37 results about "Weighted round robin" patented technology

The invention relates to the field of database reading, and particularly discloses an efficient database reading method based on multi-level cache optimization and dynamic index fragment.The efficient database reading method comprises the steps that high-frequency query data is accurately recognized through a sliding window algorithm, a dynamic cache loading and preheating strategy is implemented, query delay is effectively shortened, and the database burden is relieved; intelligent distribution of index fragments is realized by applying a consistent Hash algorithm, and the query efficiency is greatly improved in cooperation with distributed query routing and transaction processing optimization; the query load balancing ensures stable and efficient operation of the system in a high-concurrency scene by means of weighted polling and a minimum connection strategy in combination with real-time monitoring and an elastic capacity expansion and contraction mechanism. According to the method, the database access efficiency and the dynamic adaptive capacity are remarkably improved, and the method is particularly suitable for a large-scale distributed data processing environment.

The invention discloses a multi-level interconnection network based on real-time flow monitoring and dynamic priority adjustment. A multi-dimensional traffic sensing model is constructed through traffic intensity, queue depth, congestion signals and static priority dynamic indexes. A dynamic urgency value is generated by adopting a configurable weighting strategy, and a three-mode closed-loop optimization mechanism is designed. A layered interconnection structure is adopted, crossbar switches are used in computing clusters, and the computing clusters are connected through a low-diameter routing network. In a cluster gateway, a dynamic arbitration unit monitors the flow intensity in real time, calculates the urgency degree in combination with a queue state and a global congestion signal, and dynamically selects a fixed priority and a weighted polling or polling arbitration strategy. The problem that the resource utilization rate of a traditional static arbitration mechanism is reduced due to task switching is effectively solved, the risk of fairness imbalance in a mixed key data flow scene is effectively solved, and the communication adaptation capacity of a reconfigurable architecture to diversified computing loads is improved.

The invention relates to the technical field of data transmission, in particular to a CAN communication network multi-rate adaptive transmission control system, a control method and a vehicle electronic control system. According to the CAN communication network multi-rate adaptive transmission control system, dynamic rate adjustment and intelligent load balancing are realized through four core modules: a data classification module divides data into a security class, a power class, a comfort class and an entertainment class, and calculates a dynamic weight; the network state monitoring module adopts a sliding window algorithm to calculate a network load rate in real time; the rate decision module dynamically selects the transmission rate according to the priority and the load rate; and the multi-path control module selects an optimal standby path based on the comprehensive score of the path load rate, the hop count and the link quality, and distributes traffic through a weighted polling algorithm. The system breaks through the limitation of a traditional CAN fixed rate, realizes low-delay transmission of high-priority data and efficient utilization of network resources, and is particularly suitable for a high-load communication scene of an intelligent automobile.

The invention provides an unmanned cluster dynamic task allocation method, which belongs to the technical field of unmanned aerial vehicle clusters, and comprises the following steps: firstly, initializing a plurality of targets and basic information of the targets; the target broadcasts task requirements to all the unmanned aerial vehicles and dynamically adjusts the priority of tasks; the unmanned aerial vehicle calculates an evaluation value based on the bidding and submits the bidding to the target agent; distributing the tasks by adopting a weighted polling algorithm, and readjusting the priorities and distribution of the tasks when an emergency occurs; the method solves the defects of insufficient real-time performance, large communication overhead and the like of the existing contract network algorithm, and is suitable for the unmanned aerial vehicle cluster to execute the attack task under the complex condition.

The invention relates to the technical field of solid-state storage, in particular to a solid-state disk data consistency control method, and aims to solve the problems that a solid-state disk is poor in data consistency, high in power failure loss risk and difficult to balance writing efficiency and consistency under the scenes of power failure, parallel writing and garbage collection. Full-process consistency control is realized by constructing a'cache-persistent cache 'layered architecture and combining a multi-channel write synchronization, power failure emergency migration and periodic check repair mechanism: receiving data and storing the data in a cache firstly, distributing a multi-channel write task according to weighted polling and synchronizing the state; the standby capacitor is triggered to supply power during power failure, and the cache data is migrated to the persistent cache according to the FIFO; and according to a configurable period of 1-24 hours, flash memory block data are verified through SHA-256 hash, and when the flash memory block data are inconsistent, the flash memory block data are repaired from a persistent cache / backup block. According to the method, the data consistency compliance rate of the solid-state disk is larger than or equal to 99.9%, the power-down data loss rate is smaller than or equal to 0.01%, the write-in delay is reduced by 15%-20%, and the method is adaptive to consumption-level, enterprise-level and industrial-level solid-state disks.

The invention discloses a multi-AGV cooperative scheduling system for realizing load balancing. The multi-AGV cooperative scheduling system comprises a task management module, a state monitoring module, a load balancing scheduling module, an AGV cooperative scheduling module, an edge server optimization module, a task compression module, a task parallel processing module and a performance evaluation module. Wherein the load balancing scheduling module adopts a mode of combining first-stage weighted polling initial allocation and second-stage optimal allocation based on multi-agent deep reinforcement learning, the AGV collaborative scheduling module adopts an AxTD3 method to divide an Actor network into two parallel sub-networks, and joint training is carried out by sharing a Critic network. The edge server optimization module takes execution time deviation as a state, constructs a lightweight state vector in combination with a server isomerism index, and introduces federated learning and multi-Critic network architecture. According to the invention, adaptive collaborative scheduling of multiple AGVs in an edge computing environment is realized, and the system load balancing capability, the resource utilization efficiency and the task processing response speed are effectively improved.

The invention discloses a question and answer system knowledge base shunting method based on a weighted polling algorithm, and belongs to the technical field of artificial intelligence retrieval enhancement generation. In order to overcome the defects of a single knowledge base architecture of a traditional RAG system in the aspects of multi-source knowledge management, gray updating and load balancing, a cluster architecture comprising a main knowledge base, an auxiliary knowledge base and a test knowledge base is constructed, and intelligent flow distribution is achieved through a flow distribution controller. The system adopts a multi-dimensional weight calculation model, fuses a user role coefficient, a knowledge base state coefficient and a performance coefficient, and dynamically calculates the weight of each knowledge base according to a weight calculation formula; using a weighted polling algorithm to accurately distribute user requests; and automatically adjusting the weight of the knowledge base in the gray scale test period. The vector retrieval module converts questions into vectors to retrieve in a target library, and a result fusion engine resolves multi-library answer conflicts through weighted voting. After the method is applied, the update failure rate of the knowledge base is reduced from 25% to 5%, the question and answer accuracy is improved by 22%, and the response speed is improved by 40%.

This invention discloses a dynamic optimization method for network switch port WRR scheduling resources based on convex optimization modeling. Applied to the field of switching equipment scheduling and control, it addresses the problem that existing scheduling weight configurations cannot adapt to environmental changes, resulting in unstable forwarding performance of switching equipment. This invention establishes constraint functions based on port bandwidth, token resources, and queue number, collects the arrival probability distribution of data packets in each queue within the switch port to establish a distribution model, and combines the constraint functions and the collected model to construct a convex optimization problem for switch port queue resource allocation. The objective of the optimization problem is defined as minimizing the sum of queuing probabilities of the switch queues. The problem is solved using an existing convex optimization solver, and the results are configured as the weights of each port queue based on the Weighted Round Robin (WRR) scheduling algorithm. This invention solves the problem of adaptive resource allocation during dynamic network traffic exchange, enabling network switching equipment to provide stable network performance under different service loads.

The application relates to a multi-dimension dynamic weighted polling scheduling method and system for bank-enterprise direct connection query, which comprises the following steps: acquiring account information of all accounts of an enterprise and initializing a sorting value of each account; screening out accounts to be queried in a current batch according to the initialized sorting value of each account; calling a bank-enterprise direct connection interface to query the balance or transaction details of the accounts in the current batch; calculating an average transaction number interval based on the transaction number of each account and adopting a zero-elimination average method; respectively calculating a new basic weight value, a historical transaction weight value and a real-time transaction weight value by adopting a specific calculation method; generating an account weight value; generating a new sorting value for each account in the current batch according to the account weight value; screening out accounts to be queried in the next batch of the current batch according to the new sorting value; and stopping until a preset iteration number is reached, completing the polling scheduling of the enterprise account query and being suitable for efficient query scheduling scenes in a large-scale bank account environment.

The invention discloses an anti-attack smart park management system, method and device, and a storage medium, relates to the technical field of park management, and aims to solve the problem of low regulation and control efficiency of traditional park management on multi-device resources and an access control system. The system comprises an edge trusted processing layer which is used for performing feature compression and integrity verification on multi-source equipment data including energy consumption measurement data, access control biological features and equipment state data through a zero-knowledge proof algorithm, and outputting verification features carrying cryptographic proof; the block chain evidence storage network is used for constructing a distributed account book according to the verification characteristics and generating a charging voucher and an equipment health voucher; the micro-service scheduling center is used for dynamically distributing service resources based on a priority weighted polling algorithm and integrating anti-interference communication to transmit a voucher data packet comprising a charging voucher and an equipment health voucher; and the intelligent decision center is used for generating an access control instruction and an equipment regulation and control strategy based on the digital twin model and the voucher data packet.

A device includes a plurality of digital signal processors (DSPs) and analog crossbars in communication with the DSPs. The DSPs and crossbars dynamically adjust power consumption based on traffic load and / or a signal quality, leveraging traffic prediction models to optimize resource allocation. The system supports features such as powering down unused lanes, asymmetrically deactivating transmission or receiving lanes, and preemptively scaling resources during predicted traffic increases. Dynamic voltage and frequency scaling (DVFS) optimizes power usage based on link utilization and traffic priority, minimizing latency impacts during transitions. The switch controller integrates traffic prioritization algorithms, including weighted round-robin (WRR), to allocate power and bandwidth efficiently to high-priority traffic flows. Redundant crossbars handle overflow traffic or failover scenarios, transitioning between standby and active states dynamically.

The invention relates to a QOS (Quality of Service) strategy selection and queue scheduling method of a base switching network, which is characterized by comprising the following steps of: firstly, after network equipment receives a data frame, executing a QOS strategy selection operation by the network equipment, selecting a priority mapping table corresponding to the data frame from five different priority mapping tables, querying in a priority mapping table corresponding to the data frame to obtain a priority ID of the data frame; then, the network equipment queries a priority ID queue mapping table based on the priority ID of the data frame to obtain an output queue ID, and inserts the data frame into an output queue corresponding to the output queue ID; and finally, the network equipment transmits the output queue with the data frames to receiving equipment at an output port through a weighted polling scheduling (WRR) algorithm. According to the method and the device, a plurality of priority mapping tables are preset, so that the QOS model can process the data frames in various formats, and the types of compatible and supported data traffic are expanded.

The invention discloses a method for releasing a server program by one key, which realizes full-process automatic release through cooperation of six core modules: an interactive trigger module supports multi-end triggering of a Web interface, an API interface and a command line, provides lightweight operation only needing eight core configurations, and supports historical configuration multiplexing and network interrupt caching; the configuration analysis and verification module is used for presetting multiple types of application analysis rules, executing multi-level verification of a basic format, service logic and resource availability, and providing an automatic correction suggestion when verification fails; the environment adaptation and resource scheduling module remotely detects server hardware, software and network states, automatically adapts to scripts corresponding to an operating system and an application type, allocates resources through a weighted polling algorithm, and supports cross-regional network optimization; the real-time monitoring and feedback module collects logs and state indexes in a second level. The whole release process is free of manual intervention, the deployment time consumption is remarkably shortened, and the error rate is reduced to a low level.

The invention provides an alarm computing load balancing method, which belongs to the technical field of cloud computing and distributed monitoring operation and maintenance, adopts a distributed architecture, and comprises a plurality of Prometheus examples, a scheduling management module, a task distributor, an equal division algorithm executor and a shared storage component. The scheduling management module dynamically fragments the alarm rules according to rule types, data source features or time dimensions, and adjusts the number and size of fragments according to real-time loads. And the equal division algorithm executor combines the load, the performance index and the resource utilization rate of each instance, adopts strategies such as consistent Hash, polling or weighted polling to calculate an assignment relationship, and supports newly-added rule allocation and fragment migration. Dynamic management and real-time synchronization of rules are achieved through the shared storage component, stateless deployment of Prometheus instances is achieved, requests are distributed through the load balancer, and therefore the load balancing capacity, the rule management efficiency and the alarm timeliness are improved.

The invention discloses a power distribution communication network resource allocation method and device, and belongs to the technical field of network resource allocation, and the method comprises the steps: obtaining the business content of a new access business and a target resource requirement; judging the service type of the new access service based on the service content; when the service type of the new access service is an instruction control type, dividing the new access service to a core guarantee layer of the power distribution communication network, and distributing network resources to the new access service according to a target resource requirement by adopting a preemptive priority scheduling method; and when the service type of the new access service is a data processing service, dividing the new access service to a data support layer of the power distribution communication network, and distributing network resources to the new access service according to the target resource requirement by adopting a weighted polling method. Therefore, by implementing the method and the device, reasonable resource scheduling can be realized, so that the resource utilization rate is effectively improved.

The invention relates to the technical field of enterprise management systems, in particular to an enterprise management system optimization method based on background data, which comprises the following steps of: performing hierarchical encryption on sensitive data, dividing a core original layer, an aggregation middle layer and an abstract display layer, and matching a differentiation algorithm; then calculating a permission score through a hybrid model, and dynamically controlling data presentation; an encryption queue is optimized based on real-time load activation weighted polling or a minimum connection number algorithm, an access range is shrunk by combining access map pruning, and threats are isolated; switching a national cipher SM9 protocol and dual-channel redundancy transmission according to channel quality, and realizing full-life-cycle anomaly detection through a data blood relationship chain and a graph neural network; finally, the dynamic presentation rule is optimized through reinforcement learning iteration. According to the method, accurate protection of sensitive data, dynamic adaptation of system loads, controllability of authority fine grit and traceability of a data circulation full link are achieved, the service adaptability and user experience are improved, compliance requirements are met, and stable and efficient support is provided for digital transformation of enterprises.

The invention discloses a multi-protocol intelligent cockpit interconnection integrated system, which relates to the field of communication compatibility, and comprises the following steps: S1, a hardware interface module of a universal interconnection interface layer receives an access request of a mobile phone, and establishes communication connection between the mobile phone and a vehicle machine; original data transmitted by the mobile phone is subjected to standardized primary processing through a unified data transmission protocol of the software interface module; s2, a dynamic identification module of the protocol adapter layer adopts a feature code scanning technology to identify the type of an interconnection protocol accessed to a mobile phone within 100ms; the type of the interconnection protocol comprises a CarPlay, an Android Auto, a HiCar and a CarLink; calling a protocol conversion engine to carry out adaptation aiming at different protocols, and converting data in a mobile phone protocol format into data in a system general data format; and S3, a resource scheduler of the central control layer adopts a DWRR dynamic weighted polling algorithm to dynamically allocate CPU and memory resources of the vehicle machine according to the protocol type identified in the step S2 and the current service scene.

The application provides an implementation method and device for application atomization fusion arrangement based on OPENAPI, relates to the technical field of computer application, and covers service disintegration and interface definition, service registration and discovery, arrangement and adaptive execution, data processing and mapping, security control, integrated deployment and operation and maintenance steps; the application splits complex business into atomized services, defines interfaces according to the OpenAPI specification, facilitates service management, reduces integration cost, and improves system flexibility and expansibility; load balancing is realized by constructing a service registration center and combining a weighted round robin algorithm to guarantee efficient and stable service calling; API gateways, containerization technology and automatic operation and maintenance systems are used to realize rapid service deployment, elastic scaling and intelligent operation and maintenance; meanwhile, data is processed through semantic analysis and format conversion, and security is guaranteed through multi-level identity authentication and encryption.

The application discloses a time slot scheduling system and method, wherein the system comprises a time slot management module and a weighted polling scheduler; the time slot management module is used for determining the data interaction path delay between a master device and a slave device based on the time sequence delay of a hardware unit in the data transmission path between the master device and the slave device; the data interaction path delay between the master device and each slave device is used for performing the arbitration of the weighted polling scheduling authority, generating a weighted polling request scheduling signal based on the arbitration result, and sending the weighted polling request scheduling signal to the weighted polling scheduler; the weighted polling scheduler is used for polling and selecting the slave device command queue participating in the weighted polling scheduling, and determining the arbitrated slave device command queue; the application can dynamically control the scheduling time slot of the slave device based on the write command length and the data interaction path delay between the master device and the slave device, and inhibit the local congestion caused by the continuous data flow injection into the on-chip interconnection network.

The application discloses a kind of intelligent method and system based on flow, belong to distributed computing and resource scheduling technical field.The method includes: in gateway level, the business flow data of each center system is collected in real time;Maximum QPS performance benchmark based on pressure test is established for each center application;Real-time load rate of each center is dynamically calculated based on sliding time window algorithm;When load rate meets preset double threshold trigger condition, combined with performance benchmark and real-time load rate, the target allocation proportion of high-load center to low-load center is calculated by the algorithm of proportion distribution according to system processing capacity;Finally, cross-center system flow allocation operation is executed through dynamic weighted round robin mechanism.The application realizes the second-level monitoring of flow, intelligent decision and smooth scheduling, effectively solves the problems of response lag, dependence on manual intervention and unreasonable resource allocation in the prior art, significantly improves resource utilization and system stability.

The invention discloses a task processing method and device and electronic equipment, and relates to the technical field of artificial intelligence, and the method comprises the steps: determining the real-time performance information of a plurality of server nodes, and calculating the dynamic load factors of the server nodes according to the real-time performance information, the dynamic load factors being used for reflecting the current task processing capability of the server nodes. A weight value is distributed to each server node, the initial weight value of each server node is determined based on the dynamic load factor, and the weight value of each server node is dynamically updated based on a smooth weighted polling algorithm. And a task processing request is sent, the task processing request is used for requesting a target server node to execute the to-be-processed task, and the target server node is the server node with the highest current weight value. By applying the technical scheme of the invention, the large-scale multi-modal task processing efficiency can be improved, resource redundancy is reduced, and the accuracy of a processing result is enhanced.

The application discloses a large file cross-unit transmission method based on a business cooperation scene, relates to the technical field of cross-organization data transmission and resource scheduling, and takes a business context as a core to map an event type, an emergency level and an SLA deadline time into a calculable priority, and is linked with real-time network measurement results, so that scheduling no longer stays in static file attributes, but can instantly improve the service share of a key request according to events such as out-of-stock early warning and promotion start, thereby reducing the probability that a high-emergency task is blocked by regular transmission. By introducing unit-level quotas, admission and interruptible preemption on the side of a bottleneck link, long-term occupation caused by first-come-first-served is avoided, basic fairness between different units is ensured, and bandwidth can be quickly transferred to more urgent flows at a business peak moment. Weighted round-robin weights are dynamically updated according to the emergency degree, the SLA urgency degree and the quota surplus degree, hysteresis and minimum interval control are superimposed, and the agility of scheduling is improved.

This application provides a collaborative management method and system for RPA robot clusters, relating to the technical field of RPA robot management. The method includes: if a task execution request exists, parsing and generating a target task and performing task judgment; if not, re-judging after a preset interval; if the target task contains multiple subtasks, adding tags to the subtasks and matching them with the target RPA robot cluster; otherwise, adding tags to the task and matching it; selecting a target robot using a weighted round-robin algorithm based on the running data of each RPA robot in the target RPA robot cluster, and sending the task or subtask to the target robot. This application enables more refined task division and matching, improving the efficiency and accuracy of task execution, while fully utilizing the resources of the RPA robot cluster.

The embodiment of the invention provides a data packet forwarding method, gateway equipment, a storage medium and a program, the gateway equipment comprises a plurality of processors and a network card provided with a programmable chip, and the plurality of processors comprise a control processor and a plurality of data processors. The control processor sends a scheduling mode setting instruction including a first scheduling mode to the programmable chip, determines weighted polling values of the plurality of data processors according to the loads of the plurality of data processors, and sends the weighted polling values of the plurality of data processors to the programmable chip. The programmable chip receives a data packet sent by the source end equipment, determines a first data processor of which the polling frequency reaches a corresponding weighted polling value based on a first scheduling mode and the weighted polling values of the plurality of data processors, distributes the data packet to the first data processor for processing, receives a processed data packet fed back by the first data processor, and sends the processed data packet to the source end equipment; and sending the processed data packet to a destination end device. Through the scheme, the forwarding delay of the data packet can be reduced.

The invention discloses a large file cross-unit transmission method based on a service collaboration scene, and relates to the technical field of cross-organization data transmission and resource scheduling, in the invention, a service context is taken as a core, an event type, an emergency level and an SLA deadline are mapped into computable priorities, and linkage with a real-time network measurement result is carried out; scheduling does not stay at a static file attribute any more, the service share of a key request can be instantly improved along with out-of-stock early warning, promotion starting and other events, and therefore the probability that a high-emergency-degree task is blocked by conventional transmission is reduced; by introducing unit-level quota, admission and interruptible preemption at a bottleneck link side, long-term occupation caused by first arrival and first arrival is avoided, basic fairness among different units is ensured, and bandwidth can be quickly transferred to a more urgent flow at a service peak moment; the weighted polling weight is dynamically updated along with the urgency degree, the SLA urgency degree and the quota redundancy, hysteresis and minimum interval control are overlaid, and scheduling agility is improved.

The application provides a short video full-process customer acquisition system based on agent cooperation, and relates to the technical field of artificial intelligence. The input layer of the system is used for receiving task data input by a user. The scheduling layer adopts an agent scheduling engine. The agent scheduling engine is used for determining at least one subtask and a subtask execution sequence according to the task data, and distributing the subtask based on a publish / subscribe mode and allocating computing power resources based on a weighted round robin algorithm. The function layer includes an account positioning module, a benchmark account disassembly module, an account foundation building module, a topic library building module, a video creation module and a clipping and publishing module which are independently encapsulated in the form of agents. Based on the subtask execution sequence and the computing power resources, the account positioning module, the benchmark account disassembly module, the account foundation building module, the topic library building module, the video creation module and the clipping and publishing module execute corresponding subtasks. The application can automatically create short videos to acquire customers and improve efficiency.

The invention relates to a telemedicine-oriented 5G multi-link network dynamic adaptation method and device, and the method comprises the steps: building a mathematical model based on a WRR weighted polling scheduler, putting a to-be-sent message into a to-be-sent cache stack, enabling a pointer of the stack reading the message to correspond to a first message, creating two WRR queue weights to correspond to two scheduling message queues, and enabling the WRR queue weights to correspond to two scheduling message queues; a message in a cache stack is subjected to polling scheduling and split into two message queues, the two message queues correspond to two transmission links, a message reading pointer of each transmission link points to the head of the message queue, and by integrating a 5G multi-link aggregation algorithm, the communication bandwidth of a single service is increased by two times. According to the technical scheme, the bandwidth requirements of the upper and lower networks of the medical equipment in a mobile environment are met, SIM cards of different operators can be inserted, the problems that the bandwidth is limited and the time delay is too large in real-time data communication interaction faced by the current medical terminal equipment are solved, and a new technical scheme is provided for the development of remote medical services.

Provided in the embodiments of the present disclosure are a data packet forwarding method, a gateway device, a storage medium and a program. The gateway device comprises a plurality of processors and a network card provided with a programmable chip, wherein the plurality of processors comprise a control processor and a plurality of data processors. The control processor sends to the programmable chip a scheduling mode setting instruction including a first scheduling mode, determines weighted round-robin values of the plurality of data processors on the basis of loads of the plurality of data processors, and sends the weighted round-robin values of the plurality of data processors to the programmable chip. The programmable chip receives a data packet sent by a source-end device, and determines, on the basis of the first scheduling mode and the weighted round-robin values of the plurality of data processors, a first data processor, the number of round-robin cycles of which has reached a corresponding weighted round-robin value, and the programmable chip then distributes the data packet to the first data processor for processing, receives the processed data packet fed back by the first data processor, and sends the processed data packet to a destination-end device. The solution can reduce the forwarding delay of data packets.

The invention provides a channel selection method and a related device, and relates to the technical field of communication. The method comprises the following steps: under the condition that first data to be transmitted is detected, selecting a first network channel from a plurality of network channels based on a smooth weighted polling algorithm to carry out data transmission; in the process of transmitting the first to-be-transmitted data through the first network channel, if it is detected that the first network channel is abnormal, reducing the effective weight of the first network channel; and when the first to-be-transmitted data fails to be transmitted or second to-be-transmitted data is detected, network channel selection is carried out based on the smooth weighted polling algorithm and the adjusted effective weights of the plurality of network channels. According to the technical scheme provided by the invention, the problem of how to better realize load balancing for a plurality of network channels in the prior art can be solved.

This invention discloses a joint deployment optimization method for hybrid expert model inference in heterogeneous GPU environments. The method first uses offline trajectory statistics to determine expert activation frequencies and collects resource profiles of heterogeneous devices. Then, it constructs a joint optimization model with the objective of minimizing the weighted sum of inference latency surrogate terms and quantization error penalty terms. Under memory constraints, a pruned greedy search algorithm is used to solve for the placement location, number, and quantization accuracy of each expert's replicas. During the online inference phase, a routing mechanism combining local priority and load-aware weighted round-robin is used for dynamic token distribution. This invention can simultaneously address inference latency, memory usage, and accuracy preservation in heterogeneous GPU environments, improving the processing efficiency of popular experts, reducing cross-device communication overhead and slow device trailing effects, and is suitable for distributed large-scale model inference deployment scenarios.