A pipeline and industry business optical interconnection delay optimization scheduling method
By constructing a latency optimization scheduling method for optical interconnection of pipelined parallel services, the problem of uncoordinated resource allocation in existing technologies is solved, achieving efficient communication scheduling and resource utilization, reducing communication latency and pipeline bubbles, and meeting the training requirements of large-scale deep learning models.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUZHOU UNIV
- Filing Date
- 2026-05-09
- Publication Date
- 2026-06-05
AI Technical Summary
Existing optical interconnect networks neglect the execution dependencies between services and the correlation of resource allocation in pipelined parallel communication, resulting in low scheduling efficiency, insufficient resource utilization, excessive communication latency and pipeline bubbles, which cannot meet the requirements of efficient communication.
We construct pipeline constraints, service routing and scheduling constraints, service integrity transmission constraints, routing constraints, and port constraints. We optimize the joint allocation of links, time slots, and ports through a scheduling objective function. We use the maximum bottleneck pipeline first algorithm and Dijkstra's algorithm for resource allocation to ensure the coordinated optimization of service dependencies and resource competition.
It effectively improves scheduling efficiency and resource utilization, reduces communication latency and pipeline bubbles, and meets the high-efficiency communication requirements of large-scale deep learning pipeline parallel training.
Smart Images
Figure CN122160658A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of optical communication technology, and in particular to a method for optimizing the scheduling of optical interconnect delays for pipelined parallel services. Background Technology
[0002] In the field of artificial intelligence, the training of large-scale deep learning models is the core driving force for continuous technological progress. As the scale of deep learning model parameters rapidly evolves from billions to trillions, traditional single-machine, single-card training and simple data parallel training methods are no longer sufficient to meet the training needs of large-scale models. Pipeline parallelism (PP) technology, by dividing the neural network model into layers and deploying them to different computing devices to perform computational tasks, provides an important approach for the efficient training of ultra-large-scale deep learning models. However, limited by the memory capacity of a single accelerator, ultra-large-scale deep learning models cannot be stored and trained on a single accelerator. The model must be split and deployed to multiple accelerators for collaborative training. With the continuous development of model parallelism technology, communication losses and overhead between devices have become increasingly prominent, becoming a key bottleneck restricting the improvement of training efficiency for ultra-large-scale models.
[0003] Existing technologies mainly focus on routing and scheduling optimization in Dragonfly optical interconnect networks, including the following methods: 1. Static routing-based scheduling methods: Fixed optical interconnect path topologies are pre-constructed before the training task starts. Static communication paths are set according to the inter-layer dependencies of the model, and a fixed time slot allocation mechanism is used to complete inter-device communication scheduling. 2. Shortest-path-first (SP) scheduling algorithm: Parallel pipeline tasks are sorted according to the amount of communication data, prioritizing inter-layer communication tasks with smaller data volumes to reduce device queuing length and overall waiting time. 3. Shortest-path-first (SP) strategy: During communication, the shortest-path algorithm is used to prioritize PP services with the shortest communication time, aiming to minimize end-to-end transmission latency. However, these methods all ignore the fact that multiple tasks in pipelined parallel training do not run independently, but have strict inter-layer dependencies and execution constraints. Optimization through static communication paths, fixed time slot allocation, or a single dimension (such as shortest service priority scheduling or shortest path priority scheduling strategies) can easily lead to problems such as link resource contention, port scheduling conflicts, increased latency, and expanded transmission bubbles. These methods cannot meet the core requirements of low latency, low bubble size, and high resource utilization for pipelined parallel communication services.
[0004] In summary, existing routing and scheduling optimization methods in optical interconnect networks ignore the execution dependencies between services and the correlation of resource allocation during pipelined parallel communication. They cannot achieve coordinated optimization of resources such as links, time slots, and ports, and it is difficult to balance the adaptability of multi-task scheduling with pipeline constraints. Ultimately, this leads to problems such as low scheduling efficiency, insufficient resource utilization, excessive communication latency and pipeline bubbles, and failure to meet the requirements of efficient communication. Summary of the Invention
[0005] Therefore, the technical problem to be solved by the present invention is to overcome the fact that the routing and scheduling optimization methods in the existing optical interconnect network ignore the execution dependency between services and the correlation of resource allocation when pipelined parallel communication, and cannot achieve the coordinated optimization of resources such as links, time slots, and ports. It is also difficult to balance the adaptability of multi-task scheduling and pipeline constraints, which ultimately leads to low scheduling efficiency, insufficient resource utilization, communication latency and excessive pipeline bubbles, and failure to meet the requirements of efficient communication.
[0006] To address the aforementioned technical problems, this invention provides a method for optimizing the scheduling of optical interconnect latency for pipelined parallel services, comprising: Obtain the service node pairs, the number of services between each service node pair, the execution order of each service, and the time requirement of each service in the service requirement set of the MEMS-based optical switching network and pipeline parallel service. Based on the execution order, time requirements, and start time of each service, pipeline constraints are constructed; based on the fact that the number of services scheduled between optical routing nodes in the optical switching network for each service node pair is equal to the number of services between each service node pair, service number routing scheduling constraints are constructed. Based on the start time, time requirement, earliest scheduling time, and total time slots allocated in the optical switching network for each service, complete service transmission constraints are constructed for each service; based on the continuity of time slots allocated in the optical switching network for each service, service continuity constraints are constructed for each service. For services sharing the same link in an optical switching network, routing constraints are constructed based on execution order and start time; for services sharing the same MEMS port in an optical switching network, port constraints are constructed based on execution order, start time, and MEMS port reconfiguration delay time. A scheduling objective function is constructed with the goal of minimizing the completion time of all business processes and pipeline bubbles. Solving the scheduling objective function, pipeline constraints, service routing scheduling constraints, service integrity transmission constraints, service continuity constraints, routing constraints, and port constraints yields the links and time slots allocated to each service in the optical switching network.
[0007] Preferably, pipeline constraints are constructed based on the execution order, time requirements, and start time of each service, including: constructing pipeline constraints such that the start time of a service executed later is greater than or equal to the sum of the start time and time requirements of a service executed earlier.
[0008] Preferably, the pipeline constraint is expressed as: , in, Indicates business node pair Business operations executed later The start time; Indicates business node pair Prior execution of business between The start time; Indicates business Time requirements; Service routing scheduling constraints are expressed as follows: , in, Indicates the judgment of business node pair A binary variable indicating whether the service node d between them has completed scheduling; if the service node... The business d between them is scheduled. If the business node is The service d between them was not scheduled. ; Indicates business node pair The number of transactions between them; Indicates business node pair Number of available paths in an optical switching network.
[0009] Preferably, service integrity constraints for each service are constructed based on its start time, time requirement, earliest scheduling time, and total time slots allocated in the optical switching network; service continuity constraints for each service are constructed based on the continuity of time slots allocated in the optical switching network, including: The complete transmission constraints for each service are constructed based on the premise that the sum of the start time and time demand of each service is less than or equal to the total time slot allocated to each service in the optical switching network, and the start time of each service is greater than or equal to the earliest scheduling time of each service. The service continuity constraints for each service are constructed based on the following: each time slot in the total time slots allocated to each service in the optical switching network is greater than or equal to the start time of each service and less than or equal to the difference between the sum of the start time and the time requirement of each service and 1; and the number of time slots in the total time slots allocated to each service in the optical switching network is equal to the time requirement of each service.
[0010] Preferably, the service complete transmission constraint is expressed as: , , in, Indicates business node pair The start time of business d between; Indicates business node pair The time requirement of the business d between; Indicates business node pair The total time slots allocated to service d in the optical switching network between them; Indicates business node pair The earliest scheduling time of service d between them; Business continuity constraints are expressed as follows: , , , in, Indicates the available time slots in an optical switching network; This indicates whether the available time slot t has been allocated to the service node pair. The binary variable d between services, when the available time slot t is allocated to the service node pair When business d is between, When the available time slot t is not allocated to the service node pair When business d is between, ; This indicates a maximum value.
[0011] Preferably, for services sharing the same link in the optical switching network, routing constraints are constructed based on execution order and start time; for services sharing the same MEMS port in the optical switching network, port constraints are constructed based on execution order, start time, and MEMS port reconfiguration delay time, including: For any two services that share the same link in an optical switching network, a routing constraint is constructed based on the premise that the start time of the service executed later is greater than or equal to the completion time of the service executed earlier. For any two services sharing the same MEMS port in an optical switching network, a port constraint is constructed whereby the start time of the later-executed service is greater than or equal to the sum of the completion time of the earlier-executed service and the reconfiguration delay time of the MEMS port.
[0012] Preferably, the routing constraints are expressed as: , in, Indicates business node pair Prior execution of business between The start time; Indicates business node pair Business operations executed later The start time; Indicates business Time requirements; Indicates a maximum value; Indicates judgment business and business In an optical switching network, are binary variables shared on the same link? If the service... and business In an optical switching network, sharing the same link means... If business and business In optical switching networks, if the same link is not shared, then ; Indicates judgment business A binary variable indicating whether scheduling is complete; if the business... Complete scheduling. If business Scheduling not completed ; Indicates judgment business and business A binary variable representing the execution order, if the business... Execute first, business Executed later, then If business Execute first, business Executed later, then ; Port constraints are represented as: , in, Indicates the reconfiguration delay time of the MEMS port; Indicates judgment business and business In an optical switching network, are binary variables shared for the same MEMS port, if the service... and business In an optical switching network, sharing the same MEMS port means... If business and business In an optical switching network, if the same MEMS port is not shared, then ; Indicates judgment business A binary variable indicating whether scheduling is complete; if the business... Complete scheduling. If business Scheduling not completed .
[0013] Preferably, the scheduling objective function Represented as: , , , in, Indicates the completion time of all business transactions; This indicates the weight of the completion time for all business transactions; Indicates the weight of bubbles in the pipeline; Indicates business node pair The start time of business d between; Indicates business node pair The time requirement of the business d between; This represents bubbles in the production line; Indicates the number of business node pairs; Indicates business node pair The number of transactions between them; Indicates business node pair The earliest scheduling time of service d between them.
[0014] Preferably, the scheduling objective function, pipeline constraints, service number routing scheduling constraints, service integrity transmission constraints, service continuity constraints, routing constraints, and port constraints are solved to obtain the links and time slots allocated to each service in the optical switching network, including: Based on the bottleneck severity index of each service in all available candidate paths in the optical switching network, the bottleneck severity of each service is ranked. The allocation order of each business is obtained according to the bottleneck severity from high to low. The Dijkstra algorithm is used to allocate links and time slots for each service in the order of allocation. The results of link and time slot allocation for each service are obtained when the scheduling objective function, pipeline constraints, service number routing scheduling constraints, service integrity transmission constraints, service continuity constraints, routing constraints, port constraints, total delay minimization constraints, and transmission bubble minimization constraints are satisfied.
[0015] Preferably, the bottleneck index of a candidate path is the minimum bandwidth or maximum link load of all links on that candidate path.
[0016] The optical interconnect delay optimization scheduling method for pipelined parallel services provided in this application has the following advantages: 1. In the modeling phase, pipeline constraints are directly constructed based on the execution order between parallel pipeline services. The inherent sequential dependencies between services are embedded into the scheduling core. Service routing scheduling constraints ensure that all services are scheduled. Service continuity and integrity constraints limit the time slots allocated to each service to be continuous and uninterrupted, and the time slot length can meet the service execution time slot requirements. Since different services may share the same links and MEMS ports during pipeline parallel communication, this application constructs routing constraints to achieve link resource mutual exclusion and timing adaptation, and constructs port constraints to achieve MEMS port exclusivity and timing, as well as port reconfiguration delay. Late adaptation treats links, time slots, and ports as a whole for joint allocation. Through multiple constraints, it avoids port contention and link conflicts during resource scheduling, thereby achieving collaborative scheduling of multiple types of resources. Finally, it takes the total completion time and pipeline bubble as optimization objectives simultaneously, minimizing the bubble and communication latency while satisfying all constraints. It fully considers the dependencies and resource contention among multiple services during pipeline parallel communication, and can fully balance the adaptability of multi-task scheduling and pipeline constraints, thereby effectively improving scheduling efficiency and resource utilization, reducing communication latency and pipeline bubble, and meeting the high-efficiency communication requirements of large-scale deep learning pipeline parallel training. 2. When solving scheduling objectives and constraints to allocate service resources, the maximum bottleneck pipeline first algorithm is used to sort services based on their bottleneck link indicators on candidate paths. Services with higher bottleneck levels are prioritized for critical link resources, preventing them from being preempted and thus alleviating overall link congestion and reducing resource contention and conflicts. After determining the service scheduling order, path calculations are performed for each service sequentially. Dijkstra's algorithm is first used to find the shortest path that meets the constraints under the current network conditions, and then links and time slots are allocated for that service. This allows for dynamic adaptation to changes in service execution order, link load, and MEMS port reconfiguration latency. Even in complex multi-task parallel scenarios, stable and low-conflict scheduling results are maintained, better meeting the requirements of large-scale deep learning pipeline parallel training for low latency and high-reliability communication. Attached Figure Description
[0017] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 Flowchart of the optical interconnect latency optimization scheduling method for pipelined parallel services provided in this application; Figure 2 The Dragonfly optical switching network topology diagram provided for embodiments of this application. Detailed Implementation
[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.
[0019] Please see Figure 1 , Figure 1 The diagram shows the flowchart of the optical interconnect latency optimization scheduling method for pipelined parallel services provided in this application. The method specifically includes S10~S60: S10: Obtain the service node pairs in the pipelined parallel service requirement set of the MEMS-based optical switching network, the number of services between each service node pair, the execution order of each service, and the time requirement of each service.
[0020] Specifically, MEMS-based optical switching networks Within each group, multiple optical routing nodes (Routers) based on MEMS optical switching matrices are configured to enable optical connections between servers and between groups. This indicates the number of terminal servers connected to each optical routing node. This indicates the number of optical routing nodes contained in each group of the network. This indicates the number of ports used for inter-group connections by each optical routing node. Number of groups = a*h+1.
[0021] S20: Based on the execution order, time requirements, and start time of each service, construct pipeline constraints; based on the fact that the number of services scheduled between optical routing nodes in the optical switching network for each service node pair is equal to the number of services between each service node pair, construct service number routing scheduling constraints.
[0022] Furthermore, in the pipelined business set PP, the business executed later must follow the business executed earlier, thus realizing the sequential characteristic of the pipelined business. Therefore, pipeline constraints are constructed based on the premise that the start time of the business executed later is greater than or equal to the sum of the start time of the business executed earlier and the time requirement.
[0023] Specifically, pipeline constraints are represented as: , in, Indicates business node pair Business operations executed later The start time; Indicates business node pair Prior execution of business between The start time; Indicates business Time requirements; Service routing scheduling constraints are expressed as follows: , in, Indicates the judgment of business node pair A binary variable indicating whether the service node d between them has completed scheduling; if the service node... The business d between them is scheduled. If the business node is The service d between them was not scheduled. ; Indicates business node pair The number of transactions between them; Indicates business node pair Number of available paths in an optical switching network.
[0024] S30: Based on the start time, time requirement, earliest scheduling time, and total time slots allocated in the optical switching network for each service, construct the service integrity transmission constraints for each service; based on the continuity of time slots allocated in the optical switching network for each service, construct the service continuity constraints for each service.
[0025] Furthermore, step S30 specifically includes S300~S301: S300: Construct complete transmission constraints for each service based on the premise that the sum of the start time and time demand of each service is less than or equal to the total time slot allocated to each service in the optical switching network, and that the start time of each service is greater than or equal to the earliest scheduling time of each service.
[0026] S301: Construct service continuity constraints for each service based on the following: each time slot in the total time slots allocated to each service in the optical switching network is greater than or equal to the start time of each service and less than or equal to the difference between the sum of the start time of each service and the time requirement and 1; and the number of time slots in the total time slots allocated to each service in the optical switching network is equal to the time requirement of each service.
[0027] Specifically, the complete service transmission constraint is expressed as follows: , , in, Indicates business node pair The start time of business d between; Indicates business node pair The time requirement of the business d between; Indicates business node pair The total time slots allocated to service d in the optical switching network between them; Indicates business node pair The earliest scheduling time of service d between them; Business continuity constraints are expressed as follows: , , , in, Indicates the available time slots in an optical switching network; This indicates whether the available time slot t has been allocated to the service node pair. The binary variable d between services, when the available time slot t is allocated to the service node pair When business d is between, When the available time slot t is not allocated to the service node pair When business d is between, ; This indicates a maximum value.
[0028] It should be noted that time slot t was not allocated to the service node pair When there is a service d between nodes, the time slot t does not need to be greater than or equal to the service node pair. The start time of business d between them. , When the value reaches its maximum, the business continuity constraint is... It is turned off by the maximum value M.
[0029] By limiting the upper and lower limits of the time slots for service scheduling, it is ensured that the service scheduling does not start earlier than the earliest allowed time within the total time slots. At the same time, the start and end ranges of service time slot allocation, the total number of time slots occupied by the service and the service time requirements are limited to be consistent, ensuring that the service time slot occupancy is continuous and sufficient.
[0030] S40: For services sharing the same link in the optical switching network, construct routing constraints based on execution order and start time; for services sharing the same MEMS port in the optical switching network, construct port constraints based on execution order, start time, and MEMS port reconfiguration delay time.
[0031] Further, step S40 includes S400~S401: S400: For any two services that share the same link in an optical switching network, a routing constraint is constructed based on the fact that the start time of the service executed later is greater than or equal to the completion time of the service executed earlier.
[0032] S401: For any two services sharing the same MEMS port in an optical switching network, construct port constraints such that the start time of the later-executed service is greater than or equal to the sum of the completion time of the earlier-executed service and the reconfiguration delay time of the MEMS port.
[0033] Specifically, the routing constraints are expressed as: , in, Indicates business node pair Prior execution of business between The start time; Indicates business node pair Business operations executed later The start time; Indicates business Time requirements; Indicates a maximum value; Indicates judgment business and business In an optical switching network, are binary variables shared on the same link? If the service... and business In an optical switching network, sharing the same link means... If business and business In optical switching networks, if the same link is not shared, then ; Indicates judgment business A binary variable indicating whether scheduling is complete; if the business... Complete scheduling. If business Scheduling not completed ; Indicates judgment business and business A binary variable representing the execution order, if the business... Execute first, business Executed later, then If business Execute first, business Executed later, then ; Port constraints are represented as: , in, Indicates the reconfiguration delay time of the MEMS port; Indicates judgment business and business In an optical switching network, are binary variables shared for the same MEMS port, if the service... and business In an optical switching network, sharing the same MEMS port means... If business and business In an optical switching network, if the same MEMS port is not shared, then ; Indicates judgment business A binary variable indicating whether scheduling is complete; if the business... Complete scheduling. If business Scheduling not completed .
[0034] It should be noted that in routing constraints and port constraints, the maximum value M plays the same role as in service continuity constraints. For routing constraints, if the service... and business They did not share the same links, therefore, the services and business It also does not need to satisfy routing constraints. , If the value is a maximum, the routing constraint is disabled by the maximum value M; if the service and business Sharing the same link, but the business The scheduling was not completed, and the business... and business It also does not need to satisfy routing constraints. , The maximum value M disables routing constraints; due to the service... For the business to be executed first, the business To execute business logic later, if the business logic is... Execute first, business Executed later, then , When the value is at its maximum, the execution order of the two services no longer satisfies the pipeline constraint, so the routing constraint is directly turned off by the maximum value M.
[0035] This application avoids link conflicts by using routing constraints, which restricts the execution of two services on a shared link to a specific order to avoid link resource conflicts. At the same time, it restricts the execution of services that are executed later on a shared MEMS port to wait for a specified reconfiguration time slot after the execution of services that were executed earlier to start execution, thereby avoiding port resource conflicts and adapting to hardware reconfiguration features.
[0036] S50: Construct a scheduling objective function with the goal of minimizing the completion time of all business processes and pipeline bubbles.
[0037] Specifically, the scheduling objective function Represented as: , , , in, Indicates the completion time of all business transactions; This indicates the weight of the completion time for all business transactions; Indicates the weight of bubbles in the pipeline; Indicates business node pair The start time of business d between; Indicates business node pair The time requirement of the business d between; This represents bubbles in the production line; Indicates the number of business node pairs; Indicates business node pair The number of transactions between them; Indicates business node pair The earliest scheduling time of service d between them.
[0038] S60: Solve the scheduling objective function, pipeline constraints, service number routing scheduling constraints, service integrity transmission constraints, service continuity constraints, routing constraints, and port constraints to obtain the links and time slots allocated to each service in the optical switching network.
[0039] Due to the high complexity of the MILP model, AMPL cannot find its optimal solution. Furthermore, in scenarios where multiple pipelined parallel services run simultaneously, the communication data volume and resource consumption of different services vary. When multiple services share the same link, different service scheduling orders directly affect the pipeline's bubble size and the overall system communication completion time. Existing technologies have proposed the Shortest Pipeline Communication First (SP) algorithm. The SP algorithm sorts services based on their communication data volume or communication overhead, prioritizing services with lower communication demands to reduce the number of queued services in the network and lower the overall communication waiting time. However, the SP algorithm prioritizes only the volume of business communication data during sorting, completely ignoring bottleneck resources such as link bandwidth and load on candidate business paths. Even if a link is already under high load and has low remaining bandwidth, SP may still continue to allocate services to that link according to service length, further exacerbating link congestion. This leads to difficulties in routing subsequent services, an increase in bubbles, and a significant increase in overall communication completion time. Furthermore, in multi-pipeline parallel scenarios, some services have few candidate paths and scarce link resources, making them bottleneck services with high scheduling difficulty and a significant impact on overall performance. SP's priority scheduling of shorter services causes bottleneck services to be continuously postponed, gradually consuming their allocable links and time slots, ultimately resulting in no available paths and the forced insertion of numerous bubbles. Waiting for resources directly reduces the communication efficiency of the entire system. In addition, link resources in optical switching networks are exclusive, and the same link and time slot can only carry a single service. The SP algorithm lacks a mechanism to predict link bottlenecks, which can easily cause a large number of services to flood into a few popular links, resulting in resource conflicts and blockages. It is impossible to actively relieve congestion through scheduling order, which leads to an increase in pipeline bubbles and uncontrollable communication latency. Moreover, pipeline services are highly sensitive to the continuity of timing and scheduling order, and there are execution dependencies between services. SP only pursues the priority of local short services and does not balance the allocation of link resources from a global perspective. It is easy to increase the scheduling delay of long links and high-bottleneck services, which degrades the overall communication completion time and fails to achieve the coordinated optimization of total latency and pipeline bubbles.
[0040] This application designs another heuristic scheduling algorithm: Maximum Bottleneck First (MBF). It sorts services based on bottleneck link metrics (such as minimum bandwidth or maximum link load) on candidate paths, prioritizing the allocation of critical link resources to services with higher bottleneck levels to alleviate link congestion. After determining the service scheduling order, it calculates paths for each service sequentially. First, it uses Dijkstra's algorithm to find the shortest path that satisfies the constraints under the current network conditions and allocates links and time slots for that service. By repeating the above process, all services are allocated, and finally, the communication path of the service is output, along with the communication completion time and pipeline bubbles.
[0041] Specifically, step S60 includes S600~S602: S600: Based on the bottleneck severity index of each service in all available candidate paths in the optical switching network, the bottleneck severity of each service is ranked.
[0042] Optionally, the bottleneck index of a candidate path is the minimum bandwidth or maximum link load of all links on that candidate path.
[0043] For example, for all candidate paths of a service, if minimum bandwidth is used as the bottleneck level indicator, the minimum value among the minimum bandwidths of each candidate path is taken as the bottleneck level of the service; if maximum link load is used as the bottleneck level indicator, the maximum value among the maximum link loads of each candidate path is taken as the bottleneck level of the service. Furthermore, all services are sorted based on their bottleneck levels: when using minimum bandwidth as the indicator, they are sorted from smallest to largest minimum bandwidth; when using maximum link load as the indicator, they are sorted from largest to smallest maximum link load, thus obtaining the bottleneck level ranking of all services.
[0044] S601: The allocation order of each business is obtained according to the bottleneck level from high to low.
[0045] S602: The Dijkstra algorithm is used to allocate links and time slots for each service in the order of allocation, and the link and time slot allocation results for each service are obtained when the pipeline constraint, service number routing scheduling constraint, service integrity transmission constraint, service continuity constraint, routing constraint, port constraint, total delay minimization constraint and transmission bubble minimization constraint are met.
[0046] The following example illustrates the optical interconnect latency optimization scheduling method for pipelined parallel services: like Figure 2 The image shows the MEMS-based Dragonfly optical switching network constructed in this embodiment. This optical switching network consists of four groups. Each group contains 3 optical routing nodes, and each optical routing node connects to 3 server terminals, forming a data center network model architecture with a total of 36 servers and 72 GPU computing cards. At the same time, each group is allowed one Global inter-group link.
[0047] In this embodiment, the number of service pairs is selected every 5 from 5 to 20. The data size of each service is randomly generated in a uniform distribution between 1Gbps and 10Gbps. The duration of each service is taken as 1 to 5 port switching delays. The source and destination nodes for the service are also randomly selected from 36 servers by sequence number, while avoiding duplication of source and destination nodes.
[0048] All services are scheduled according to the strategy provided in this application. At the same time, the three different schemes are compared together, the service pipeline bubble and TCT under the three schemes are calculated, and finally the optimal service scheduling scheme and its performance indicators are selected.
[0049] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0050] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0051] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.
[0052] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0053] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.
Claims
1. A delay optimization scheduling method for optical interconnects in pipelined parallel services, characterized in that, include: Obtain the service node pairs, the number of services between each service node pair, the execution order of each service, and the time requirement of each service in the service requirement set of the MEMS-based optical switching network and pipeline parallel service. Based on the execution order, time requirements, and start time of each service, pipeline constraints are constructed; based on the fact that the number of services scheduled between optical routing nodes in the optical switching network for each service node pair is equal to the number of services between each service node pair, service number routing scheduling constraints are constructed. Based on the start time, time requirement, earliest scheduling time, and total time slots allocated in the optical switching network for each service, complete service transmission constraints are constructed for each service; based on the continuity of time slots allocated to each service in the optical switching network, service continuity constraints are constructed for each service. For services sharing the same link in an optical switching network, routing constraints are constructed based on execution order and start time; for services sharing the same MEMS port in an optical switching network, port constraints are constructed based on execution order, start time, and MEMS port reconfiguration delay time. A scheduling objective function is constructed with the goal of minimizing the completion time of all business processes and pipeline bubbles. Solving the scheduling objective function, pipeline constraints, service routing scheduling constraints, service integrity transmission constraints, service continuity constraints, routing constraints, and port constraints yields the links and time slots allocated to each service in the optical switching network.
2. The optical interconnect delay optimization scheduling method for pipelined parallel services according to claim 1, characterized in that, Based on the execution order, time requirements, and start time of each business, pipeline constraints are constructed, including: the start time of a business executed later is greater than or equal to the sum of the start time and time requirements of a business executed earlier.
3. The optical interconnect delay optimization scheduling method for pipelined parallel services according to claim 1, characterized in that, Pipeline constraints are represented as: ， in, Indicates business node pair Business operations executed later The start time; Indicates business node pair Prior execution of business between The start time; Indicates business Time requirements; Service routing scheduling constraints are expressed as follows: ， in, Indicates the judgment of business node pair A binary variable indicating whether the service node d between them has completed scheduling; if the service node... The business d between them is scheduled. If the business node is The service d between them was not scheduled. ; Indicates business node pair The number of transactions between them; Indicates business node pair Number of available paths in an optical switching network.
4. The optical interconnect delay optimization scheduling method for pipelined parallel services according to claim 1, characterized in that, Based on the start time, time requirement, earliest scheduling time, and total time slots allocated in the optical switching network for each service, complete service transmission constraints are constructed for each service; based on the continuity of time slots allocated to each service in the optical switching network, service continuity constraints are constructed for each service, including: The complete transmission constraints for each service are constructed based on the premise that the sum of the start time and time demand of each service is less than or equal to the total time slot allocated to each service in the optical switching network, and the start time of each service is greater than or equal to the earliest scheduling time of each service. The service continuity constraints for each service are constructed based on the following: each time slot in the total time slots allocated to each service in the optical switching network is greater than or equal to the start time of each service and less than or equal to the difference between the sum of the start time and the time requirement of each service and 1; and the number of time slots in the total time slots allocated to each service in the optical switching network is equal to the time requirement of each service.
5. The optical interconnect delay optimization scheduling method for pipelined parallel services according to claim 1, characterized in that, The complete transmission constraint of the service is represented as follows: ， ， in, Indicates business node pair The start time of business d between; Indicates business node pair The time requirement of the business d between; Indicates business node pair The total time slots allocated to service d in the optical switching network between them; Indicates business node pair The earliest scheduling time of service d between them; Business continuity constraints are expressed as follows: ， ， ， in, Indicates the available time slots in an optical switching network; This indicates whether the available time slot t has been allocated to the service node pair. The binary variable d between services, when the available time slot t is allocated to the service node pair When business d is between, When the available time slot t is not allocated to the service node pair When business d is between, ; This indicates a maximum value.
6. The optical interconnect delay optimization scheduling method for pipelined parallel services according to claim 4, characterized in that, For services sharing the same link in an optical switching network, routing constraints are constructed based on execution order and start time; for services sharing the same MEMS port in an optical switching network, port constraints are constructed based on execution order, start time, and MEMS port reconfiguration delay, including: For any two services that share the same link in an optical switching network, a routing constraint is constructed based on the premise that the start time of the service executed later is greater than or equal to the completion time of the service executed earlier. For any two services sharing the same MEMS port in an optical switching network, a port constraint is constructed whereby the start time of the later-executed service is greater than or equal to the sum of the completion time of the earlier-executed service and the reconfiguration delay time of the MEMS port.
7. The optical interconnect delay optimization scheduling method for pipelined parallel services according to claim 1, characterized in that, Routing constraints are expressed as follows: ， in, Indicates business node pair Prior execution of business between The start time; Indicates business node pair Business operations executed later The start time; Indicates business Time requirements; Indicates a maximum value; Indicates judgment business and business In an optical switching network, are binary variables shared on the same link? If the service... and business In an optical switching network, sharing the same link means... If business and business In optical switching networks, if the same link is not shared, then ; Indicates judgment business A binary variable indicating whether scheduling is complete; if the business... Complete scheduling. If business Scheduling not completed ; Indicates judgment business and business A binary variable representing the execution order, if the business... Execute first, business Executed later, then If business Execute first, business Executed later, then ; Port constraints are represented as: ， in, Indicates the reconfiguration delay time of the MEMS port; Indicates judgment business and business In an optical switching network, are binary variables shared for the same MEMS port, if the service... and business In an optical switching network, sharing the same MEMS port means... If business and business In an optical switching network, if the same MEMS port is not shared, then ; Indicates judgment business A binary variable indicating whether scheduling is complete; if the business... Complete scheduling. If business Scheduling not completed .
8. The optical interconnect delay optimization scheduling method for pipelined parallel services according to claim 1, characterized in that, Scheduling objective function Represented as: ， ， ， in, Indicates the completion time of all business transactions; This indicates the weight of the completion time for all business transactions; Indicates the weight of bubbles in the pipeline; Indicates business node pair The start time of business d between; Indicates business node pair The time requirement of the business d between; Indicates air bubbles in the production line; Indicates the number of business node pairs; Indicates business node pair The number of transactions between them; Indicates business node pair The earliest scheduling time of service d between them.
9. The optical interconnect delay optimization scheduling method for pipelined parallel services according to claim 1, characterized in that, Solving for the scheduling objective function, pipeline constraints, service routing scheduling constraints, service integrity transmission constraints, service continuity constraints, routing constraints, and port constraints yields the links and time slots allocated to each service in the optical switching network, including: Based on the bottleneck severity index of each service in all available candidate paths in the optical switching network, the bottleneck severity of each service is ranked. The allocation order of each business is obtained according to the bottleneck severity from high to low. The Dijkstra algorithm is used to allocate links and time slots for each service in the order of allocation. The results of link and time slot allocation for each service are obtained when the scheduling objective function, pipeline constraints, service number routing scheduling constraints, service integrity transmission constraints, service continuity constraints, routing constraints, port constraints, total delay minimization constraints, and transmission bubble minimization constraints are satisfied.
10. The optical interconnect delay optimization scheduling method for pipelined parallel services according to claim 9, characterized in that, The bottleneck index of a candidate path is the minimum bandwidth or maximum link load of all links on that candidate path.