Cloud native platform container migration method based on improved African buffalo algorithm

Abstract

The invention discloses a cloud native platform container migration method based on an improved African buffalo algorithm, and the method comprises the steps: constructing a multi-dimensional resource evaluation system of a video memory-computing power coupling model, a communication cost factor and a dynamic risk factor, and combining an improved African buffalo optimization algorithm of discrete-continuous hybrid coding; and collaborative scheduling of GPU / video memory resources across the data center is realized. According to the method, a position updating mechanism of an African buffalo algorithm is reconstructed, and the collaborative optimization problem of discrete variables (target data center selection and GPU number) and continuous variables (video memory allocation proportion) is solved through a hierarchical updating strategy (leader group / common group), dynamic attenuation weight and a disturbance mechanism. A bidirectional elastic migration strategy of a high-load node active emigration container and a low-load node intelligent absorption container is designed, and the limitation of traditional one-way migration is broken through. A local decision agent is deployed in each data center, an optimization algorithm is independently executed, and the communication overhead of a centralized architecture is avoided.

Description

Technical Field

[0001] This invention relates to a cloud-native platform container migration method based on an improved African buffalo algorithm, belonging to the technical field of cloud computing containerized resource scheduling. Background Technology

[0002] When individual users deploy intelligent computing services, the high procurement cost of dedicated accelerators such as GPUs poses a significant barrier to technological accessibility. However, elastic resource pooling solutions based on cloud computing technology can effectively achieve computing resource sharing. Currently, intelligent computing container migration technology in cloud-native platforms is relatively mature, basically meeting the elastic needs of artificial intelligence computing through hardware resource sharing, elastic resource allocation, and dynamic migration mechanisms. However, existing systems have several optimization flaws in container migration decisions: First, they use single-dimensional indicators or static threshold triggering mechanisms, failing to establish a multi-dimensional coupling model between memory and computing power; second, they lack dynamic quantitative assessment of resource overload risks, leading to the failure of resource combination optimization; third, traditional migration strategies optimize container selection and placement strategies separately, lacking a global load balancing perspective.

[0003] Existing research largely employs metaheuristic algorithms to design migration strategies. These algorithms, with their strong versatility, can effectively handle complex nonlinear combinatorial optimization problems in cloud-native platforms. Typical applications include dynamic scheduling frameworks based on ant colony optimization and distributed resource allocation using the African Buffalo Optimization (ABO) algorithm. Current mainstream solutions manage container-datacenter matching through global proxy nodes, seeking the optimal migration strategy among multiple combinations. However, this approach has two limitations: first, the centralized architecture leads to a surge in proxy node communication overhead, generating additional operating costs; second, while the algorithm's time complexity is better than traditional methods, it still grows exponentially, making it difficult to meet the real-time scheduling requirements of ultra-large-scale clusters. While distributed metaheuristic algorithms can alleviate these problems, key technical challenges such as multi-resource coupling optimization and local-global optimal solution coordination need to be addressed.

[0004] The African buffalo optimization algorithm, as an emerging swarm intelligence algorithm, simulates the ecological behavioral characteristics of buffalo herds foraging and migrating, constructing an adaptive mechanism based on a "calling" scoring mechanism, demonstrating unique advantages in dynamic equilibrium exploration and development. However, this algorithm has adaptability defects in intelligent computing container migration scenarios: the native ABO algorithm is an abstract general algorithm, and its position update formula relies on continuous space vector operations, which cannot directly handle discrete decisions such as data center selection; its unified update strategy is difficult to coordinate the collaborative optimization of discrete variables (such as the number of GPUs) and continuous variables (such as the memory allocation ratio), resulting in low solution space search efficiency. Algorithm reconstruction is needed for discrete variables (computable / non-computable), mixed variables (discrete-continuous composite), and the complex constraints of container migration, requiring the establishment of a distributed optimization framework oriented towards heterogeneous resource collaboration. Therefore, how to effectively combine metaheuristic algorithms to rationally design intelligent computing container migration strategies under cloud-native platforms has become a major research problem for the efficient development of artificial intelligence. Summary of the Invention

[0005] This invention aims to address the problems of low efficiency in multi-dimensional resource coordination, lack of dynamic risk management, and insufficient algorithm architecture adaptability in intelligent computing container migration within cloud-native platforms. It proposes a cloud-native platform container migration method based on an improved African buffalo algorithm. This method constructs a coupled update model of GPU quantity and memory ratio, a multi-dimensional resource-computing power coupling model, a multi-GPU parallel loss model, a dynamic overload risk quantification assessment system, and a container selection-placement collaborative decision-making mechanism. Combined with a discrete-continuous hybrid variable optimization framework, it achieves simultaneous improvement in resource utilization and scheduling efficiency in ultra-large-scale clusters. Specifically addressing the high cost of GPU resources, it innovatively designs a bidirectional elastic migration strategy, breaking through the limitations of traditional unidirectional "migration out" decisions. It simultaneously optimizes the pressure offloading of high-load nodes and the resource acquisition capabilities of low-load nodes, significantly reducing GPU resource idle rates and improving overall cost-effectiveness. This distributed local decision-making scheme reduces the exponential computational complexity of global centralized optimization by allowing each node to make its own migration decisions. This method effectively solves the shortcomings of traditional solutions, such as resource fragmentation, suboptimal migration decisions, and limited system scalability, providing key technical support for the artificial intelligence industry.

[0006] The technical solution adopted by this invention to solve its technical problem is: a cloud-native platform container migration method based on an improved African buffalo algorithm. This method achieves global optimization of cross-datacenter intelligent computing container migration by combining a multi-dimensional resource evaluation model with an improved swarm intelligence algorithm. The technical implementation process is based on a memory-computing power coupling model. It dynamically evaluates the comprehensive performance of heterogeneous computing resources through memory influence factors and computing power efficiency factors. Simultaneously, it introduces a communication cost composite factor to nonlinearly correlate network transmission time with the training cycle, forming a multi-dimensional migration decision evaluation system. This system integrates training time, communication cost, and resource risk factors to generate a comprehensive benefit score for the migration scheme. Based on this, to address the adaptability problem of the traditional African buffalo algorithm in discrete variable processing and hybrid optimization scenarios, a discrete-continuous hybrid encoding mechanism is designed. This mechanism achieves coordinated optimization of three types of parameters: datacenter selection, GPU quantity configuration, and memory allocation ratio through hash pointer mapping and dynamic step-size encoding. Furthermore, the design of risk factors and resource constraint factors in the fitness function must ensure that candidate solutions only emerge from the feasible solution space that satisfies the resource surplus constraints of the target node, thereby fundamentally avoiding the risk of migration failure.

[0007] This invention represents a significant improvement on the leader-following random mechanism. Leader group individuals learn from the optimal individual and randomly selected leader group individuals, while ordinary individuals dynamically choose to follow or explore strategies based on historical performance. The leader group employs a dynamic decay probability weighting strategy for parameter updates, while the ordinary group implements a three-stage cross-perturbation mechanism, preserving advantageous genes while enhancing population diversity. By establishing a fitness stability counter and dual convergence criteria, computational redundancy is significantly reduced while ensuring optimization accuracy. Addressing the unique resource fluctuation characteristics of cloud-native platforms, a dynamic bidirectional migration decision mechanism is constructed. When the local data center load factor exceeds a preset threshold, a pressure offloading mechanism is triggered for high-load nodes to actively migrate out of the optimal container, and a smart absorption mechanism is triggered for low-load nodes to intelligently absorb high-scoring containers from overloaded nodes. Global load balancing is achieved through bidirectional elastic migration.

[0008] To support the real-time scheduling requirements of the cluster and effectively reduce coordination overhead, the system adopts an adaptive distributed architecture design: each data center deploys a local decision agent to independently execute container optimization algorithms. When a container migration event occurs, a new round of the method of this invention is triggered. Adaptive global equilibrium is achieved by deciding on migration based on the algorithm results.

[0009] Method flow:

[0010] Step 1: Obtain the global data center list, the local data center and its communicable data centers, and obtain the local candidate migration container and solution list.

[0011] Step 2: Obtain the general attributes of the container that has been migrated to the local machine, and transfer the acquisition of sensitive attributes and algorithmic decisions to the internal operation of the container.

[0012] Step 3: Generate and initialize the buffalo population, and start counting at the same time.

[0013] Step 4: Traverse the population and calculate the fitness of individuals.

[0014] Step 5: Update the population identity based on fitness and update the count.

[0015] Step 6: Determine whether to stop updating the population based on the counting results. If you decide to stop updating, proceed to step 13; otherwise, proceed to step 7.

[0016] Step 7: Update the data center attributes of the individual buffalo population fitness leader group.

[0017] Step 8: Update the data center attributes of individuals in the normal fitness group of the buffalo population.

[0018] Step 9: Update the proposed GPU allocation attribute for each individual in the buffalo population fitness leader group.

[0019] Step 10: Update the proposed GPU allocation attribute for individuals in the normal fitness group of the buffalo population.

[0020] Step 11: Update the memory allocation ratio attribute of the buffalo population fitness leader group individuals.

[0021] Step 12: Update the memory allocation ratio attribute of the normal group of buffalo population fitness, and jump to step 4.

[0022] Step 13: Return the optimal individual data of the buffalo population to the data center, and update the candidate migration containers and their schemes based on the returned data and the various attributes of the migration container.

[0023] Step 14: Determine whether to move out of the container based on the load factor. If it is necessary to move out of the container, proceed to step 15; otherwise, proceed to step 16.

[0024] Step 15: Access the list of candidate containers and solutions corresponding to the communicable data center, select the most suitable container and migrate it out according to its configuration scheme, and then jump to step 14.

[0025] Step 16: Determine whether it is possible to migrate into the container based on the load factor. If it is possible to migrate into the container, proceed to step 17; otherwise, end the algorithm.

[0026] Step 17: Traverse the list of communicable data centers, check if there are suitable containers to migrate locally, and if so, select the most suitable container and migrate it according to its configuration scheme.

[0027] Step 18: Determine if there is a suitable container that needs to perform the active container migration operation. If there is, go to step 2; otherwise, end the algorithm.

[0028] Beneficial effects:

[0029] 1. This invention establishes a multi-dimensional resource evaluation model that includes computing power efficiency factors, memory impact factors, communication cost factors, and risk factors. Combined with an improved African buffalo optimization algorithm, it achieves dynamic optimization of container migration schemes. At the same time, it constructs a bidirectional adaptive migration mechanism to achieve global optimization configuration of heterogeneous GPU resources and load balancing across data centers while ensuring the execution efficiency of container tasks.

[0030] 2. This invention achieves collaborative scheduling of GPU / memory resources across data centers by constructing a multi-dimensional resource evaluation system that includes a video memory-computing power coupling model, communication cost factors, and dynamic risk factors, combined with an improved African buffalo optimization algorithm based on discrete-continuous hybrid coding. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method of the present invention.

[0032] Figure 2 This is a deployment diagram of the method of the present invention. Detailed Implementation

[0033] The invention will now be described in further detail with reference to the accompanying drawings.

[0034] like Figure 1 and Figure 2 As shown, this invention provides a cloud-native platform container migration method based on an improved African buffalo algorithm, which includes the following steps:

[0035] Step 1: Obtain the global data center list Dc = {Dc1, Dc2, ..., Dc...} n}, obtain local data center dcl ocal (dc local Total GPU resources dc (∈Dc) local .G total GPU resource availability (DC) local .G available Total video memory resources (DC) local .V total Memory resource availability (DC) local .V available Bandwidth resource margin (dc) local .BW available Get DC from DC local List of communicable data centers DL = {dc1, dc2, ..., dc i, ..., dc n} and the list length DL.size, get each dc i The total GPU resources dc corresponding to (i∈[1,DL.size]) i .G total GPU resource availability (DC) i .G available Single GPU computing power DC i .G flops Total video memory resources (DC) i .V total Memory resource availability (DC) i .V available Bandwidth resource margin (dc) i .BW available Get DC local Candidate migration to DC i The list of containers and their solutions, IC_List i ={ic1, ..., ic j , ..., ic n}, where the j-th record ic j Includes IC j The unique identifier of the container recorded (ic) j .icid, the number of GPUs the container intends to allocate. j .N GPU The container is intended to allocate a certain percentage of single-GPU video memory. j .P VRAM ,dc i With IC j fitness ic j .Sound.

[0036] Step 2: When the container IC local Migration to data center DC local At that time, obtain IC local unique identifier ic local. icid, container data volume to be migrated (ic) local .S container Through the interface, DC local .BW available DL, DL.size, and each dc in DL i dc corresponding to (i∈[1,DL.size]) i .G total ,dc i .G available ,dc i .G flops ,dc i .V total ,dci .V availabve ,dc i .BW available ic local .S container and the improved African buffalo algorithm function input ic local The specific algorithm operations are shown in steps 3 to 13 below. In the container ic local Internally, by IC local Obtain the GPU computation floating-point number S for a single epoch of the intelligent computing task being carried. flopsepoch Maximum preset number of rounds N for the task maxepoch The number of training rounds for the task is N. trainedepoch The definition of prediction still requires N training rounds. epoch ,

[0037] N epoch =N maxepoch -N trainedepoch .

[0038] Step 3: In IC local Internal generation size is population Among them, each individual x i (i∈[1,N) Buf ]) all store data elements (dc) T N GPU P VRAM (Sound), each x i Initialize its data items separately: Randomly select a data center from the DL as the target data center entity dc T Randomly select an integer N from the interval [1, 16] as the number of GPUs to be allocated. GPU A real number is randomly selected from the interval [0.3, 1] as the proposed single-GPU memory allocation ratio P. VRAM The fitness (Sound) is initialized to 0. Define the number of updates (steps) for population X and the number of times the optimal Sound attribute of X stabilizes (count). Sound And all of them are initialized to 0.

[0039] Step 4: Traverse the population X, for each x i (i∈[1,N) Buf Perform the following operation: Let x be the value of x. i DC T The data center in the DL corresponding to the data item is dc k Define the available communication bandwidth BWef.

[0040] BWef=min{dc k .BW available DC local.BW available}

[0041] Define memory impact factor α (P) VRAM ),

[0042]

[0043] Define G flops Efficiency factor η(N) GPU ),

[0044]

[0045] Define the training time T for a single round of the task. epoch ,

[0046]

[0047] Define the data center resource constraint factor P resource ,

[0048]

[0049] Define the risk factor Risk.

[0050]

[0051] Define the communication evaluation time T comu ,

[0052]

[0053] Define the communication impact factor β.

[0054]

[0055] Calculate fitness sound,

[0056]

[0057] Step 5: Increment the steps count by 1. Apply the Topk algorithm to all individuals in X to filter out the data items before the Sound data items. For individuals with higher stats, the selected individuals are swapped to the front of X. One position. And in X[1] to Select the individual with the highest Sound value from the pool and swap it to position X[1]. If the current individual in X[1] has a Sound value of 0, then swap it to position X[1]. (steps) Sound fitness of individuals before updating using the Topk algorithm (steps-1) The ratio satisfies Then count Sound Increment the count by 1; otherwise, increment the count. SoundSet to 0 and start counting again.

[0058] Step 6: If steps > 150 and count sound If the value is greater than 30, or steps = 500, then proceed to step 13; otherwise, proceed to step 7.

[0059] Step 7: Traverse X[2] to X in X All individuals x stored i And update DC T The attribute is updated as follows: Select an individual from X[1] as x best Define dynamic probability weights pα1(steps) and pβ1(steps), where,

[0060] pα1(steps)=0.55e -0.005steps +0.1

[0061] pβ1(steps)=0.2(1-e -0.01steps )

[0062] Each iteration Individual x i Perform the following operation: Define random numbers r1 ~ U(0, 1),

[0063] And randomly select Individual x lrand ,calculate

[0064]

[0065] Step 8: Traverse X To X[N Buf All individuals x stored i And update DC T The attribute is updated as follows: Select an individual from X[1] as x best Each time a new instance is traversed... Individual x i Perform the following operation: Define a random number r2 ~ U(0, 1), in Individuals are randomly selected as x lrand ,exist Individuals are randomly selected as x frand Calculate, if but

[0066]

[0067] otherwise,

[0068]

[0069] Step 9: Traverse X[2] to X in X All individuals x stored i And update N GPU The attribute is updated as follows: Select an individual from X[1] as x best Define dynamic probability weights pα²(steps) and pβ²(steps), where,

[0070] pα²(steps) = 0.7e -0.005steps +0.15

[0071] pβ2(steps)=0.3(1-e -0.01steps )

[0072] Each iteration Individual x i Perform the following operation: Individuals are randomly selected as x lrand Calculate intermediate variables

[0073]

[0074] Using N GPUround calculate

[0075]

[0076] Step 10: Traverse X To X[N Buf All individuals x stored i And update N GPU The attribute is updated as follows: Select an individual from X[1] as x best Each time a new instance is traversed... Individual x i Perform the following operation: Define a random number r3 ~ U(-2,2), in Individuals are randomly selected as x lrand .if Then calculate the intermediate variables.

[0077]

[0078] Otherwise, calculate intermediate variables.

[0079]

[0080] Using N GPUround calculate

[0081]

[0082] Step 11: Traverse X[2] to X in X All individuals x stored i And update P VRAM The attribute is updated as follows: Select an individual from X[1] as x best Define dynamic probability weights pα3(steps) and pβ3(steps), where,

[0083] pα3(steps)=0.6e -0.005steps +0.15

[0084] pβ3(steps)=0.2(1-e -0.01steps )

[0085] Each iteration Individual x i Perform the following operation: Individuals are randomly selected as x lrand Define the disturbance amplitude δ p ,

[0086]

[0087] The memory ratio corresponding to a single GPU should be defined as N. GPU The change is ΔP VRAM The calculation method is as follows: If Then ΔP VRAM ~U(-δ p , 0); otherwise, ΔP VRAM ~U(0,δ p ). Calculate intermediate variables

[0088]

[0089] Using N VRAMround calculate

[0090]

[0091] Step 12: Traverse X To X[N Buf All individuals x stored i And update P VRAM The attribute is updated as follows: Select an individual from X[1] as x best Each time a new instance is traversed... Individual x i Perform the following operation: Individuals are randomly selected as x lrand Define the disturbance amplitude δ p ,

[0092]

[0093] The memory ratio corresponding to a single GPU should be defined as N. GPU The change is ΔP VRAM The calculation method is as follows: If Then ΔP VRAM ~U(-δ p , 0); otherwise, ΔP VRAM ~U(0,δ p ).if Then calculate the intermediate variables.

[0094]

[0095] Otherwise, calculate intermediate variables.

[0096]

[0097] Using N VRAMround calculate

[0098]

[0099] Calculation complete. Proceed to step 4.

[0100] Step 13: Connect the container IC via the interface. local Data elements (dc) of individuals in inner X[1] T N GPU P VRAM (Sound) is transmitted to the data center (dc) local Traverse the deep learning (DL) using the pointer target to find the dc of the data element. T For each data item, find the corresponding DC in the DL. target and IC_List target If the IC_List target If empty, then in the IC_List target Insert a new record at the beginning of the list; otherwise, traverse the IC_List using a pointer. target Insert a new record before the record whose fitness Sound data item is lower than the record containing the current data element's Sound data item. In the new record, the container's unique identifier is assigned the value ic. local .icid, the number of GPUs the container intends to allocate is assigned to this data element as N. GPU The data item, the container intends to allocate a single GPU memory ratio to assign the value P to this data element. VRAM Data item, dc target With IC local The fitness value is assigned to the Sound data item of this data element.

[0101] Step 14: Update DC local .G available and DClocal .V available Calculate DC local Load factor local ,

[0102]

[0103] If Load local If the value is greater than 0.8, proceed to step 15; otherwise, proceed to step 16.

[0104] Step 15: Traverse the DL and find the record IC_List with the largest Sound data item. i .ic1 and its corresponding DC i Move the container in the record to this DC. i Configure the migrated N according to this record. GPU and P VRAM Proceed to step 14.

[0105] Step 16: If Load local If the value is less than 0.4, proceed to step 17; otherwise, end the algorithm.

[0106] Step 17: Define and initialize the immigration flag `tagin = false`. Iterate through the deep learning flow (DL) for each dc. i (i∈[1,DL.size]) Calculate the load factor Load i ,

[0107]

[0108] If Load exists i >2Load local Then update tagin = true, and load all conditions that meet the condition. i >2Load local DC i The corresponding candidate migration to DC local From the list of containers and their schemes, select the record with the highest value for the Sound data item, and move the container in that record into DC. local Configure the migrated N according to this record. GPU and P VRAM Proceed to step 18.

[0109] Step 18: If tagin = true, then go to step 2; otherwise, end the algorithm.

[0110] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A cloud-native platform container migration method based on an improved African buffalo algorithm, characterized in that, The method includes the following steps: Step 1: Obtain the global data center list, the local data center and its communicable data centers, and obtain the local candidate migration container and solution list; Step 2: Obtain the general attributes of the container that has been migrated to the local machine, and transfer the acquisition of sensitive attributes and algorithmic decisions to the internal operation of the container; Step 3: Generate and initialize the buffalo population, and start counting simultaneously; Step 4: Traverse the population and calculate the fitness of individuals; Step 5: Update population identity based on fitness and update the count; Step 6: Determine whether to stop updating the population based on the counting results. If you decide to stop updating, proceed to step 13; otherwise, proceed to step 7. Step 7: Update the data center attributes of the buffalo population fitness leader group individuals; Step 8: Update the data center attributes of individuals in the normal fitness group of the buffalo population; Step 9: Update the proposed GPU allocation attribute for each individual in the buffalo population fitness leader group; Step 10: Update the proposed GPU allocation attribute for individuals in the normal fitness group of the buffalo population; Step 11: Update the memory allocation ratio attribute of the fitness leader group individuals in the buffalo population; Step 12: Update the memory allocation ratio attribute of individuals in the normal group of buffalo population fitness, and jump to step 4; Step 13: Return the optimal individual data of the buffalo population to the data center, and update the candidate migration containers and their solutions based on the returned data and the various attributes of the migration container; Step 14: Determine whether to move out of the container based on the load factor. If it is necessary to move out of the container, proceed to step 15; otherwise, proceed to step 16. Step 15: Access the list of candidate containers and solutions corresponding to the communicable data center, select the most suitable container and migrate it out according to its configuration scheme, and then jump to step 14; Step 16: Determine whether it is possible to migrate into the container based on the load factor. If it is possible to migrate into the container, proceed to step 17; otherwise, end the algorithm. Step 17: Traverse the list of communicable data centers, check if there are suitable containers to migrate locally, and if so, select the most suitable container and migrate it according to its configuration scheme. Step 18: Determine if there is a suitable container that needs to perform the active container migration operation. If there is, go to step 2; otherwise, end the algorithm.

2. The cloud-native platform container migration method based on the improved African buffalo algorithm according to claim 1, characterized in that, The method involves leaders learning from the optimal individual and random leaders, while ordinary individuals dynamically choose to follow or explore based on their historical performance. The leader group uses a dynamic decay probability weighting strategy for parameter updates, and the ordinary group implements a three-stage cross-perturbation mechanism to both preserve advantageous genes and enhance population diversity. By establishing a fitness stability counter and dual convergence criteria, computational redundancy is significantly reduced while ensuring optimization accuracy. Targeting the unique resource fluctuation characteristics of cloud-native platforms, a dynamic bidirectional migration decision mechanism is constructed. When the load factor of the local data center exceeds a preset threshold, a pressure unloading mechanism for high-load nodes to actively migrate out of the optimal container and an intelligent absorption mechanism for low-load nodes to absorb high-scoring containers from overloaded nodes are triggered, respectively. Global load balancing is achieved through bidirectional elastic migration.

3. The cloud-native platform container migration method based on the improved African buffalo algorithm according to claim 1, characterized in that, Step 1 includes: obtaining a global data center list Dc = {Dc1, Dc2, ..., Dc...} n }, obtain local data center (dc) local (dc local Total GPU resources dc (∈Dc) local .G total GPU resource availability (DC) local .G available Total video memory resources (DC) local .V total Memory resource availability (DC) local .V available Bandwidth resource margin d clocal .BW available Get DC from DC local List of communicable data centers DL = {dc1,dc2,...,dc i ,...,dc n } and the list length DL.size, get each dc i The total GPU resources dc corresponding to (i∈[1,DL.size]) i .G total GPU resource availability (DC) i .G available Single GPU computing power DC i .G flops Total video memory resources (DC) i .V total Memory resource availability (DC) i .V available Bandwidth resource margin (dc) i .BW available Get DC local Candidate migration to DC i The list of containers and their solutions, IC_List i ={ic1, ..., ic j , ..., ic n }, where the j-th record ic j Includes IC j The unique identifier of the container recorded (ic) j .icid, the number of GPUs the container intends to allocate. j .N GPU The container is intended to allocate a certain percentage of single-GPU video memory. j .P VRAM ,dc i With IC j fitness ic j .Sound.

4. The cloud-native platform container migration method based on the improved African buffalo algorithm according to claim 1, characterized in that, Step 2 includes: when the container ic loval Migration to data center DC local At that time, obtain IC loval unique identifier ic local .icid, container data volume to be migrated (ic) local .S container Through the interface, DC local .BW available DL, DL.size, and each dc in DL i dc corresponding to (i∈[1,DL.size]) i .G total ,dc i .G available ,dc i .G flops ,dc i .V total ,dc i .V available ,dc i .BW available i clocal .S container And the improved African buffalo algorithm function input i clocal The specific algorithm operations are shown in steps 3 to 13 below, in the container ic local Internally, by IC local Obtain the GPU computation floating-point number S for a single epoch of the intelligent computing task being carried. flopsepoch Maximum preset number of rounds N for the task maxepoch The number of training rounds for the task is N. trainedepoch The definition of prediction still requires N training rounds. epoch N epoch =N maxepoch -N trainedepoch .

5. A cloud-native platform container migration method based on an improved African buffalo algorithm according to claim 1, characterized in that, Step 3 includes: in ic local Internal generation size is population Among them, each individual x i (i∈[1,N) Buf ]) all store data elements (dc) T N GPU P VRAM (Sound), each x i Initialize its data items separately: Randomly select a data center from the DL as the target data center entity dc T Randomly select an integer N from the interval [1, 16] as the number of GPUs to be allocated. GPU A real number is randomly selected from the interval [0.3, 1] as the proposed single-GPU memory allocation ratio P. VRAM The fitness Sound is initialized to 0. The number of updates (steps) and the number of times the optimal Sound attribute of X is stable (count) are defined for population X. Sound And all of them are initialized to 0.

Images