or simply take the average value of task requirements as the basis. This processing method cannot truly reflect the actual resource requirements of the task, which can easily lead to mismatch in resource allocation, affect the overall system performance, and even cause task failure during peak demand. The present invention introduces a scenario sampling method to convert the time-varying requirements of the task into multiple representative scenarios. Each scenario reflects the specific demand distribution of the task under certain uncertain conditions, thereby providing a more refined reference basis for task resource allocation.

[0022] In the specific operation, step 2 first compares the macroscopic available resource vector obtained by solving the partial differential equation in step 1 and spatial integration with the task demand vector to form a feasibility constraint. The essence of this process is to check whether the node resources can meet the time-varying requirements of the task within the task execution period to ensure the feasibility of task scheduling. If the node resources are understood as a time series vector, then the task demand is also a time series vector. The two need to be compared one by one at each time point to ensure that the resource requirements of the task at all times during the execution cycle can be met. This resource and demand matching mechanism enables the scheduling method of the present invention to accurately control the relationship between resource supply and demand in the time dimension, avoiding the problem of insufficient resources during task execution. In traditional technologies, since the task demand is simplified to a fixed value or an average value, it is difficult to achieve dynamic matching of time-varying demands, resulting in a greatly reduced accuracy of resource allocation. However, in a multi-cloud heterogeneous environment, due to the uncertainty of task demand, it is obviously not enough to rely solely on a single task demand vector for scheduling. The uncertainty of task demand may come from many factors, such as the complexity of the task itself, the randomness of input data, the fluctuation of external load, and the instability of the execution environment. This uncertainty often causes actual task requirements to deviate from the estimated values. If this deviation is not taken into account, the scheduling solution is prone to performance degradation or even task failure when facing extreme demand scenarios. Therefore, in order to deal with this uncertainty, step 2 further introduces a scenario sampling method, which generates a set of demand scenarios representing different possible situations by scenario-based modeling of task requirements. Each scenario corresponds to a specific task demand curve, which describes the resource requirements of the task under different uncertain conditions. This scenario sampling can not only cover a variety of possible situations of task requirements, but also reflect the probability of occurrence of different scenarios through a certain probability distribution, thereby providing data support for subsequent risk control.

[0023] On this basis, step 2 uses the risk control indicator of conditional value at risk (CVaR) to quantify and optimize the uncertainty of task requirements. CVaR is an advanced risk measurement method that not only focuses on the overall impact of the uncertainty of task requirements on resource allocation, but also emphasizes the impact of the worst-case task requirements on system performance. Specifically, CVaR calculates the worst-case scenario of the task. -The average loss value in the per-quantile scenario, so as to accurately assess the risk of task requirements under extreme conditions. The core idea of ​​this process is that even in the most unfavorable scenario, the system can still ensure the normal execution of the task and avoid task failure or system crash due to insufficient resources. Compared with the traditional average risk measurement method, CVaR can more effectively control the risks in extreme scenarios and improve the robustness and stability of the scheduling scheme. Through the combination of scenario sampling and CVaR risk control, step 2 can not only comprehensively model the uncertainty of task requirements, but also accurately constrain the resource requirements of tasks in the worst case, so that the scheduling scheme can maintain high reliability and performance in a multi-cloud heterogeneous environment. This method breaks through the limitations of the prior art on the assumption of deterministic task requirements, and solves the problem that traditional scheduling methods are prone to performance degradation when facing uncertain requirements. By introducing uncertain scenarios and risk control, the present invention can effectively balance the performance and risk of resource allocation, ensuring that the task scheduling scheme has high execution efficiency and resource utilization in various possible demand scenarios. In addition, this method can flexibly adjust the confidence level parameters of CVaR according to the risk tolerance and service level agreement (SLA) requirements of different tasks, thereby achieving a more refined trade-off between performance and risk.

[0024] Step 3: Construct a comprehensive optimization objective function; the objectives of the comprehensive optimization objective function include resource mismatch, delay and energy consumption; integrate all constraints into a large nonlinear optimization problem; the constraints include: equality constraints after discretization of partial differential equations of node resources, feasibility constraints and allocation constraints; In an actual multi-cloud heterogeneous environment, there are significant differences in the hardware configuration and resource capabilities of different nodes. At the same time, the demand for resources during task execution is also time-varying and uncertain, which leads to the scheduling system having to balance and optimize between resource allocation, task execution delay, and system energy consumption. There is a complex coupling relationship between the three. Resource mismatch directly affects the execution efficiency of tasks, delay affects the response time of tasks, and energy consumption is directly related to the operating cost and energy efficiency of the system. Therefore, the key to constructing a comprehensive optimization objective function is to quantify and uniformly model these three core indicators so that they can be solved within the same optimization framework, thereby achieving global optimization of computing resource scheduling in multiple dimensions. Specifically, resource mismatch refers to the difference between the resources that a node can provide and the resources required for a task. This difference may lead to waste or insufficiency of resources during task execution, directly affecting the efficiency of task execution and the utilization rate of the overall system resources. In the present invention, the quantification of resource mismatch measures the deviation between the availability of node resources and task requirements to achieve precise control of resource supply and demand matching. By comparing the supply and demand of resources in the time dimension point by point, this deviation is incorporated into the objective function in a mathematically quantified form, thereby minimizing the negative impact of resource mismatch during the optimization process and ensuring that the task can fully utilize node resources while meeting demand.

[0025] On the other hand, delay is one of the performance indicators that need to be considered in the task scheduling process in a multi-cloud environment, especially in terms of the time overhead required for data transmission, task initialization, and resource allocation decisions. Network delays between nodes, bandwidth limitations, and differences in resource initialization of heterogeneous nodes will have a significant impact on the overall execution time of the task. In the present invention, the delay cost is quantified by combining the initial resource requirements of the task, the network bandwidth of the node, and the delay characteristics between the nodes, and is converted into an optimizable objective function term, so that the delay cost is minimized through a reasonable task allocation strategy during the solution process, and the response speed of the task and the overall execution efficiency of the system are improved. In addition, energy consumption optimization is one of the important components of the objective function of the present invention, mainly considering the relationship between the actual energy consumption of node resources and resource utilization. In a multi-cloud heterogeneous environment, different types of hardware resources (such as CPU, GPU, memory, etc.) have different energy consumption characteristics under different load conditions, and the actual occupancy rate of the task to the resource is positively correlated with the energy consumption of the node. In the present invention, the resource occupancy rate of the node is associated with the actual power consumption through the energy consumption coefficient, so that the overall energy consumption of the system is minimized as one of the goals during the optimization process. This not only helps to reduce the operating cost of the system, but also can improve energy utilization efficiency to a certain extent, which is in line with the current technical trends of green computing and sustainable development. In addition to the construction of the above objective function, the present invention also integrates the constraints involved in step 1 and step 2 into the same large nonlinear optimization problem, including the equality constraints after the discretization of the partial differential equations of node resources, the resource feasibility constraints of the tasks, and the task allocation completeness constraints. The introduction of these constraints ensures the physical feasibility and practical operability of the optimization results. For example, the PDE discretization constraints of node resources ensure that the continuity of the resource evolution process is consistent with the real physical laws, so that the optimization results can reflect the actual distribution and available status of resources. At the same time, the task allocation completeness constraint ensures that each task can be assigned to a suitable node for execution, avoiding the problem of missing or repeated assignment of tasks, while the resource feasibility constraint ensures that the actual available resources of the node can meet the needs of the task, avoiding the phenomenon of resource overload or shortage in the resource allocation process. By integrating these objective functions and constraints into a large nonlinear optimization problem, the present invention realizes the global optimization of multi-cloud heterogeneous computing power resource scheduling, and solves the problem of isolated treatment and conflict of each optimization target in the traditional scheduling method. This multi-objective unified modeling method can not only find the best balance between resource supply and demand matching, task delay and system energy consumption, but also ensure the overall efficiency and performance of the scheduling scheme through a global optimization solution. In addition, by introducing a nonlinear optimization method, the present invention enables the entire system to find a near-optimal scheduling scheme under highly complex constraints, thereby further improving the accuracy and computing efficiency of resource utilization in a multi-cloud heterogeneous environment.

[0026] Step 4: Use the primal-dual interior point method to construct the KKT system, and combine it with the large-scale nonlinear optimization problem to solve the overall problem of the partial differential equation containing node resources to obtain the solution result; when the solution result converges, discretize it according to the final solution result value to obtain the final scheduling plan.

[0027] The complexity of the multi-cloud heterogeneous computing power scheduling problem stems from the coupling of multiple factors, including the dynamic nature of node resources, the time-varying nature of task requirements, and the nonlinear matching relationship between resources and tasks, which makes it difficult for traditional linear programming or heuristic algorithms to effectively solve such problems. The primal-dual interior point method adopted in the present invention is an efficient algorithm for solving large-scale nonlinear constrained optimization problems. By introducing dual variables and barrier functions, the constraints of the original problem are integrated into the objective function to form a Lagrangian function, thereby converting the constrained optimization problem into an unconstrained problem for solution. The entire solution process is based on the iterative idea of ​​the primal-dual interior point method. By alternating optimization between the original variables and the dual variables and gradually reducing the obstacle parameters, the solution can be kept within the feasible domain in each iteration, and finally converges to the optimal solution that meets the original constraints. Specifically, the primal-dual interior point method first integrates the objective function, PDE constraints of node resources, resource feasibility constraints, and task allocation constraints in the nonlinear optimization problem into a unified framework by constructing a Lagrangian function. Dual variables are introduced into the Lagrangian function to represent the "shadow price" of the constraints. These dual variables can reflect the degree of influence of the constraints on the optimal solution, thereby guiding the adjustment direction of the original variables during the optimization process. In addition, in order to deal with the binary characteristics of the task allocation decision variables, the present invention uses a barrier function to relax the task allocation variables so that they are restricted to the continuous space of (0,1) during the solution process, and by gradually reducing the barrier parameters, it is ensured that the final solution can gradually approach the boundary of the binary decision. The introduction of the barrier function not only ensures the feasibility of the solution, but also effectively avoids the problem of infeasible solutions caused by overly tight constraints during the optimization process.

[0028] In the solution process, the primal-dual interior point method optimizes the primal and dual variables simultaneously by solving the KKT system (Karush-Kuhn-Tucker conditions). The KKT system is a mathematical framework that describes the optimality conditions of optimization problems, which includes the primal feasibility conditions, the dual feasibility conditions, and the complementary relaxation conditions. In each iteration, the algorithm obtains the update amount of the primal and dual variables by linearizing the KKT system and solving its Newton direction. In this process, the primal variables (including task allocation variables and PDE discrete solutions) are adjusted according to the optimization direction of the objective function, while the dual variables are updated according to the degree of relaxation of the constraints, so as to ensure that the solution gradually approaches the optimal target value under the premise of satisfying the constraints. At the same time, in order to ensure the stability and convergence of the iteration, the algorithm will perform step size control at each update, and ensure that the new solution is still within the feasible domain through line search to avoid destroying the feasibility of the constraints due to excessive step size. An important feature of the primal-dual interior point method is that it can efficiently handle large-scale nonlinear constraint problems in continuous space and find solutions close to the global optimal by stepwise approximation. In the present invention, the PDE constraints of node resources are discretized into a set of linear equations, which describe the dynamic distribution and evolution of node resources in time and space, while resource feasibility constraints and task allocation constraints ensure the feasibility and integrity of the scheduling scheme. In the solution process, the original dual interior point method effectively integrates these constraints to achieve simultaneous optimization of multidimensional constraints, so that the entire scheduling scheme achieves the best balance between resource supply and demand matching, task delay minimization and system energy consumption optimization. When the algorithm converges to the vicinity of the optimal solution, the continuous solution of the task allocation variable will gradually approach 0 or 1. At this time, the task allocation variable is discretized by setting a threshold, that is, the solution close to 1 is regarded as a sign that the task is assigned to the corresponding node, and the solution close to 0 indicates that the task is not assigned to the node. Through this discretization operation, the present invention can convert the continuous relaxation solution into an actual executable task scheduling scheme, ensure that the scheduling scheme can be implemented in the actual system, and meet the matching requirements of task requirements and resource availability.

[0029] Example 2: Assume that the task set is ;No. Tasks There is execution time ; The value range is 1 to ; is the total number of tasks; Representative tasks; let the multi-cloud collection be ; is the number of clouds in the multi-cloud set; For the cloud; Cloud There is a node collection ; The value range is 1 to ; For the Cloud The total number of nodes; Indicates Cloud No. nodes; Cloud The node resources in the feature space The continuous distribution, Indicates Cloud No. Nodes The hardware resource feature domain is obtained by mapping the CPU core topology, GPU array and memory channel into a continuous domain; the node resources include: CPU, GPU and MEM; CPU represents the processor occupancy rate, which ranges from 0 to 1; GPU represents the video memory occupancy rate, which ranges from 0 to 1; MEM represents the memory occupancy rate, which ranges from 0 to 1.

[0030] Specifically, in traditional computing resource scheduling methods, CPU, GPU and memory resources are usually simply regarded as discrete units. For example, CPU resources are described as several independent cores, GPU is described as several computing units, and memory is described as fixed storage space. Although this discrete modeling method is simple to calculate, it cannot fully reflect the spatial topological structure of hardware resources and the dynamic distribution characteristics within the node. In an actual multi-cloud heterogeneous environment, the computing resources of the node are not static. They have significant spatial structure and dynamic migration characteristics, such as topological interconnection between CPU cores, parallel computing distribution of GPU arrays, and data transmission bandwidth of memory channels. These physical structures and spatial characteristics have a profound impact on the performance of task scheduling. If these factors are ignored, scheduling decisions may lead to resource mismatch and even system performance degradation. The mapping of CPU core topology refers to abstracting the core structure of the CPU in the node, including the number of cores, the interconnection topology between cores, the cache hierarchy and the frequency characteristics, into a dimension in a continuous domain. In traditional methods, the CPU core is described as a single independent entity, while in the present invention, through continuous mapping, the availability of CPU resources is regarded as a field quantity distributed continuously in space. For example, in a multi-core CPU, the cores are connected to the high-speed interconnection network through cache, and the load migration and computing scheduling of different cores will produce dynamic changes in resource occupancy. Through continuous domain modeling, the interconnection relationship between the cores and the resource distribution state can be described as a continuous function, thereby more accurately reflecting the occupancy and availability of CPU resources inside the node. Similar to the CPU, the mapping of the GPU array is to abstract the computing units, video memory access bandwidth and data transmission structure inside the GPU into another dimension in the continuous domain. In the GPU, there are parallel computing characteristics between computing units (such as CUDA cores or stream processors), but the load and video memory occupancy of different computing units are not uniform, especially when executing large-scale parallel tasks, resource occupancy has obvious spatial distribution characteristics. For example, some computing units may be in a high-load state, while other units are in an idle or low-load state. In the present invention, through the mapping of the continuous domain, the resource occupancy state of the GPU array is regarded as a continuously distributed vector field quantity, so that the system can model and optimize the distribution state of GPU resources through mathematical tools, thereby achieving more refined task scheduling. For the mapping of memory channels, the present invention reflects the dynamic distribution of memory resource occupancy and access bandwidth by abstracting the memory controller, number of channels, and data bandwidth characteristics into a continuous domain. In a multi-cloud heterogeneous environment, the access performance of memory channels has a key impact on the efficiency of task execution, especially in high data throughput tasks, where the bottleneck of memory bandwidth often becomes a limiting factor for system performance. In traditional methods, memory resources are usually simply regarded as fixed capacity or bandwidth, while ignoring the dynamic distribution of data access.In the present invention, by mapping the memory channel into a continuous domain, the memory occupancy state can be described as a continuous function in time and space, thereby revealing the actual usage state of memory resources and providing a more realistic memory performance evaluation for task scheduling.

[0031] Embodiment 3: In step 1, the partial differential equation of node resources is used to describe the distribution of each node resource evolving over time through the following formula: ; ; ; in, is a resource vector field describing the nodes At feature space point Place and time Available node resources; is the coordinate point in the feature space; is a time variable; It is a positive definite symmetric diffusion tensor obtained by measuring and modeling the node hardware topology and resource scheduling mechanism; represents the spatial gradient operator, represents the divergence operator; is the dissipation rate function, which is a scalar function or a diagonal matrix function; It is an exogenous regulatory item; is the initial condition function, giving Available node resources, which are obtained through benchmark testing and system initial status detection; is the maximum execution time; Representation Node At feature space point Place and time Available processor usage at the time Representation Node At feature space point Place and time Available video memory usage at the time Representation Node At feature space point Place and time Available memory usage at time ; This is the transpose operation of a vector.

[0032] Specifically, Contains the available ratio of CPU, GPU, MEM and other resources of the node at this location (feature space coordinate). For example, if (corresponding to CPU, GPU, and MEM respectively), it means that at the coordinate Place, time The CPU availability is 80%, the GPU availability is 50%, and the memory bandwidth availability is 90%. This reflects the dynamic distribution of resource availability of different hardware units in the node in time and space. Is a node The feature space is an abstract mapping of the hardware structure inside the node, which represents the CPU core topology position, GPU unit distribution, memory channels, etc. in a continuous space domain. Represents a point in the space, corresponding to a virtual "resource location" or "resource fragment" inside the node. As time goes on, the availability of node resources will change dynamically. For example, node cooling, turbo frequency mechanism, and task load changes will cause resource occupancy to change over time. Describe this dynamic process that evolves over time. Describes the ability of resources to diffuse in different feature space directions within a node. Just like the diffusion coefficient in heat conduction, resource availability can also be "diffused" or balanced within a node. For example, if there is a high-speed interconnect between CPU cores, the total availability can be evenly distributed in the adjacent area. It describes the resource balancing capability determined by the hardware interconnection structure. This tensor is obtained by measuring and modeling the node hardware topology and resource scheduling mechanism. Represents the rate of change of resource availability in space. It indicates the resource reallocation process caused by uneven spatial distribution. Resources will spread from high availability areas to low availability areas to achieve a certain balance. Describes the extent to which resources automatically weaken over time without external replenishment. Just like heat dissipates on its own, resource availability here will decrease without external maintenance. For example, the system's underlying maintenance process, hardware aging, and frequency reduction strategy will cause the actual available ratio of resources to decrease over time. This value can be obtained by fitting the power consumption and frequency change data of the node in the no-load state. Indicates active intervention of the hypervisor on node resources. For example, At any given moment, the dispatch system can (For example, a group of CPU cores) increase the frequency (equivalent to Inject positive resource increments into the node), or open more SM modules in some GPU units to increase availability. These external controls come from the automation policy controller, which is dynamically adjusted according to the current load, energy consumption strategy and SLA optimization goals. When the node is powered on, initialized or the task scheduling strategy is started, each feature space point is measured at this time. The available ratio of basic resources. For example, the CPU frequency of a newly started node may be at the baseline level (CPU availability is high), the GPU unit is not loaded when it is just activated (GPU availability is close to 1), and the memory bandwidth is not used intensively at the beginning. Obtained through benchmark testing and system initial status detection .

[0033] Embodiment 4: In step 1, the partial differential equation of the node resources is solved by the following formula and then the macroscopic available resource vector is obtained by spatial integration: ; in, is the macro available resource vector; The feature space The coordinate point integration variable in .

[0034] Specifically, Indicates that the node Internal feature space The resource utilization rate of the distribution field The overall resource availability vector is obtained after spatial integration. In other words, it aggregates the fine-grained resource distribution information within the node into a macro metric, allowing us to use a simple vector in the subsequent steps Indicates that the node is at time For example, Indicates at time The total amount of CPU, GPU, and MEM resources that the node can devote to task scheduling.

[0035] Embodiment 5: In step 1, the following formula is used, and in step 2, the macro available resource vector is compared with the measured task time-varying demand through the following formula to form a feasibility constraint: ; in, Representation Node In time The total amount of processor occupancy available for assignment to tasks; Representation Node In time The total amount of video memory available for allocation; Representation Node In time The total amount of memory available for allocation; is the resource requirement vector, representing the task In time Requirements for processor occupancy, video memory occupancy, and memory occupancy; For the task In time Required processor resource utilization; For tasks in time Required video memory occupancy; For tasks in time Required memory usage.

[0036] Specifically, For example, For the task In time Required CPU resource usage. For example, 0.6 means that the task requires the node to allocate 60% CPU availability. This requirement is usually collected and fitted by historical operation logs and performance analysis tools (such as perf, GPU profiler, memory bandwidth counter).

[0037] Example 6: In step 1, the following formula is used, and in step 2, the following formula is used to introduce uncertainty scenarios, sample scenarios for time-varying task requirements, and use conditional risk value to control risks: ; in, Describe the worst case scenario under uncertainty -Percentile expected loss; different scenarios under uncertainty Represents possible implementation methods of task requirements; It is the cost incurred if resources are insufficient or mismatched in a given scenario; ); Therefore, when resources are insufficient to meet the scenario When the demand is will become larger; is the L2 norm; It is an auxiliary optimization variable introduced in the definition of conditional risk value, which is used to linearize the definition of conditional risk value and is defined as a quantile or critical loss level; when Exceed hour, Taking into account excess losses, the definition of conditional value at risk is ensured to be an optimizable convex problem; is the confidence level used to determine the proportion of the worst-case scenario being considered; is the total number of scenes.

[0038] Specifically, It is an indicator function of conditional value-at-risk. It measures the worst case scenario. - Average loss per quantile (i.e., in the most adverse scenario exceeding a certain loss level). When considering the resource misallocation loss of tasks under uncertain scenarios hour, Indicate the situations where the loss is the most serious (probability mass is 1- The expected loss value under the most unfavorable scenarios. By minimizing , which ensures good performance and robustness even in the “worst” scenarios. It's in the scene Next task Assign to Node Under uncertainty, different scenarios Represents possible implementation methods of task requirements (such as high or low task load situations). It is the cost (such as increased computing delay and increased risk of execution failure) incurred if resources are insufficient or mismatched in a given scenario. This value is calculated using the method defined above (for example: ). Therefore, when resources are insufficient to meet the scenario When the demand is will become larger. Uncertainty comes from inaccurate prediction of task requirements, temporary load fluctuations, hardware status changes, etc. By sampling or scenario analysis of uncertainty, we obtain a set of scenarios . Indicates possible demand scenarios, each of which gives a possible demand curve This is similar to the sample approximation of a random variable in stochastic programming. The scenario samples the uncertain distribution of task requirements. The larger it is, the more detailed the characterization of uncertainty is, but the computational complexity will also increase. A variety of possible demand curves can be generated using historical data sets or Monte Carlo simulations. Usually a higher value (such as 0.95 or 0.99) is taken, indicating that the worst loss situation we are concerned about is in the worst 1- The closer it is to 1, the more we pay attention to the worst-case scenario. Intuitively, when We consider the average expected loss under the worst 5% scenario. Used to linearize the CVaR definition, it can be understood as a “quantile” or critical loss level. Exceed hour, Excess losses are taken into account to ensure that the definition of CVaR is an optimizable convex problem. This auxiliary variable is adaptively adjusted in the optimization to find a suitable cutoff point so that the CVaR calculation can be achieved through linear programming techniques. If Less than ,but , Equal to 0, indicating that the loss of the scene does not exceed the critical value .if Greater than , there is a positive excess loss value. This piecewise function is conducive to calculating CVaR and focusing on the scenario losses that exceed the threshold. is a high confidence level (such as 0.95), then represents the worst 5% of the scenario subset of interest. CVaR focuses on this part of the worst scenario to calculate the average loss, thereby ensuring better performance in extreme cases. In the CVaR calculation, we only calculate the worst The scenes are averaged. Since we have When we estimate CVaR, we need to weight the excess loss item. This ensures that the weighted average has the correct probabilistic meaning, that is, the expected value is calculated only for the worst subset of scenarios.

[0039] Example 7: In step 3, a comprehensive optimization objective function is constructed by the following formula: ; in, is a binary decision variable. Assign to Node hour, ; otherwise 0; It's in the scene Next, the task At the initial moment of the task The resource demand vector of represents the L1 norm; Is a node The delay factor of Is a node The average available bandwidth to the data source; For Node Specific node resources Energy consumption coefficient; Includes: CPU, GPU and MEM; Is a node In time For node resources The macro available resources vector.

[0040] Specifically, is a binary decision variable. Assign to Node hour, ; otherwise 0. This variable is used to indicate the scheduling decision result. Finally dispatched to the node This ensures that in the final solution, each task can only be executed by one node. It is used to determine the critical loss level when calculating CVaR, and is used to transform the definition of CVaR into an optimizable linear (or convex) form. In this objective function, will with Appear together, Helps to robustly optimize resource mismatch losses in uncertain scenarios. The triple index uniquely locates the Task 1 and Task 2 The first cloud provider The scheduling relationship between the nodes. is the total number of tasks, is the number of cloud service providers, It is These are problem scale parameters that describe the overall topology and resource pool structure of the system. Metrics at worst -Expected resource misallocation loss in the quantile scenario Here, it is used as part of the optimization objective to control the risk under uncertain demand and ensure that the decision-making still maintains good performance in the worst case scenario. ) usually requires loading initial data or initialization. In a more "average" or expected scenario The total amount of resources required at the start of the task (for example, a certain amount of data needs to be loaded, CPU / GPU startup warm-up or memory initialization). This provides the basis for the subsequent latency item to be associated with bandwidth. : It represents the total resource demand of the task demand vector at the initial moment. For example, if the task initially requires 0.3CPU, 0.2GPU, and 0.4MEM occupancy, then the 1-norm is 0.3+0.2+0.4=0.9, which represents the overall demand for resources when the task starts. Describe from data center to node Obtain initial data and configure the latency characteristics required by the environment. For example, if the network latency is high or the node takes a long time to initialize a specific acceleration unit, This parameter is obtained by measuring network bandwidth, RTT (round-trip time), node initialization delay and other indicators, and is used to estimate the impact of initial data loading time on final performance. Characterizes the transmission rate capability. For example, if the bandwidth is larger, the transmission time will be shorter for the same initial data volume, thus reducing the startup delay. The term converts the initial demand into relative delay cost through bandwidth and delay factors. (such as CPU, GPU, MEM), The coefficient that represents the relationship between resource utilization and power consumption. Through power consumption metering and performance counters, we can fit the increased energy consumption when resource utilization increases. This coefficient allows us to Converted into energy consumption cost and used to reflect energy-saving requirements in optimization objectives. Indicates that in the task During the execution cycle, the node provides resource types for the task The cumulative occupancy of the resource type. It reflects the usage of the task during the entire execution time. Multiplying them together gives the energy consumption of the resource type. We need to ensure that resource requirements are met during this time and calculate the related consumption. It can be obtained from the mission performance model, user requirements or service level agreement (SLA).

[0041] Embodiment 8: In step 3, all constraints are integrated into a large nonlinear optimization problem through the following process: The equality constraint after discretization of the partial differential equation of the node resources is expressed by the following formula: ; in, for The discretization result is that the step size is used in space or meshing, using time steps in time To divide, represents the time step index; Is a node Approximate solution of resource vector fields at discrete grid points and discrete time levels; is the right-hand side of the discretized partial differential equation of node resources; The coefficient matrix obtained after discretization of the partial differential equation of node resources; the allocation constraint is expressed as: ; The feasibility constraint is expressed as: ; in, Yes Yes Node In time For node resources The macro-available resource vector; Because it is in the scene Next, the task On task at time The resource demand vector of . Specifically, for is a discrete version of . Using the step size in space or meshing, using time steps in time , Represents the time step index. For example . Is a node The resource availability vector approximates the solution at the discrete grid point and discrete time level. It is the discrete result of the PDE solution and is used for subsequent calculations . The continuous feature space The spatial step size used when dividing into finite elements or meshes. The smaller the time interval, the finer the spatial discretization and the higher the accuracy of the numerical solution, but the amount of calculation also increases. After discreteness, Represents a time point in a time series. By discrete time, PDE is transformed from solving continuous time to solving algebraic equations of a series of time layers. The core operator that transforms continuous PDE into a set of algebraic equations reflects the discrete approximation of the physical laws of resource changes within the node (diffusion, dissipation, exogenous supply). Indicates that the resource field is The influence of external inputs and fixed conditions.

[0042] Example 9: In step 4, the KKT system is constructed by using the primal-dual interior point method to solve the overall problem of the partial differential equation containing node resources and obtain the solution result: Construct the Lagrangian function The comprehensive optimization objective function, the partial differential equation of node resources, and the large nonlinear optimization problem that integrates all constraints are integrated to obtain the following formula: ; in, is the dual pairing operator; is the barrier function, which is used to ensure The feasible region does not deviate from during the solution process of the primal-dual interior point method; ; Due to the final need It is a 0-1 integer decision, but it is too difficult to solve the 0-1 problem directly. , in the initial stage of the solution, it is allowed In the interval (0,1), as the barrier parameter decreases, It will force the solution to approach 0 or 1, so that the primal-dual interior point method approaches the binary solution on the basis of continuous relaxation; is the dual variable of the allocation constraint, which requires that each task must be assigned to exactly one node. It measures the sensitivity of relaxing the allocation constraints to the Lagrangian function near the optimal solution; It is the dual variable of the equality constraint after the partial differential equation of the node resources is discretized, and is used to measure the influence of the equality constraint after the partial differential equation of the node resources is discretized on the Lagrangian function if it is relaxed or tightened; for The dual variable of If the resource constraint of a node is tight, that is, it just satisfies the equation, will become larger, indicating that the node resource has become a bottleneck; if there are surplus node resources, is 0, indicating that the constraint is not a tight constraint problem; the KKT condition is solved iteratively by the primal-dual interior point method, and the following process is performed in each iteration: using the current Approximate Update ; Use Newton's method to solve the linearized KKT system to find the optimal , and ; When making When the value of is the largest, stop the iteration and get The final value of ;right Perform a threshold determination, and if it is greater than the set threshold, it is assigned to the node.

[0043] Specifically, as mentioned above, Indicates that the task Assign to Node , otherwise it is 0. In the interior point method, it is temporarily allowed It can be solved by continuous relaxation and finally restored to a binary decision through a threshold. In optimization with scene loss Together, the CVaR constraint and the objective function remain convex and solvable. By optimizing the loss factor, we can find the appropriate “threshold” loss level and then obtain the CVaR value. is the intermediate variable in solving the PDE. As one of the original variables, the physical consistency of node resource dynamics is achieved by satisfying the constraints of discrete PDE equations. For The constraint requires that each task must be assigned to exactly one node. It measures the sensitivity of relaxing this constraint to the objective function near the optimal solution (shadow price). It is very large, indicating that the allocation constraints of the task have a greater impact on the optimization direction of the problem. Represents the dual variable introduced for constraints such as PDE systems and resource feasibility inequalities (it may be a scalar, vector or function, corresponding to the discrete version of the continuous constraint). Dual to the PDE discrete equation system: used to measure the impact of relaxing or tightening the PDE constraints on the objective function. When the value changes, it means that the satisfaction degree of PDE constraints is different and the intrinsic requirements for resource field distribution change. Dual to the resource feasibility inequality: Dual variables. In the optimal solution, if a resource constraint is tight (just satisfies the equation), will be large, indicating that the resource has become a bottleneck; if there are surplus resources, Usually 0 or small, indicating that the constraint does not tighten the problem. It is the core object of the original dual interior point method. By solving the primal and dual problems simultaneously, we can find the optimal solution that satisfies the KKT conditions.

[0044] Embodiment 10: A multi-cloud heterogeneous computing power scheduling system, the system comprising: an available resource calculation unit, used to use the partial differential equation of node resources to describe the distribution of each node resource evolving over time; after solving the partial differential equation of node resources, a macroscopic available resource vector is obtained by spatial integration; a scenario and constraint construction unit, used to compare the macroscopic available resource vector with the measured time-varying demand of the task to form a feasibility constraint; introduce uncertainty scenarios, sample scenarios for the time-varying demand of the task and use conditional risk value to control risks; an optimization target construction unit, used to construct a comprehensive optimization objective function; the objectives of the comprehensive optimization objective function include resource mismatch, delay and energy consumption; all constraints are integrated into a large nonlinear optimization problem; the constraints include: equality constraints, feasibility constraints and allocation constraints after the partial differential equation of node resources is discretized; a scheduling unit, used to construct a KKT system using the primal-dual interior point method, and solve the overall problem of the partial differential equation containing node resources in combination with the large nonlinear optimization problem to obtain a solution result; when the solution result converges, discretization is performed according to the value of the final solution result to obtain a final scheduling plan.

[0045] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. A multi-cloud heterogeneous computing power scheduling method, characterized in that: The method comprises: Step 1: Use the partial differential equation of node resources to describe the distribution of each node resource over time; solve the partial differential equation of node resources and obtain the macroscopic available resource vector through spatial integration; Step 2: Compare the macro-available resource vector with the measured time-varying requirements of the task to form feasibility constraints; introduce uncertainty scenarios, sample scenarios for the time-varying requirements of the task, and use conditional risk value to control risks; Step 3: Construct a comprehensive optimization objective function; the objectives of the comprehensive optimization objective function include resource mismatch, delay and energy consumption; integrate all constraints into a large nonlinear optimization problem; the constraints include: equality constraints after discretization of partial differential equations of node resources, feasibility constraints and allocation constraints; Step 4: Use the primal-dual interior point method to construct the KKT system, and combine it with the large-scale nonlinear optimization problem to solve the overall problem of the partial differential equation containing node resources to obtain the solution result; when the solution result converges, discretize it according to the final solution result value to obtain the final scheduling plan.

2. The multi-cloud heterogeneous computing power scheduling method according to claim 1, characterized in that: Let the task set be ;No. Tasks There is execution time ; The value range is 1 to ; is the total number of tasks; Representative tasks; let the multi-cloud collection be ; is the number of clouds in the multi-cloud set; For the cloud; Cloud There is a node collection ; The value range is 1 to ; For the Cloud The total number of nodes; Indicates Cloud No. nodes; Cloud The node resources in the feature space The continuous distribution, Indicates Cloud No. Nodes Hardware resource feature domain; The hardware resource feature domain is obtained by mapping the CPU core topology, GPU array and memory channel into a continuous domain; assuming that the node resources include: CPU, GPU and MEM; CPU represents the processor occupancy rate, ranging from 0 to 1; GPU represents the video memory occupancy rate, ranging from 0 to 1; MEM represents the memory occupancy rate, ranging from 0 to 1.

3. The multi-cloud heterogeneous computing power scheduling method according to claim 2, characterized in that: In step 1, the partial differential equation of node resources is used to describe the distribution of each node resource over time through the following formula: ； ； ； in, is a resource vector field describing the nodes At feature space point Place and time Available node resources; is the coordinate point in the feature space; is a time variable; It is a positive definite symmetric diffusion tensor obtained by measuring and modeling the node hardware topology and resource scheduling mechanism; represents the spatial gradient operator, represents the divergence operator; is the dissipation rate function, which is a scalar function or a diagonal matrix function; It is an exogenous regulatory item; is the initial condition function, giving Available node resources, which are obtained through benchmark testing and system initial status detection; is the maximum execution time; Representation Node At feature space point Place and time Available processor usage at the time Representation Node At feature space point Place and time Available video memory usage at the time Representation Node At feature space point Place and time Available memory usage at time ; This is the transpose operation of a vector.

4. The multi-cloud heterogeneous computing power scheduling method according to claim 2, characterized in that: In step 1, the partial differential equation of node resources is solved by the following formula and then the macro available resource vector is obtained by spatial integration: ； in, is the macro available resource vector; The feature space The coordinate point integration variable in .

5. The multi-cloud heterogeneous computing power scheduling method according to claim 4, characterized in that: In step 1, the following formula is used to compare the macro-available resource vector with the measured task time-varying requirements to form a feasibility constraint in step 2: ； in, Representation Node In time The total amount of processor occupancy available for assignment to tasks; Representation Node In time The total amount of video memory available for allocation; Representation Node In time The total amount of memory available for allocation; is the resource requirement vector, representing the task In time Requirements for processor occupancy, video memory occupancy, and memory occupancy; For the task In time Required processor resource utilization; For tasks in time Required video memory occupancy; For tasks in time Required memory usage.

6. The multi-cloud heterogeneous computing power scheduling method according to claim 5, characterized in that: In step 1, the following formula is used. In step 2, the following formula is used to introduce uncertainty scenarios, sample scenarios for time-varying task requirements, and use conditional risk value to control risks: ； in, Describe the worst case scenario under uncertainty -Percentile expected loss; different scenarios under uncertainty Represents possible implementation methods of task requirements; It is the cost incurred if resources are insufficient or mismatched in a given scenario; ); Therefore, when resources are insufficient to meet the scenario When the demand is will become larger; is the L2 norm; It is an auxiliary optimization variable introduced in the definition of conditional risk value, which is used to linearize the definition of conditional risk value and is defined as a quantile or critical loss level; when Exceed hour, Taking into account excess losses, the definition of conditional value at risk is ensured to be an optimizable convex problem; is the confidence level used to determine the proportion of the worst-case scenario being considered; is the total number of scenes.

7. The multi-cloud heterogeneous computing power scheduling method according to claim 6, characterized in that: In step 3, the comprehensive optimization objective function is constructed by the following formula: ； in, is a binary decision variable. Assign to Node hour, ; otherwise 0; It's in the scene Next, the task At the initial moment of the task The resource demand vector of represents the L1 norm; Is a node The delay factor of Is a node The average available bandwidth to the data source; For Node Specific node resources Energy consumption coefficient; Includes: CPU, GPU and MEM; Is a node In time For node resources The macro available resources vector.

8. The multi-cloud heterogeneous computing power scheduling method according to claim 7, characterized in that: In step 3, all constraints are integrated into a large nonlinear optimization problem through the following process: The equality constraints after discretization of the partial differential equations of the node resources are expressed as follows: ； in, for The discretization result is that the step size is used in space or meshing, using time steps in time To divide, represents the time step index; Is a node Approximate solution of resource vector fields at discrete grid points and discrete time levels; is the right-hand side of the discretized partial differential equation of node resources; The coefficient matrix obtained after discretization of the partial differential equation of node resources; the allocation constraint is expressed as: ； The feasibility constraint is expressed as: ； in, Yes Yes Node In time For node resources The macro-available resource vector; Because it is in the scene Next, the task On task at time The resource demand vector.

9. The multi-cloud heterogeneous computing power scheduling method according to claim 8, characterized in that: In step 4, the KKT system is constructed using the primal-dual interior point method through the following process to solve the overall problem of the partial differential equation containing node resources and obtain the solution result: Construct the Lagrangian function The comprehensive optimization objective function, the partial differential equation of node resources, and the large nonlinear optimization problem that integrates all constraints are integrated to obtain the following formula: ； in, is the dual pairing operator; is the barrier function, which is used to ensure The feasible region does not deviate from during the solution process of the primal-dual interior point method; ; Due to the final need It is a 0-1 integer decision, but it is too difficult to solve the 0-1 problem directly. , in the initial stage of the solution, it is allowed In the interval (0,1), as the barrier parameter decreases, It will force the solution to approach 0 or 1, so that the primal-dual interior point method approaches the binary solution on the basis of continuous relaxation; is the dual variable of the allocation constraint, which requires that each task must be assigned to exactly one node. It measures the sensitivity of relaxing the allocation constraints to the Lagrangian function near the optimal solution; It is the dual variable of the equality constraint after the partial differential equation of the node resources is discretized, and is used to measure the influence of the equality constraint after the partial differential equation of the node resources is discretized on the Lagrangian function if it is relaxed or tightened; for The dual variable of If the resource constraint of a node is tight, that is, it just satisfies the equation, will become larger, indicating that the node resource has become a bottleneck; if there are surplus node resources, is 0, indicating that the constraint is not a tight constraint problem; the KKT condition is solved iteratively by the primal-dual interior point method, and the following process is performed in each iteration: using the current Approximate Update ; Use Newton's method to solve the linearized KKT system to find the optimal , and ; When making When the value of is the largest, stop the iteration and get The final value of ;right Perform a threshold determination, and if it is greater than the set threshold, it is assigned to the node.

10. A multi-cloud heterogeneous computing power scheduling system, characterized in that: The system includes: an available resource calculation unit, which is used to use the partial differential equation of node resources to describe the distribution of each node resource evolving over time; after solving the partial differential equation of node resources, a macroscopic available resource vector is obtained by spatial integration; a scenario and constraint construction unit, which is used to compare the macroscopic available resource vector with the measured time-varying demand of the task to form a feasibility constraint; introduce uncertainty scenarios, sample scenarios for the time-varying demand of the task and use conditional risk value to control risks; an optimization target construction unit, which is used to construct a comprehensive optimization objective function; the objectives of the comprehensive optimization objective function include resource mismatch, delay and energy consumption; all constraints are integrated into a large nonlinear optimization problem; the constraints include: equality constraints, feasibility constraints and allocation constraints after the partial differential equation of node resources is discretized; a scheduling unit, which is used to construct a KKT system using the primal-dual interior point method, and solve the overall problem of the partial differential equation containing node resources in combination with a large nonlinear optimization problem to obtain a solution result; when the solution result converges, discretization is performed according to the value of the final solution result to obtain a final scheduling plan.

Images