A communication path selection method and system for multi-chip acceleration platform

Abstract

This invention discloses a communication path selection method and system for multi-chip acceleration platforms, relating to the field of IoT chip technology. It includes a request characterization and same-target fragment communication group identification module, used to collect communication requests to be sent from each source chip within a short acquisition window, generate request description items, and identify same-target fragment communication groups pointing to the same destination chip from the communication requests to be sent. A merging chip selection and main path planning module is used to determine candidate merging chips for the same-target fragment communication groups. This invention identifies same-target fragment communication groups pointing to the same destination chip within a short acquisition window, and then selects a merging chip for the same-target fragment communication group, allowing multiple source chips to merge short-range first, and then transmit them far along a main path with continuous idle time slot intervals in one go. Path jitter is suppressed through path reservation control messages and path holding mechanisms.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) chip technology, specifically to a communication path selection method and system for multi-chip acceleration platforms. Background Technology

[0002] A chip is a pre-manufactured functional die unit with specific functions that can work collaboratively with other functional dies through inter-die interconnects within its package. Multiple chips are combined to form a multi-chip acceleration platform. With the increasing trend of deploying multi-chip acceleration platforms in IoT terminals, edge intelligent devices, and automotive sensing devices, computing chips, storage chips, and I / O chips perform parameter transfer, activation transmission, result write-back, and control response through inter-die interconnects. Inter-die communication has become a critical component of the operational chain of edge intelligent systems. Recent publicly available research indicates that multi-chip platforms can support large-scale and sparse model inference at the edge, but they also face significant inter-die data flow challenges. Especially in scenarios with low-batch requests, short-response links, and multiple types of heterogeneous interconnects, the communication organization itself directly affects the throughput, latency, and resource utilization performance of edge multi-chip systems.

[0003] In path selection for multi-chip acceleration platforms, routing is typically performed based on individual packets, single destination lookups, or instantaneous congestion feedback. While this approach works under typical load scenarios, it presents a rare but challenging drawback in the low-batch, asynchronous, and sparsely activated scenarios unique to the IoT edge: multiple source chips generate several fragmented communication requests with small payloads, no single repetition, but the same destination chip within a very short acquisition window. Traditional packet-by-packet routing treats these fragmented requests, which essentially point to the same destination, as independent streams and sends them separately. This causes the forward link of the same destination chip to be repeatedly broken down, resulting in micro-burst queuing, path jitter, instantaneous overload of cold destination links while other links are idle, and loss of continuity after the main path is interrupted. Especially in inter-chip networks with heterogeneous bandwidth domains and inconsistent bandwidth-delay products, these problems are more likely to induce backpressure propagation and head-of-line blocking, forcing fragmented communications that could have converged at the near end and sent out once at once to undergo multiple long-distance traversals. Summary of the Invention

[0004] The purpose of this invention is to provide a communication path selection method and system for multi-chip acceleration platforms to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a communication path selection method and system for a multi-chip acceleration platform, including a request characterization and same-target fragment communication group identification module, used to collect communication requests to be sent from each source chip within a short-time acquisition window, generate a request description item, and identify same-target fragment communication groups pointing to the same target chip from the communication requests to be sent. The converging core selection and main path planning module is used to determine candidate converging cores for the same target fragment communication group, select the target converging core, and select a main path with continuous idle time slot intervals between the target converging core and the target core. The path reservation and holding control module is used to issue path reservation control messages on the main path, establish continuous forwarding records corresponding to the same target fragment communication group, maintain the stable use of the main path in multiple consecutive evaluation cycles, and switch to the backup confluence core or re-execute path planning when the switching conditions are met.

[0006] According to the above technical solution, the request characterization and target fragment communication group identification module includes a request acquisition submodule, used to receive communication requests to be sent output by each source core; a description item generation submodule, used to generate a request description item for each communication request to be sent, including source core identifier, destination core identifier, payload size, remaining delay budget, service category, sequence index, and generation time slot number; and a flow group construction submodule, used to group multiple communication requests to be sent into the same target fragment communication group based on destination core consistency, service category consistency, generation time slot number proximity, and remaining delay budget proximity. The merging core selection and main path planning module includes a candidate merging generation submodule, used to generate multiple candidate merging cores around the forward neighbor region of the target core; a merging delay calculation submodule, used to calculate the merging delay, transmission delay, queuing delay and expected congestion delay of each candidate merging core; and a main path selection submodule, used to select the target main path from multiple reachable paths from the target merging core to the target core based on the length of the continuous idle time slot interval and the path hop count. The path reservation and holding control module includes a path reservation control submodule, used to send path reservation control messages on the target main path and establish continuous forwarding records; a path holding determination submodule, used to determine whether to hold the current path based on the flow group stability, queuing status of the rendezvous point, and the continuous idle time slot interval condition of the main path in the current evaluation period; a backup switching submodule, used to select a replacement rendezvous core from the backup rendezvous core set and re-establish the main path when the current rendezvous core no longer meets the holding conditions; and a status feedback submodule, used to write the path usage results of the current period back to the link status table and the rendezvous status table. The link status table is used to record the number of continuous idle transmission time slots and the expected number of blocked time slots for each link in the future evaluation window. The rendezvous status table is used to record the number of queued bytes, the path reservation control message transmission results, and the current path holding status for each candidate rendezvous core.

[0007] A converged communication path selection method for multi-chip acceleration platforms in the Internet of Things includes the following steps: S1. Collect the communication requests to be sent from each source chip within the current short-term acquisition window, and generate a request description for each communication request to be sent; based on the destination chip, service category, generation time slot number and remaining delay budget, group multiple communication requests to be sent into the same target fragment communication group; output the source chip set, destination chip and total payload scale corresponding to each same target fragment communication group. S2. Generate a set of candidate merging cores for each fragment communication group with the same target, and calculate the merging delay for each candidate merging core; select the target merging core based on the merging delay; output the target merging core and its corresponding main path planning input conditions; S3. Enumerate multiple candidate paths between the target converging core and the destination core, and calculate the length of the continuous idle time slot interval that each candidate path can provide within the future evaluation window; select the main path from the candidate paths that meet the continuous transmission requirements; when there are no candidate paths that meet the continuous transmission requirements in the current period, retain the same target fragment communication group and enter the next short-time acquisition window to continue evaluation; S4. First, send a path reservation control message along the main path so that each intermediate forwarding node on the main path establishes a continuous forwarding record corresponding to the same target fragment communication group; each source core sends the payload belonging to the same target fragment communication group to the target merging core; the target merging core merges the payload into a column of continuously sent messages according to the sequential index, and sends it to the destination core in one go along the main path. S5. Detect the stability of the current target fragment communication group, the queuing status of the target merging core, and the status of the continuous idle time slot interval of the main path within multiple consecutive evaluation cycles; when the maintenance condition is continuously met, maintain the current target merging core and the main path unchanged; when the switching condition is met, switch to the backup merging core or re-execute the candidate merging selection and main path planning.

[0008] According to the above technical solution, step S1 includes the following sub-steps: S1-1. In order to convert the original communication requests to be sent from each source chip into a unified input structure that can be directly used for subsequent group decision-making and path planning, in the first... Short-time data acquisition window Internal collection of communication request sets to be sent and for any communication request to be sent Generate request description item: ,in, Indicates the communication request to be sent Source chip identification, Indicates the communication request to be sent Purpose core identification, Indicates the communication request to be sent Payload bytes Indicates the communication request to be sent The remaining delay budget at the start of the current short-time acquisition window. Indicates the communication request to be sent Business categories, Indicates the communication request to be sent Original transmission sequence index within the source core Indicates the communication request to be sent The corresponding generated time slot number within the current short-term acquisition window will be used for subsequent unified calculation with the number of time slots required for forwarding; S1-2, simultaneously constrain the service type, arrival time proximity, and remaining delay budget proximity, and perform same-target fragment communication group determination according to the following conditions, and send two communication requests to be sent. and Grouped into the same target fragment communication group The judgment criteria are set as follows: , , , ,in, This represents the maximum allowed time difference of arrival threshold for the same target fragment communication group. This represents the maximum allowed delay budget difference threshold for the same target fragment communication group. The meaning of the above determination method is that only when multiple communication requests to be sent have the same purpose, the same service category, and are close enough in time are they considered to be the same target fragment communication group suitable for first merging and then sending. S1-3, For any fragment communication group with the same target The total load capacity is calculated according to the following formula: ,in Indicates the same target fragment communication group The total payload bytes; the purpose of calculating this total payload bytes is to provide direct input for subsequently determining the number of consecutive transmission time slots required by the same target fragment communication group on the main path, and to provide a basis for the payload scale for the continuous long-distance transmission after the merged cores are executed in an ordered manner, while also indexing according to the order in each request description item. A corresponding sending sequence table is formed so that the original order can be restored after the aggregated core particles have completed their aggregation.

[0009] According to the above technical solution, step S2 includes the following sub-steps: S2-1. To enable the fragment payloads from each source core to converge near the target core, thereby reducing the long-distance transmission after convergence, and to avoid introducing new relay delays by setting the convergence location too far away, for any target fragment communication group... target core In the target core Surrounding structure candidate confluence core set: ,in Indicates candidate confluence cores To the target core The topological hop count, This represents the maximum forward neighbor hops of the candidate merging core relative to the target core; the meaning of this formula is to limit the candidate merging core to the vicinity of the target core to ensure that the long-distance transmission backbone path after merging is not too long. S2-2, For any candidate merged core Calculate the distance from each source core to the candidate merge core. Total estimated import latency: And calculate candidate confluence cores To the target core Expected long-distance transmission delay: ,in, Indicates from the core To the core The expected transmission delay, expressed in the number of transmission time slots; Indicates the same target fragment communication group Each source chip merges into a candidate merge chip. The total estimated inflow delay, Indicates candidate confluence cores To the target core The estimated long-distance transmission delay, after obtaining the inbound path delay and the long-distance transmission path delay after merging of the candidate merging core, also needs to be further considered whether there is a queue backlog of the candidate merging core, and whether the candidate merging core faces existing blockage when forward transmission to the destination core, thus forming the complete total merging delay. S2-3, Combine candidate core particles Current number of bytes in queue Converted to queuing delay: ,in, This indicates the number of bytes that a single transmit slot can carry. Indicates candidate confluence cores Additional queuing slots required; simultaneously read candidate merge cores from the link state table. To the target core Expected number of blocked time slots Based on the aforementioned total expected inflow delay, expected long-distance transmission delay, and queuing delay, candidate merging cores are calculated. Total rendezvous delay: The principle of this formula is to uniformly convert the source-end merging delay, the post-merging transmission delay, the queuing delay at the merging point, and the forward blocking delay into a number of transmission time slots for comparison, so as to select the target merging core that is truly suitable for undertaking the merging task from multiple candidate merging cores; from the set of candidate merging cores In the middle, choose to make the total rendezvous delay The candidate merging core with the minimum value is selected as the target merging core. Finally, the candidate merging core that satisfies the following formula is selected as the target merging core: .

[0010] According to the above technical solution, step S3 includes the following sub-steps: S3-1, For the target confluence core particle With target core Construct a candidate path set and in the A future assessment window Internally read the length of consecutive idle time slots for each link on each candidate path; S3-2. To ensure that the selected main path is not interrupted during actual continuous long-distance transmission due to insufficient continuous idle time slots of any link, the link with the weakest continuous transmission capability on the candidate path needs to be used as the characterization of the continuous transmission capability of the entire candidate path. For any candidate path... The capacity of its continuous idle time slot interval is calculated according to the following formula: ,in Indicate candidate path Any link on, Indicates the link Future assessment window The number of consecutive idle transmission slots available within. Indicate candidate path The maximum number of consecutive transmission time slots that can be transmitted without interruption; this formula uses the shortest consecutive idle time slot interval link on the path as the continuous transmission capacity of the entire path to ensure that the main path selection result can truly reflect the continuous transmission upper limit of the entire path. S3-3. Calculate the target fragment communication group according to the following formula. Number of continuous transmission demand time slots: ,in, This indicates the completion of the same target fragment communication group. The minimum number of consecutive transmission slots required for a single long-distance transmission; among all candidate paths that satisfy the following formula, select the candidate path with the smallest topology hop count as the main path. : When no candidate path meets the above conditions, select the one that... The candidate path that achieves the maximum value is selected as the priority candidate path for the next evaluation cycle, and this candidate path is written into the path candidate cache. Simultaneously, the same target fragment communication group is... The fragmented payloads that should have been merged and sent to the next short acquisition window will be retained for further evaluation to avoid sending them separately in the current cycle.

[0011] According to the above technical solution, step S4 includes the following sub-steps: S4-1. First, calculate the target fragment communication group according to the following formula. At the target confluence core The latest estimated arrival time slot for completing the rendezvous: ,in Indicates the same target fragment communication group All fragment loads reached the target confluence core. The latest estimated time slot; then on the main path It is determined no earlier than The earliest continuous transmission start time slot This makes the main path Each link in the interval All of them have the conditions for continuous transmission. After obtaining the latest rendezvous completion time slot of the same target fragment communication group and the continuous transmission start time slot of the main path, the actual payload is not sent immediately. Instead, a path reservation control message is sent to the intermediate forwarding nodes on the main path so as to establish a continuous forwarding record in advance before the actual payload arrives, thereby reducing the interruption caused by intermediate nodes re-competing for forwarding resources during the actual continuous transmission process. S4-2, Constructing a path reservation control message: ,in, Indicates the target fragment communication group identifier; the path reservation control message Along the main path Send the data first, so that each intermediate forwarding node on the main path establishes a continuous forwarding record: ,in, This indicates the end time slot of continuous forwarding; the function of the continuous forwarding record is to indicate the intermediate forwarding nodes on the main path at the start time slot. until the end of the gap Within the interval, for those with the same flow group identifier The message queues are maintained in a continuous forwarding order in the same path direction, and are not reallocated to other candidate paths within that interval; S4-3, Each source chip will belong to the same target fragment communication group. The fragment payload is sent to the target confluence core. The target confluence core Indexed in order The fragmented payloads are sequentially merged to form a continuously transmitted message train, and this is done in the initial time slot. until the end of the gap within the interval along the main path Send to the destination chip in one go .

[0012] According to the above technical solution, step S5 includes the following sub-steps: S5-1. Within the current assessment period, calculate the communication group of the same target fragment according to the following formula. Stability: ,in This indicates that the target chip identifier and service category remain consistent with the previous assessment period, and the chips continue to be classified into the same target fragment communication group. The number of communication requests to be sent. This indicates that the fragment belongs to the same target communication group during the current evaluation period. The total number of communication requests to be sent; the stability Used to characterize whether the current flow group is still suitable to use the confluence core and main path of the previous evaluation cycle; S5-2. To avoid path jitter caused by rerouting in every evaluation cycle, first determine whether the current target fragment communication group is still suitable to continue using the rendezvous relationship and main path established in the previous cycle. Then, perform path preservation determination based on the following preservation conditions. When all three of the following conditions are met, maintain the current target rendezvous core. With the main path constant: , , ,in This represents the minimum stability threshold for path preservation. This represents the maximum allowed queuing time slot threshold at the merging point; the meaning of the above retention criterion is that the current path organization result will only be maintained if the flow group composition is sufficiently stable, the queuing at the merging point has not deteriorated, and the main path still has a sufficient number of continuous idle time slot intervals. S5-3, Surrounding the target core particle Pre-configured spare merge core assembly The spare confluence core assembly To surround the target core A pre-configured set of candidate merge core particles is used to perform a fast replacement selection when the current target merge core particle no longer meets the retention conditions. A backup merge switchover is performed when any one of the following four switchover conditions occurs: First, ;second, ;third, And continue The evaluation cycle; fourth, the number of failed path reservation control message transmissions. Reaching the threshold During standby merge handover, the standby merge core assembly is selected according to the following formula. Replacement cores selected from among them: and the alternative confluence core. Repeat steps S2 to S4; if the continuous transmission condition still cannot be met after the backup rendezvous handover, regenerate the candidate rendezvous core set. And execute a new round of complete path planning.

[0013] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: First, by first identifying the same target fragment communication group and then merging the small payload communications of multiple source ends before long-distance transmission, the number of repeated long-distance traversals of the same target chip can be significantly reduced, thereby reducing the micro-burst impact of cold destination links from the root. Second, by using continuous idle time slot intervals rather than single-point instantaneous idle as the main path selection criterion, the interruption of the main path and path jitter can be reduced, improving the continuity of inter-chip long-distance transmission. Third, by establishing lightweight continuous forwarding records through path reservation control messages instead of using heavy fixed circuits, the adaptive capability of the inter-chip network is preserved, and sufficient short-term continuous passage conditions are provided for the same target fragment communication group. Fourth, through path holding criteria and backup merging chip switching mechanism, the merging point itself can be prevented from evolving into a new hotspot, thus making this scheme applicable not only to inter-chip data transmission between computing chips and storage chips, but also to various short-term burst communications such as control response, feature transport, and intermediate result write-back in IoT multi-chip platforms. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall modular structure of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1 The present invention provides a technical solution: a communication path selection method and system for a multi-chip acceleration platform, including a request characterization and same-target fragment communication group identification module, which is used to collect the communication requests to be sent from each source chip within a short acquisition window, generate a request description item, and identify the same-target fragment communication group pointing to the same target chip from the communication requests to be sent. The merge core selection and main path planning module is used to determine candidate merge cores for the same target fragment communication group, select the target merge core, and select a main path with continuous idle time slot intervals between the target merge core and the target core. The path reservation and hold control module is used to issue path reservation control messages on the main path, establish continuous forwarding records corresponding to the same target fragment communication group, maintain the stable use of the main path in multiple consecutive evaluation cycles, and switch to the backup convergence core or re-execute path planning when the switching conditions are met. The request characterization and target fragment communication group identification module includes a request acquisition submodule, used to receive communication requests to be sent from each source core; a description item generation submodule, used to generate a request description item for each communication request to be sent, including source core identifier, destination core identifier, payload size, remaining delay budget, service category, sequence index, and generation time slot number; and a flow group construction submodule, used to group multiple communication requests to be sent into the same target fragment communication group based on destination core consistency, service category consistency, generation time slot number proximity, and remaining delay budget proximity. The merge particle selection and main path planning module includes a candidate merge generation submodule, which generates multiple candidate merge particles around the forward neighbor region of the target particle; a merge delay calculation submodule, which calculates the merge delay, transmission delay, queuing delay, and expected congestion delay for each candidate merge particle; and a main path selection submodule, which selects the target main path from multiple reachable paths from the target merge particle to the target particle based on the length of the continuous idle time slot interval and the path hop count. The path reservation and hold control module includes a path reservation control submodule, used to send path reservation control messages on the target main path and establish continuous forwarding records; a path hold determination submodule, used to determine whether to hold the current path based on the flow group stability, queuing status of the merge point, and the continuous idle time slot interval of the main path in the current evaluation period; a backup switching submodule, used to select a replacement merge core from the backup merge core set and re-establish the main path when the current merge core no longer meets the hold conditions; and a status feedback submodule, used to write the path usage results of the current period back to the link status table and the merge status table. The link status table records the number of continuous idle transmission time slots and the expected number of blocked time slots for each link in the future evaluation window, and the merge status table records the number of queued bytes, the path reservation control message transmission results, and the current path hold status for each candidate merge core. A converged communication path selection method for multi-chip acceleration platforms in the Internet of Things includes the following steps: S1. Collect the communication requests to be sent from each source chip within the current short-term acquisition window, and generate a request description for each communication request to be sent; based on the destination chip, service category, generation time slot number and remaining delay budget, group multiple communication requests to be sent into the same target fragment communication group; output the source chip set, destination chip and total payload scale corresponding to each same target fragment communication group. S2. Generate a set of candidate merging cores for each fragment communication group with the same target, and calculate the merging delay for each candidate merging core; select the target merging core based on the merging delay; output the target merging core and its corresponding main path planning input conditions; S3. Enumerate multiple candidate paths between the target confluence core and the destination core, and calculate the length of the continuous idle time slot interval that each candidate path can provide in the future evaluation window; select the main path from the candidate paths that meet the continuous transmission requirements; when there are no candidate paths that meet the continuous transmission requirements in the current period, retain the communication group with the target fragment to enter the next short-time acquisition window for continued evaluation; S4. First, send a path reservation control message along the main path to enable each intermediate forwarding node on the main path to establish a continuous forwarding record corresponding to the target fragment communication group; each source core will send the payload belonging to the same target fragment communication group to the target merging core; the target merging core will merge the payload into a column of continuously sent messages according to the sequential index, and send it to the destination core in one go along the main path. S5. Detect the stability of the current target fragment communication group, the queuing status of the target merging core, and the status of the continuous idle time slot interval of the main path within multiple consecutive evaluation cycles; when the maintenance condition is continuously met, maintain the current target merging core and the main path unchanged; when the switching condition is met, switch to the backup merging core or re-execute the candidate merging selection and main path planning. Step S1 includes the following sub-steps: S1-1. In order to convert the original communication requests to be sent from each source chip into a unified input structure that can be directly used for subsequent group decision-making and path planning, in the first... Short-time data acquisition window Internal collection of communication request sets to be sent and for any communication request to be sent Generate request description item: ,in, Indicates a communication request to be sent. Source chip identification, Indicates a communication request to be sent. Purpose core identification, Indicates a communication request to be sent. Payload bytes Indicates a communication request to be sent. The remaining delay budget at the start of the current short-time acquisition window. Indicates a communication request to be sent. Business categories, Indicates a communication request to be sent. Original transmission sequence index within the source core Indicates a communication request to be sent. The corresponding generated time slot number within the current short-term acquisition window will be used for subsequent unified calculation with the number of time slots required for forwarding; S1-2, simultaneously constrain the service type, arrival time proximity, and remaining delay budget proximity, and perform same-target fragment communication group determination according to the following conditions, and send two communication requests to be sent. and Grouped into the same target fragment communication group The judgment criteria are set as follows: , , , ,in, This represents the maximum allowed time difference of arrival threshold for the same target fragment communication group. This represents the maximum allowed delay budget difference threshold for the same target fragment communication group. The meaning of the above determination method is that only when multiple communication requests to be sent have the same purpose, the same service category, and are close enough in time are they considered to be the same target fragment communication group suitable for first merging and then sending. S1-3, For any fragment communication group with the same target The total load capacity is calculated according to the following formula: ,in Indicates the same target fragment communication group The total payload bytes; the purpose of calculating this total payload bytes is to provide direct input for subsequently determining the number of consecutive transmission time slots required by the same target fragment communication group on the main path, and to provide a basis for the payload scale for the continuous long-distance transmission after the merged cores are executed in an ordered manner, while also indexing according to the order in each request description item. A corresponding sending sequence table is formed so that the original order can be restored after the aggregated core particles have completed their aggregation. This invention does not immediately route each of the multiple pending communication requests independently. Instead, it first determines whether these requests actually point to the same destination chip, whether they belong to the same service category, and whether they are sufficiently close in time. This approach is because in multi-chip acceleration platforms, many factors affecting link stability are not caused by a single large flow, but by the superposition of multiple small load requests occurring consecutively within a short period. If conventional methods were used to route these small load requests separately, each request, though seemingly small, would repeatedly traverse the link leading to the same destination chip within a short timeframe, causing frequent short-term occupancy and intermittent congestion on the forward path of that destination chip.

[0017] This implementation first identifies requests that meet certain conditions as fragmented communication groups with the same target. Its working principle is not simply merging data, but rather transforming the originally scattered, fragmented, and independently decided transmission behaviors into transmission units that can be uniformly scheduled. The purpose of this is to establish subsequent convergence core selection and main path planning on the basis of group communication, rather than on individual small flows. Small payload requests with the same destination, consistent business semantics, and adjacent times are treated as the same scheduling object, fundamentally changing the object of subsequent path selection, enabling subsequent steps to achieve a communication organization mode of convergence first, followed by continuous long-distance transmission.

[0018] Step S2 includes the following sub-steps: S2-1. To enable the fragment payloads from each source core to converge near the target core, thereby reducing the long-distance transmission after convergence, and to avoid introducing new relay delays by setting the convergence location too far away, for any target fragment communication group... target core In the target core Surrounding structure candidate confluence core set: ,in Indicates candidate confluence cores To the target core The topological hop count, This represents the maximum forward neighbor hops of the candidate merging core relative to the target core; the meaning of this formula is to limit the candidate merging core to the vicinity of the target core to ensure that the long-distance transmission backbone path after merging is not too long. S2-2, For any candidate merged core Calculate the distance from each source core to the candidate merge core. Total estimated import latency: And calculate candidate confluence cores To the target core Expected long-distance transmission delay: ,in, Indicates from the core To the core The expected transmission delay, expressed in the number of transmission time slots; Indicates the same target fragment communication group Each source chip merges into a candidate merge chip. The total estimated inflow delay, Indicates candidate confluence cores To the target core The estimated long-distance transmission delay, after obtaining the inbound path delay and the long-distance transmission path delay after merging of the candidate merging core, also needs to be further considered whether there is a queue backlog of the candidate merging core, and whether the candidate merging core faces existing blockage when forward transmission to the destination core, thus forming the complete total merging delay. S2-3, Combine candidate core particles Current number of bytes in queue Converted to queuing delay: ,in, This indicates the number of bytes that a single transmit slot can carry. Indicates candidate confluence cores Additional queuing slots required; simultaneously read candidate merge cores from the link state table. To the target core Expected number of blocked time slots Based on the aforementioned total expected inflow delay, expected long-distance transmission delay, and queuing delay, candidate merging cores are calculated. Total rendezvous delay: The principle of this formula is to uniformly convert the source-end merging delay, the post-merging transmission delay, the queuing delay at the merging point, and the forward blocking delay into a number of transmission time slots for comparison, so as to select the target merging core that is truly suitable for undertaking the merging task from multiple candidate merging cores; from the set of candidate merging cores In the middle, choose to make the total rendezvous delay The candidate merging core with the minimum value is selected as the target merging core. Finally, the candidate merging core that satisfies the following formula is selected as the target merging core: ; This implementation method does not employ the conventional path selection approach of directly sending data from the closest available core or prioritizing the least available core. Instead, it specifically sets up a convergence core for communication groups of fragments sharing the same target. This implementation first filters candidate convergence cores near the target core, then comprehensively assesses the capacity of different candidate convergence cores to handle the loads from each source core and the smoothness of continued transmission to the target core. Its working principle is to allow multiple small payloads from different sources to converge at a suitable location, and then initiate subsequent transmissions to the target core from that location. The effect of this is to transform multiple scattered transmissions into a single, concentrated transmission after near-end convergence. The convergence core is used as an actively configured intermediate organizing point to change the structure of the entire communication path, rather than simply selecting the currently optimal path from the existing paths.

[0019] Step S3 includes the following sub-steps: S3-1, Targeted Merging Core Particles With target core Construct a candidate path set and in the A future assessment window Internally read the length of consecutive idle time slots for each link on each candidate path; S3-2. To ensure that the selected main path is not interrupted during actual continuous long-distance transmission due to insufficient continuous idle time slots of any link, the link with the weakest continuous transmission capability on the candidate path needs to be used as the characterization of the continuous transmission capability of the entire candidate path. For any candidate path... The capacity of its continuous idle time slot interval is calculated according to the following formula: ,in Indicate candidate path Any link on, Indicates link Future assessment window The number of consecutive idle transmission slots available within. Indicate candidate path The maximum number of consecutive transmission time slots that can be transmitted without interruption; this formula uses the shortest consecutive idle time slot interval link on the path as the continuous transmission capacity of the entire path to ensure that the main path selection result can truly reflect the continuous transmission upper limit of the entire path. S3-3. Calculate the target fragment communication group according to the following formula. Number of continuous transmission demand time slots: ,in, Indicates completion of communication group with target fragment The minimum number of consecutive transmission slots required for a single long-distance transmission; among all candidate paths that satisfy the following formula, select the candidate path with the smallest topology hop count as the main path. : When no candidate path meets the above conditions, select the one that... The candidate path that achieves the maximum value is selected as the priority candidate path for the next evaluation cycle and written into the path candidate cache. Simultaneously, it is communicated with the target fragment group. Retain the data until the next short acquisition window for further evaluation to avoid splitting up and sending fragmented payloads that should have been merged and transmitted in the current cycle; When determining the main path, this implementation method does not rely solely on path length, local idleness, or instantaneous availability as in conventional approaches. Instead, it focuses on whether a path can provide continuous transmission over a subsequent period. This method pre-determines whether candidate paths possess sufficient continuous transmission capacity within the future evaluation window, thus prioritizing paths capable of supporting the entire long-distance transmission process. If no path currently meets the continuous transmission requirements, it does not hastily split the communication group with the same target for separate transmission but retains it for further evaluation in the next short-term acquisition window. This avoids breaking up communication that could be organized as a whole in pursuit of immediate transmission, thereby disrupting the effects of previous grouping and rendezvous. Compared to conventional technologies that select paths based solely on instantaneous status, this method uses continuous availability as the core criterion for main path usability, ensuring that the main path is not merely the optimal channel at a given moment, but a stable channel capable of fully supporting a grouped long-distance transmission process.

[0020] Step S4 includes the following sub-steps: S4-1. First, calculate the target fragment communication group according to the following formula. At the target confluence core The latest estimated arrival time slot for completing the rendezvous: ,in Indicates the same target fragment communication group All fragment loads reached the target confluence core. The latest estimated time slot; then on the main path It is determined no earlier than The earliest continuous transmission start time slot This makes the main path Each link in the interval All of them have the conditions for continuous transmission. After obtaining the latest rendezvous completion time slot of the same target fragment communication group and the continuous transmission start time slot of the main path, the actual payload is not sent immediately. Instead, a path reservation control message is sent to the intermediate forwarding nodes on the main path so as to establish a continuous forwarding record in advance before the actual payload arrives, thereby reducing the interruption caused by intermediate nodes re-competing for forwarding resources during the actual continuous transmission process. S4-2, Constructing a path reservation control message: ,in, Indicates the target fragment communication group identifier; Path reservation control message Along the main path Send the data first, so that each intermediate forwarding node on the main path establishes a continuous forwarding record: ,in, This indicates the end time slot of continuous forwarding; the purpose of the continuous forwarding record is to indicate the intermediate forwarding nodes on the main path at the start time slot. until the end of the gap Within the interval, for those with the same flow group identifier The message queues are maintained in a continuous forwarding order in the same path direction, and are not reallocated to other candidate paths within that interval; S4-3, Each source chip will belong to the same target fragment communication group. The fragment payload is sent to the target confluence core. , from target converging core particles Indexed in order The fragmented payloads are sequentially merged to form a continuously transmitted message train, and the process begins in the initial time slot. until the end of the gap within the interval along the main path Send to the destination chip in one go ; In this implementation, the path reservation control message does not carry service data, but rather indicates in advance to intermediate forwarding nodes on the main path that a series of incoming messages will need to pass through in a predetermined direction. Continuous forwarding conditions are established for the subsequent message series before the actual payload enters the main path. This prevents intermediate nodes from processing subsequent message series individually as ordinary, scattered messages, and instead allows them to maintain a consistent forwarding direction and order within the continuous transmission interval. By using a lightweight control action to pre-organize the forwarding conditions required for continuous long-distance transmission, the converged one-time long-distance transmission can be truly completed continuously, rather than being broken up again midway.

[0021] Step S5 includes the following sub-steps: S5-1. Within the current assessment period, calculate the communication group of the same target fragment according to the following formula. Stability: ,in This indicates that the target chip identifier and service category remain consistent with the previous assessment period, and the chips continue to be classified into the same target fragment communication group. The number of communication requests to be sent. This indicates that the fragments are classified into the same target communication group during the current evaluation period. Total number of pending communication requests; stability Used to characterize whether the current flow group is still suitable to use the confluence core and main path of the previous evaluation cycle; S5-2. To avoid path jitter caused by rerouting in every evaluation cycle, first determine whether the current target fragment communication group is still suitable to continue using the rendezvous relationship and main path established in the previous cycle. Then, perform path preservation determination based on the following preservation conditions. When all three of the following conditions are met, maintain the current target rendezvous core. With the main path constant: , , ,in This represents the minimum stability threshold for path preservation. This represents the maximum allowed queuing time slot threshold at the merging point; the meaning of the above retention criterion is that the current path organization result will only be maintained if the flow group composition is sufficiently stable, the queuing at the merging point has not deteriorated, and the main path still has a sufficient number of continuous idle time slot intervals. S5-3, Surrounding the target core particle Pre-configured spare merge core assembly Backup consolidation core assembly To surround the target core A pre-configured set of candidate merge core particles is used to perform a fast replacement selection when the current target merge core particle no longer meets the retention conditions. A backup merge switchover is performed when any one of the following four switchover conditions occurs: First, ;second, ;third, And continue The evaluation cycle; fourth, the number of failed path reservation control message transmissions. Reaching the threshold During standby merge handover, the standby merge core assembly is selected according to the following formula. Replacement cores selected from among them: and alternative confluence cores Re-execute steps S2 to S4; if the continuous transmission condition still cannot be met after the backup rendezvous handover, regenerate the candidate rendezvous core set. And execute a new round of complete path planning.

[0022] This implementation addresses this problem by combining decision-making and backup rendezvous switching. Its working principle is as follows: when the current path organization is still valid, the existing organization result is maintained first; only when the maintenance condition is actually broken does it switch to a backup rendezvous core or re-execute the complete plan. This avoids path rigidity while preventing overly sensitive and frequent reselections. By separating path maintenance, backup switching, and replanning into different levels, the system can maintain the continuous transmission effect formed by previous steps while also quickly moving away from unsuitable rendezvous positions and main paths when necessary.

[0023] Through the above implementation method, small-load communication requests that originally appeared scattered among multiple source cores and repeatedly impacted the forward path of the same destination core in a short period of time are reorganized into fragmented communication groups with the same target that can be processed uniformly. These communication groups are no longer sent out independently, but are first gathered at a suitable merging core, and then sent to the destination core in one go along the main path with continuous transmission conditions. After this processing, the forward region of the destination core no longer frequently suffers from fragmented traffic that repeatedly enters, and intermediate forwarding nodes no longer temporarily organize continuous forwarding conditions when the actual message arrives, but instead establish continuous forwarding arrangements in advance after the path reservation control message arrives.

[0024] Furthermore, during subsequent evaluation cycles, the system will not repeatedly change routes due to slight fluctuations in local states, but will instead handle these processes in a layered manner, including maintenance, backup switching, and replanning. The final result is not simply that one path is better at a certain moment, but rather that the entire process of communication with the target fragments—from source identification, near-end convergence, continuous transmission, and subsequent maintenance—forms a coherent organizational structure, making it more suitable for handling short-term fragmented, destination-centralized inter-chip communication scenarios in multi-chip acceleration platforms.

[0025] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0026] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A communication path selection system for multi-chip acceleration platforms, characterized in that: include: The request characterization and same-target fragment communication group identification module is used to collect the communication requests to be sent from each source chip within a short acquisition window, generate a request description item, and identify the same-target fragment communication group pointing to the same target chip from the communication requests to be sent. The converging core selection and main path planning module is used to determine candidate converging cores for the same target fragment communication group, select the target converging core, and select a main path with continuous idle time slot intervals between the target converging core and the target core. The path reservation and holding control module is used to issue path reservation control messages on the main path, establish continuous forwarding records corresponding to the same target fragment communication group, maintain the stable use of the main path in multiple consecutive evaluation cycles, and switch to the backup confluence core or re-execute path planning when the switching conditions are met.

2. The communication path selection system for a multi-chip acceleration platform according to claim 1, characterized in that: The request characterization and target fragment communication group identification module includes a request acquisition submodule, which is used to receive communication requests to be sent from each source chip. The description item generation submodule is used to generate a request description item for each communication request to be sent, which includes source core identifier, destination core identifier, payload size, remaining delay budget, service category, sequence index and generation time slot number; the flow group construction submodule is used to group multiple communication requests to be sent into the same target fragment communication group based on destination core consistency, service category consistency, generation time slot number proximity and remaining delay budget proximity. The merged core selection and main path planning module includes a candidate merge generation submodule, which is used to generate multiple candidate merged cores around the forward neighbor region of the target core. The merging delay calculation submodule is used to calculate the merging delay, remote transmission delay, queuing delay and expected congestion delay of each candidate merging core; the main path selection submodule is used to select the target main path from multiple reachable paths from the target merging core to the destination core based on the length of the continuous idle time slot interval and the number of path hops. The path reservation and hold control module includes a path reservation control submodule, used to send path reservation control messages on the target main path and establish a continuous forwarding record; a path hold determination submodule, used to determine whether to hold the current path based on the flow group stability, queuing status of the merge point, and the continuous idle time slot interval of the main path within the current evaluation period; and a backup switching submodule, used to select a replacement merge core from the backup merge core set and re-establish the main path when the current merge core no longer meets the hold conditions. The status feedback submodule is used to write the path usage results of the current cycle back to the link status table and the convergence status table. The link status table is used to record the number of consecutive idle transmission slots and the number of expected blocked slots for each link in the future evaluation window. The convergence status table is used to record the number of queued bytes, the path reservation control message transmission results, and the current path hold status for each candidate convergence chip.

3. A converged communication path selection method for an IoT multi-chip acceleration platform based on the system described in claim 1 or 2, characterized in that: Includes the following steps: S1. Collect the communication requests to be sent from each source chip within the current short-term acquisition window, and generate a request description item for each communication request to be sent; based on the destination chip, service category, generation time slot number and remaining delay budget, group multiple communication requests to be sent into the same target fragment communication group. Output the source core set, destination core, and total payload size for each communication group with the target fragment; S2. Generate a set of candidate merging cores for each fragment communication group with the same target, and calculate the merging delay for each candidate merging core; select the target merging core based on the merging delay; output the target merging core and its corresponding main path planning input conditions; S3. Enumerate multiple candidate paths between the target converging core and the destination core, and calculate the length of the continuous idle time slot interval that each candidate path can provide within the future evaluation window; select the main path from the candidate paths that meet the continuous transmission requirements; when there are no candidate paths that meet the continuous transmission requirements in the current period, retain the same target fragment communication group and enter the next short-time acquisition window to continue evaluation; S4. First, send a path reservation control message along the main path so that each intermediate forwarding node on the main path establishes a continuous forwarding record corresponding to the same target fragment communication group; each source core sends the payload belonging to the same target fragment communication group to the target merging core; the target merging core merges the payload into a column of continuously sent messages according to the sequential index, and sends it to the destination core in one go along the main path. S5. Detect the stability of the current target fragment communication group, the queuing status of the target merging core, and the status of the continuous idle time slot interval of the main path within multiple consecutive evaluation cycles; when the maintenance condition is continuously met, maintain the current target merging core and the main path unchanged; when the switching condition is met, switch to the backup merging core or re-execute the candidate merging selection and main path planning.

4. The communication path selection method for a multi-chip acceleration platform according to claim 3, characterized in that: Step S1 includes the following sub-steps: S1-1, in the Short-time data acquisition window Internal collection of communication request sets to be sent and for any communication request to be sent Generate request description item: ,in, Indicates the communication request to be sent Source chip identification, Indicates the communication request to be sent Purpose core identification, Indicates the communication request to be sent Payload bytes Indicates the communication request to be sent The remaining delay budget at the start of the current short-time acquisition window. Indicates the communication request to be sent Business categories, Indicates the communication request to be sent Original transmission sequence index within the source core Indicates the communication request to be sent The corresponding generated time slot number within the current short-term acquisition window will be used for subsequent unified calculation with the number of time slots required for forwarding; S1-2, simultaneously constrain the service type, arrival time proximity, and remaining delay budget proximity, and perform same-target fragment communication group determination according to the following conditions, and send two communication requests to be sent. and Grouped into the same target fragment communication group The judgment criteria are set as follows: , , , ,in, This represents the maximum allowed time difference of arrival threshold for the same target fragment communication group. This represents the maximum allowable delay budget difference threshold for the same target fragment communication group; S1-3, For any fragment communication group with the same target The total load capacity is calculated according to the following formula: ,in Indicates the same target fragment communication group The total payload bytes; the purpose of calculating this total payload bytes is to provide direct input for subsequently determining the number of consecutive transmission time slots required by the same target fragment communication group on the main path, and to provide a basis for the payload scale for the continuous long-distance transmission after the merged cores are executed in an ordered manner, while also indexing according to the order in each request description item. A corresponding sending sequence table is formed.

5. The communication path selection method for a multi-chip acceleration platform according to claim 4, characterized in that: Step S2 includes the following sub-steps: S2-1, For any fragment communication group with the same target target core In the target core Surrounding structure candidate confluence core set: ,in Indicates candidate confluence cores To the target core The topological hop count, This represents the maximum forward neighbor hop count of the candidate confluence core relative to the target core; S2-2, For any candidate merged core Calculate the ratio of each source core to the candidate merge core. Total estimated inbound latency: And calculate candidate confluence cores To the target core Expected long-distance transmission delay: ,in, Indicates from the core To the core The expected transmission delay, expressed in the number of transmission time slots; Indicates the same target fragment communication group Each source chip merges into a candidate merge chip. The total estimated inflow delay, Indicates candidate confluence cores To the target core The estimated long-distance transmission delay, after obtaining the inbound path delay and the long-distance transmission path delay after merging of the candidate merging core, also needs to be further considered whether there is a queue backlog of the candidate merging core, and whether the candidate merging core faces existing blockage when forward transmission to the destination core, thus forming the complete total merging delay. S2-3, Combine candidate core particles Current number of bytes in queue Converted to queuing delay: ,in, This indicates the number of bytes that a single transmit slot can carry. Indicates candidate confluence cores Additional queuing slots required; simultaneously read candidate merge cores from the link state table. To the target core Expected number of blocked time slots Based on the aforementioned total expected inflow delay, expected long-distance transmission delay, and queuing delay, candidate merging cores are calculated. Total rendezvous delay: The principle of this formula is to uniformly convert the source-end merging delay, the post-merging transmission delay, the queuing delay at the merging point, and the forward blocking delay into a number of transmission time slots for comparison, so as to select the target merging core that is truly suitable for undertaking the merging task from multiple candidate merging cores; from the set of candidate merging cores In the middle, choose to make the total rendezvous delay The candidate merging core with the minimum value is selected as the target merging core. Finally, the candidate merging core that satisfies the following formula is selected as the target merging core: .

6. The communication path selection method for a multi-chip acceleration platform according to claim 5, characterized in that: Step S3 includes the following sub-steps: S3-1, For the target confluence core particle With target core Construct a candidate path set and in the A future assessment window Internally read the length of consecutive idle time slots for each link on each candidate path; S3-2. Using the link with the weakest continuous transmission capability on the candidate path as the representation of the continuous transmission capability of the entire candidate path, for any candidate path... The capacity of its continuous idle time slot interval is calculated according to the following formula: ,in Indicate candidate path Any link on, Indicates the link Future assessment window The number of consecutive idle transmission slots available within. Indicate candidate path The maximum number of consecutive transmission slots that can be transmitted without interruption. S3-3. Calculate the target fragment communication group according to the following formula. Number of continuous transmission demand time slots: ,in, This indicates the completion of the same target fragment communication group. The minimum number of consecutive transmission slots required for a single long-distance transmission; among all candidate paths that satisfy the following formula, select the candidate path with the smallest topology hop count as the main path. : When no candidate path meets the above conditions, select the one that... The candidate path that achieves the maximum value is selected as the priority candidate path for the next evaluation cycle, and this candidate path is written into the path candidate cache. Simultaneously, the same target fragment communication group is... The fragmented payloads that should have been merged and sent to the next short acquisition window will be retained for further evaluation to avoid sending them separately in the current cycle.

7. The communication path selection method for a multi-chip acceleration platform according to claim 6, characterized in that: Step S4 includes the following sub-steps: S4-1. First, calculate the target fragment communication group according to the following formula. At the target confluence core The latest estimated arrival time slot for completing the rendezvous: ,in Indicates the same target fragment communication group All fragment loads reached the target confluence core. The latest estimated time slot; then on the main path It is determined no earlier than The earliest continuous transmission start time slot This makes the main path Each link in the interval All of them have the conditions to be continuously transmittable. After obtaining the latest rendezvous completion time slot of the same target fragment communication group and the continuous transmission start time slot of the main path, the actual payload is not immediately transmitted. Instead, a path reservation control message is first sent to the intermediate forwarding nodes on the main path. S4-2, Constructing a path reservation control message: ,in, Indicates the target fragment communication group identifier; the path reservation control message Along the main path Send the data first, so that each intermediate forwarding node on the main path establishes a continuous forwarding record: ,in, Indicates the gap when continuous forwarding ends; S4-3, Each source chip will belong to the same target fragment communication group. The fragment payload is sent to the target confluence core. The target confluence core Indexed in order The fragmented payloads are sequentially merged to form a continuously transmitted message train, and this is done in the initial time slot. until the end of the gap within the interval along the main path Send to the destination chip in one go .

8. The communication path selection method for a multi-chip acceleration platform according to claim 7, characterized in that: Step S5 includes the following sub-steps: S5-1. Within the current assessment period, calculate the communication group of the same target fragment according to the following formula. Stability: ,in This indicates that the target chip identifier and service category remain consistent with the previous assessment period, and the chips continue to be classified into the same target fragment communication group. The number of communication requests to be sent. This indicates that the fragment belongs to the same target communication group during the current evaluation period. The total number of communication requests to be sent; S5-2. To avoid path jitter caused by rerouting in every evaluation cycle, first determine whether the current target fragment communication group is still suitable to continue using the rendezvous relationship and main path established in the previous cycle. Then, perform path preservation determination based on the following preservation conditions. When all three of the following conditions are met, maintain the current target rendezvous core. With the main path constant: , , ,in This represents the minimum stability threshold for path preservation. This indicates the maximum allowed queuing time slot threshold at the rendezvous point; S5-3, Surrounding the target core particle Pre-configured spare merge core assembly The spare confluence core assembly To surround the target core A pre-configured set of candidate merge core particles is used to perform a fast replacement selection when the current target merge core particle no longer meets the retention conditions. A backup merge switchover is performed when any one of the following four switchover conditions occurs: First, ;second, ; third, And continue The evaluation cycle; fourth, the number of failed path reservation control message transmissions. Reaching the threshold During standby merge handover, the standby merge core assembly is selected according to the following formula. Replacement cores selected from among them: and the alternative confluence core. Repeat steps S2 to S4; if the continuous transmission condition still cannot be met after the backup rendezvous handover, regenerate the candidate rendezvous core set. And execute a new round of complete path planning.

Images