9 results about "Fair scheduling" patented technology

The invention relates to a shared GPU resource fair scheduling method and system based on multi-dimensional priority, and belongs to the field of cloud computing and container orchestration.The method comprises the steps that pooling and virtualization segmentation are conducted on physical GPU resources in a cluster, and a unified GPU resource pool is formed; calculating the comprehensive priority weight of the task according to the user level, the task type and the service quality in combination with a multi-dimensional priority model; a scheduling decision is made based on a weighted dominant resource fair algorithm, task priority weights and dominant resource proportions are comprehensively considered, and fair scheduling and optimal distribution of GPU resources are achieved; through a resource monitoring and feedback mechanism, the GPU utilization rate, the task execution state and the resource allocation result are monitored in real time, priority parameters and scheduling weights are dynamically adjusted, and finally fair scheduling of shared GPU resources is achieved. According to the method, intelligent allocation, resource sharing and fair scheduling of tasks can be realized in a GPU virtualization environment, and the scheduling efficiency and the overall cluster utilization rate are improved.

The invention discloses a fair resource management method and system for multi-user shared large language model reasoning, and the method is characterized in that the method employs a cold data recognition elimination mechanism and a fair cache distribution mechanism, carries out the unified quantification of the resource consumption of a request in GPU calculation, decoding and cache transmission, and carries out the fair scheduling based on the accumulated CPI value of each user; the system comprises a user request management module, a CPI fair scheduling module, a model reasoning execution module, a resource monitoring module and a cache management module. Compared with the prior art, the method has the advantages that users can share computing and caching resources fairly, efficient operation performance of the system is guaranteed, users with high cache hit rate are prevented from excessively occupying GPU execution opportunities due to low apparent computing overhead, the overall reasoning efficiency and the cache hit rate of the system are remarkably improved, and the user experience is improved. Unified and fair distribution of computing resources and cache resources is achieved, the effect is prominent especially in a multi-tenant high-load scene, and the method has good application prospects and commercial development value.

The invention relates to the technical field of computer system resource scheduling, in particular to a dynamic fair scheduling method and system and a computer readable storage medium. According to the method, when a producer generates a new message, the message is stored in a corresponding exclusive message queue; then, whether the identifier corresponding to the producer exists or not is inquired in a polling queue, if not, the identifier corresponding to the producer is added to the tail of the polling queue, and the polling queue is used for registering the identifier of the producer with the message to be processed currently; and finally, during message scheduling, taking out a producer identifier from the queue head of the polling queue, taking out a message from the exclusive message queue corresponding to the producer for processing, and resetting the producer identifier into the queue tail of the polling queue if the exclusive message queue is not empty after the processing is completed. According to the invention, through a cooperation mechanism of the dynamic polling queue and the tenant exclusive queue, self-adaptive resource scheduling without physical isolation is realized.

The invention discloses a deep model resource fair scheduling method for multi-task service quality assurance in a cloud-edge collaborative environment. The method comprises the following steps: S1, a task request sensing module; s2, an online model evaluation module; s3, a model fair selection module; s4, an isomorphic model migration module; s5, a heterogeneous model collaborative scheduling module; and S6, a service quality fairness guarantee module. The method has high engineering practical value, and can be applied to various real-time reasoning scenes such as intelligent medical treatment, intelligent security and protection, industrial internet and the like.

The invention discloses a task request scheduling method and device, equipment, a medium and a program product. In response to the task scheduling instruction, a plurality of first users whose task queues are not empty are determined, and one user corresponds to one task queue; in the dynamic priorities of the users participating in scheduling, the dynamic priority of the first user is obtained, the dynamic priority is generated based on the static basic weights of the users and the dynamic credit points of the users, the dynamic priority and the static basic weights are in positive correlation, and the dynamic priority and the dynamic credit points are in positive correlation; in the plurality of first users, determining the first user corresponding to the highest dynamic priority as a target user; and executing the target task request in the task queue corresponding to the target user. According to the embodiment of the invention, fair scheduling in a user dimension in a real sense can be realized, and system stability and service quality are guaranteed.

The invention relates to the technical field of wireless optical communication, and discloses a hybrid fair scheduling method and system based on a wireless optical communication network. Time slots are dynamically distributed by comprehensively considering user channel capacity and throughput requirements on the basis of different throughput and time delay requirements of interactive users and transmission users; meanwhile, a scheduling structure is reconstructed according to different service types; for the interactive users, through an iteration mechanism of preferentially allocating a time slot to the interactive user with the maximum service guiding priority, and through presetting the minimum reallocation time slot, resources are rapidly obtained, it is ensured that the time slot to be recombined is preferentially consumed, and time delay accumulation caused by waiting for the long time slot of the transmission type user is avoided; and for the transmission type users, after the time slots of the interactive type users are distributed, the time slots are distributed to the transmission type users in a centralized manner according to the service oriented priority, so that the throughput loss caused by context switching is reduced, the requirement of completing large-volume data transmission within several seconds is met, and the network real-time performance and the resource utilization rate are integrally improved.

When scheduling a task to be executed by a target virtual machine, a scheduler first obtains at least one top-priority task in a to-be-executed task queue of the first target virtual machine to first ensure that the top-priority task is preferentially executed, determines a first task with a minimum virtual runtime from the at least one top-priority task, and controls the target virtual machine to execute the first task. This ensures fair task scheduling of at least one task with the same priority. The scheduler can ensure both preferential scheduling of a high-priority task and fair scheduling of tasks with the same priority when scheduling the task on the first target virtual machine.

The invention discloses an RISC-V CLIC pre-arbitration method and system based on hierarchical priority and cycle perception, and belongs to the technical field of embedded systems. The method comprises the following steps of: firstly, dividing an interrupt source into a hard real-time domain L1 and a fair scheduling domain L2 according to real-time and security requirements, directly preempting the interrupt source L1 in a hard manner according to original semantics, and ensuring a deterministic response of a security-level task; the L2 interrupt source can select three dynamic arbitration logics, and flexible scheduling is realized through parameter configuration in combination with token bucket current limiting, waiting time compensation and a period sensing mechanism. Two implementation modes of an external preposed arbitration module or an internal integrated dynamic arbitration sub-module are supported, a shift addition lightweight hardware design is adopted, low overhead and high real-time performance are both considered, and compatibility with a CLIC interface is achieved; and the method can be applied to scenes such as a vehicle regulation MCU (Microprogrammed Control Unit), industrial control, an edge computing gateway and the like.