A congestion-aware aggregated communication scheduling method and system for intelligent computing center networks
By constructing a communication cost graph and a group-structure joint closed loop, network congestion can be perceived in real time, and the aggregate communication scheduling can be optimized. This solves the communication bottleneck and long-tail effect caused by network congestion and improves the training efficiency of large-scale clusters.
Patent Information
- Application Number
- CN202611552252.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-09
- Publication Date
- 2026-11-17
AI Technical Summary
Existing ensemble communication scheduling methods cannot detect network congestion in real time, leading to communication bottlenecks and long-tail effects, especially affecting training efficiency in large-scale clusters, and lacking mutual exclusion coordination for congested edges in multi-task shared clusters.
By periodically collecting network status data to construct a communication cost graph, congestion labels are generated. Before executing aggregated communication, high-bandwidth domains are identified and reduced locally to generate initial group constraints. The minimum communication structure is selected to form a group-structure joint closed loop. Scheduling is performed by combining bucket granularity and task priority to ensure mutual exclusion time slot reservation and dual feedback adjustment during congestion.
It effectively reduces the impact of network congestion on communication, shortens the critical path across domains, reduces long-tail uncertainty, and improves training efficiency and resource utilization.
Smart Images

Figure CN123123261A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of network communication technology, and in particular to a method and system for collective communication scheduling with network congestion awareness in intelligent computing centers. Background Technology
[0002] Distributed training is the dominant paradigm for large-scale artificial intelligence model training. Each iteration typically requires ensemble communication operations, such as global gradient reduction, between multiple computing nodes participating in the training to synchronize gradient information and update parameters. Accompanied communication is characterized by massive communication volume and strong synchronization: all participating nodes must complete the current round of communication before proceeding to the next iteration. Congestion in any network link will slow down the completion time of the entire iteration, resulting in a significant long-tail effect in communication. The larger the cluster size, the more significant this long-tail effect becomes on the overall training efficiency.
[0003] Existing aggregated communication libraries typically employ fixed communication algorithms and static node grouping schemes that are independent of the real-time network status. When a link in the network experiences temporary congestion or slowdown due to other traffic, hardware aging, or cable quality issues, the fixed algorithm and fixed grouping cannot detect this change and continue to execute communication according to the preset path and structure, causing the congested link to become the bottleneck of the entire round of communication. Current solutions generally adopt serial communication or bucketing adjustment methods based solely on a single factor of bandwidth / computation time, failing to incorporate path congestion status into the bucketing granularity closed loop. In multi-task shared cluster scenarios, different task communication flows compete for resources on shared links, lacking mutual exclusion coordination for congested edges.
[0004] Among publicly available approaches to aggregated communication scheduling and traffic engineering, TE-CCL and TACCL abstract aggregated communication into a multi-commodity flow / traffic engineering or topology-aware synthesis problem, exhibiting off-line or semi-static one-time output scheduling and lacking joint grouping-structure back-injection. MSCCL, SCCL, TACOS, and Blink focus on algorithm synthesis, time-extended matching, or path construction, belonging to the synthesis / construction layer, but lacking orchestration closure. AdapCC and others analyze and reconstruct the aggregated communication graph at runtime, primarily focusing on continuous bandwidth / latency optimization; NetHint and others provide tenants with white-box bandwidth hints to select overlay topologies, but these hints are soft hints rather than hard constraints, lacking failure degradation closure. Therefore, how to use a communication cost graph with congestion annotations as a unified online state source to connect intra-domain reduction and cross-domain hard feasible domain structure selection, grouping-structure joint closure, bucketed dual feedback, and multi-task congestion edge mutual exclusion reservation into a failure-degradable runtime scheduling causal chain is a pressing technical problem that needs to be solved. Summary of the Invention
[0005] To overcome the above problems and shortcomings, this invention proposes a congestion-aware aggregated communication scheduling method for intelligent computing center networks, comprising the following steps:
[0006] Step S1: Periodically collect queue depth, bandwidth utilization of each link, and end-to-end latency metrics for each switch in the intelligent computing center network, and construct a communication cost graph based on the cluster physical topology. Where V is the set of compute nodes and switch vertices, and E is the set of physical link edges; for each edge e∈E, the edge weight is calculated by weighting the normalized queue depth, bandwidth utilization, and end-to-end delay. And generate congestion labels according to preset rules. ∈{0,1}, where the congested edge corresponds to =1; The communication cost map is refreshed incrementally with the acquisition cycle; When the telemetry index expires or is missing, the previous valid cost map is used and the telemetry degradation state is set;
[0007] Step S2: Before executing the aggregated communication task, identify high-bandwidth domains and perform local reduction on participating nodes within the domain to compress the amount of cross-domain data transmission; for the cross-domain transmission stage, first generate initial grouping constraints from the congested connectivity components and high-weight cut edges in the communication cost graph, and then calculate the total structural cost of each candidate communication structure such as ring, tree, and hierarchical structures under these constraints. and forcibly excluded An edge with a value of 1 is in the feasible region. Solve the above Minimal communication structure ;
[0008] like If the space is empty, it enters the airspace degradation state, adopts a safe ring structure that avoids congestion edges or performs flow limiting, and records the degradation state.
[0009] Step S3: Select the communication structure in step S2 Next, the packet boundaries are refined based on the communication cost graph: Within the rack, the boundaries are formed by... When the congestion-connected component induced by the edge with weight = 1 is split into multiple subgroups, the lightly loaded neighborhood can be merged when there are no high-weight cut edges and the total structural cost does not increase after merging; the refined grouping result is back-injected as the input constraint for the candidate structure enumeration in step S2 of the next scheduling cycle, forming a group-structure joint closed loop; it is forbidden to divide the group boundary only according to the physical rack static division.
[0010] Step S4: Divide the gradient data to be synchronized into buckets according to model layers. For the buckets corresponding to layers that have completed gradient calculation during backpropagation, immediately initiate aggregate communication and run them in parallel with the calculation process of layers that have not yet completed gradient calculation. The bucketing granularity is based on the overlap ratio. The congestion integral of the scheduled edges on the cost graph Joint regulation: when Reduce the number of bins when the value is too low; when When the current is too high, reduce the number of buckets or limit the flow; when Approaching the upper limit and When the value is too low, the number of buckets can be increased to reduce scheduling overhead and form a double feedback closed loop.
[0011] Step S5: When multiple training tasks are undertaken on the same physical cluster, staggered communication windows are allocated according to the priority and iteration rhythm of each task; Shared edges with a value of 1 are reserved for mutual exclusion time slots; when there is a window conflict, high-priority preemption occurs, while low-priority delay or reduction of concurrent groups and backpressure.
[0012] It should be noted that the causal order of steps S2 and S3 forms a joint closed loop that cannot be arbitrarily interchanged: (i) the initial grouping constraint is generated solely from the congested connected components and high-weight cut edges in the communication cost graph, without depending on the selected final communication structure; (ii) then, under this constraint, within the feasible region... Solve above (iii) Step S3 in the selected The above grouping and refining process is only used as the input constraint for step S2 of the next scheduling cycle, and is not used to rewrite the selected steps within the current cycle. This sequence differs from synthesizers like TACCL, TE-CCL, and TACOS, which produce offline or semi-static schedules in a single operation, and from synthesizers like AdapCC, which drive whole-graph reconstruction with continuous metrics and do not rely on static scheduling. The runtime modification diagram, which uses the hard feasible region and spatial domain degradation state machine as its core, is also different from the hierarchical heuristic that is only statically divided according to the physical rack. The above comparisons are all descriptions of structural differences and do not constitute performance assertions for any publicly available system.
[0013] Furthermore, the edge weights described in step S1 are calculated according to the formula... = + + Calculation, where , , These represent the normalized queue depth, bandwidth utilization, and end-to-end latency of edge e, respectively. , , The weight coefficients are non-negative and + + =1; the stated The depth is preferentially obtained by normalizing the egress or ingress queue depth of the switch; It can be obtained by normalizing one or more of the following: link bandwidth utilization, explicit congestion notification (ECN) tagging rate, congestion notification (CNP) arrival frequency, in-band network telemetry (INT) link occupancy rate, or the gap between network interface card (NIC) throughput and rated bandwidth; It can be obtained by normalizing the end-to-end probe delay, round-trip delay, or their sliding window statistics; this step consumes the above telemetry and congestion signals to construct a cost graph without rewriting the transport layer congestion control loop itself; when any component expires or is missing, the previous valid component on that edge is used. , ) and set the telemetry to a degraded state; the condition for exiting the telemetry degraded state is: when the telemetry becomes valid again, stop using the previous valid edge ( , And updated according to current telemetry; the congestion label is preferably updated using a dual-threshold hysteresis update: = ( ≥ ), For a single congestion threshold: when ≥ Time setting =1, when ≤ Time setting =0, when < < Maintain the previous cycle Unchanged, among which , Preset upper and lower thresholds and > The incremental refresh includes updating only the edge weights and congestion labels that have changed in telemetry within the current period.
[0014] Further, the total structural cost described in step S2 is defined as... = ,in For candidate communication structures, for The set of edges occupied; the feasible region ={ ∈ ∣∀e∈ , =0}, = ;in It is a set of candidate communication structures; hard exclusion means prohibiting the use of any one of them. An edge with a value of 1, or equivalently, an edge congested with a prohibition cost. The edge weights are incorporated and then minimized, where M is a prohibition constant much larger than the magnitude of ordinary edge weights; the secure ring structure is in A ring communication topology is constructed on the edge set where = 0; the rate limiting is performed to reduce cross-domain concurrent traffic or lengthen communication time slots; when When the space is empty, the system enters a degraded airspace state: it employs the aforementioned safety ring structure or rate limiting, and broadcasts a degraded flag to the bucketing and multi-task modules; the condition for exiting the degraded airspace state is: after a double-threshold hysteresis update, a non-empty space exists. And on the critical path =1 edge has been cleared or is below the preset recovery count; candidate set The structure and rules can be derived from the union of ring, tree, and hierarchical structures provided by the offline topology-aware synthesis library; this step is based on the runtime... right Perform hard feasible region pruning and output Instead of a one-time offline scheduling, it is a degraded state.
[0015] Furthermore, the candidate set in step S2 and total structural cost Differentiated by the type of set communication primitives: reduction-broadcast primitives for AllReduce or equivalents. according to Sum the edge weights of occupied edges and prioritize enumerating circular, tree, and hierarchical candidates; for the Reduce-Scatter and AllGather pipeline combinations, candidates... It can include phased heterogeneous components and accumulate the edge weights of each phase separately; for AlltoAll primitives, it can improve the sensitivity of many-to-one convergence edges. Effective weights or priority hard exclusion in =1 and is located at the convergence point, then in Solve above The above differentiation remains unchanged. =1 Hard exclusion and spatial degradation state machine.
[0016] Furthermore, when the physical topology undergoes edge additions or deletions, node hot migrations, or link jitter, the communication cost graph only performs local incremental updates and recalculations on the affected edges and their adjacent cut edges. Grouping constraints and The replan is by Hard constraint trigger: When a new constraint is added =1 edge falls into the current Occupied set or Re-enumerate immediately when the graph changes from non-empty to empty, instead of re-optimizing the entire graph just because of small fluctuations in continuous edge weights; re-planning and... Updates are subject to / Hysteresis constraints are applied, and a minimum hold time is set to avoid structural overturning oscillations.
[0017] Furthermore, the initial grouping constraints described in step S2 are derived solely from the cost graph. The congested connected components induced by edges with weights equal to 1, and the high-weight cut edges connecting different group candidates with edge weights exceeding a preset cut edge threshold, are determined without relying on the already selected final communication structure. The refining process in step S3 includes: splitting the rack into multiple sub-groups when there are congested connected components within the rack; and refining the current lightly loaded neighborhoods when there are no high-weight cut edges between adjacent lightly loaded neighborhoods and the merged neighborhoods are in the current... Merging is performed when the total structural cost in the sense of the time does not increase; the refinement result only serves as an input constraint for the candidate structure enumeration in step S2 of the next scheduling cycle, and the selected structure is not rewritten within the cycle. Together with step S2, they form a grouping-structure joint closed loop, rather than a unidirectional static partition.
[0018] Furthermore, the overlap ratio mentioned in step S4 Defined as = / max( , ),in The overlap time between communication and parallel computing, For the execution duration of the communication task, The congestion integral represents the parallel computation time. Defined as = or = Where P is the set of edges traversed by the currently scheduled communication path; bucket granularity according to ← · Updated, the segmentation rule is: if < or > Then f = <1; if ≥ and ≤ Then f = >1; otherwise f=1; , For the lower and upper bounds of the overlapping targets, , For congestion warnings at lower and upper limits, , The preset adjustment coefficient is used; when in telemetry degradation or airspace degradation state, it is prohibited to press [the button]. Enlarge the buckets and force a conservative adjustment to keep f≤1 until the corresponding downgraded state exits.
[0019] Furthermore, the mutual exclusion time slot reservation mentioned in step S5 only applies to... Establish a time slot occupancy table for shared edges with a value of 1. Edges with a value of 0 are not subject to mutual exclusion; at most one set of tasks is allowed to occupy the congested edge in the same time slot, i.e., for ∀e: =1 and ∀time slot s satisfy ≤1, where ∈{0,1} indicates whether task t occupies edge e in time slot s, where T is the set of concurrent tasks; when pre-allocation would violate the above mutual exclusion constraint, the high-priority task preempts the conflicting time slot already occupied by the low-priority task; the backpressure is coupled to step S4: notifying the preempted or delayed low-priority task to modify the bucket granularity. (according to Reduce or freeze the amplification), reduce the number of concurrent groups or lower the bucket trigger frequency to form a closed loop between multi-task mutual exclusion decision and single-task bucket dual feedback; tasks in the airspace degradation or telemetry degradation state are prohibited from preempting the upgrade cross-domain concurrent bandwidth.
[0020] Furthermore, the local reduction described in step S2 requires a cross-domain structure selection solution. Previously completed to ensure that only representative data after reduction within the domain is transmitted during the cross-domain phase; the local reduction uses summation or average aggregation; the high-bandwidth domain is a set of nodes identified based on the physical topology hierarchy and the distribution of edge weights within the domain in the communication cost graph, specifically a set of nodes mounted on the same access switch or in the same rack within a high-speed interconnection domain, and whose edge weights within the domain are all below a preset domain threshold; the candidate communication structure includes at least two of the following: ring structure, tree structure, and hierarchical structure, and the candidate set... Must pass =1 Hard exclusion from falling into the feasible region Output can be done later ;when When the airspace is empty, it enters a degraded airspace state, adopts a safe ring structure that avoids congestion edges or performs flow limiting, and records and broadcasts the degraded state.
[0021] This invention also proposes a congestion-aware aggregated communication scheduling system for intelligent computing center networks, used to execute the above method, comprising: a congestion-labeled cost graph module for executing step S1; an intra-domain reduction and cross-domain structure selection module for executing step S2; a cost graph-driven dynamic grouping module for executing step S3; a bucketing overlap and dual-feedback scheduling module for executing step S4; and a multi-task peak shifting and mutual exclusion reservation module for executing step S5; wherein, the congestion-labeled cost graph module outputs a communication cost graph with edge weights and congestion labels and telemetry degradation status to the other modules; the intra-domain reduction and cross-domain structure selection module outputs... Alternatively, the spatial domain may be in a degraded state, and a degraded flag may be broadcast to the bucketing and multi-task modules; a grouping-structure joint closed-loop interface may be established between the intra-domain reduction and cross-domain structure selection module and the cost graph-driven dynamic grouping module: the former generates initial grouping constraints from the cost graph and outputs them. The latter uses the refining result as the input constraint for the next cycle; the bucket overlap and dual feedback scheduling module receives the overlap ratio. Congestion Integral It also writes back the bucket granularity; the multi-task peak shifting and mutual exclusion reservation module maintains the congestion edge time slot occupancy table and outputs the preemption backpressure signal.
[0022] The present invention also provides a computer-readable storage medium on which a computer program stored is executed by a processor to implement the above-described method.
[0023] To achieve the above-mentioned congestion-aware scheduling, this invention relates to the following main symbols: the communication cost diagram is denoted as... The queue depth, bandwidth utilization, and end-to-end latency of edge e, after normalization, are denoted as follows: , , The edge weight is The congestion threshold is (or dual threshold) , The threshold within the domain is denoted as (with the edge cut threshold) (Distinguish), congestion label is ∈{0,1} and congested edges correspond to =1; the candidate communication structure is denoted as Its occupied edge set is The total cost of the structure is The candidate set is The feasible region is The overlap ratio is denoted as The communication duration is recorded as The parallel computing time is denoted as The overlap duration is recorded as The set of scheduled path edges is denoted as P, and the congestion integral is denoted as... The particle size of the bins is recorded as The adjustment function is denoted as f; in a multi-task scenario, the task set is denoted as T, and the task priority is denoted as f. The off-peak window is recorded as The occupancy indicator variable is denoted as ∈{0,1}.
[0024] The edge weight calculation is written as Equation (1), the congestion label is written as Equation (2), the total structural cost is written as Equation (3), the optimal structure selection is written as Equation (4), the overlap ratio is written as Equation (5), the congestion integral is written as Equation (6), the bucketed double feedback update is written as Equation (7), and the multi-task mutual exclusion constraint is written as Equation (8):
[0025] Equation (1): w e = Q e + U e + D e
[0026]
[0027] = + + ,in , , ≥0 and + + =1, , , For telemetry components normalized to [0,1].
[0028] Equation (2): e = (w e ≥ ); lag e ← H(w e ; , , e (prev)
[0029]
[0030]
[0031] Simplified form is = ( ≥ That is, when the edge weight reaches or exceeds a preset threshold. The congestion label is set to 1 if the time threshold is not met, and 0 otherwise; the preferred double-threshold hysteresis form is: if ≥ but =1, if ≤ but =0, if < < but Maintain the value from the previous cycle to suppress The shaking caused the structure to flip frequently.
[0032] Equation (3): = w e
[0033]
[0034] For candidate structures Define the total cost of the structure = .
[0035] Equation (4): = , = { ∈ ∣ ∀e∈ , e =0}
[0036]
[0037]
[0038] feasible region ={ ∈ ∣∀e∈ , =0}; where Given a set of candidate communication structures; solve over the feasible region. = ;like If empty, then it will be downgraded to... Construct a safe ring or perform rate limiting on the edge set = 0, and record the degradation state. Hard exclusion can also be equivalently expressed as replacing the edge weights with + Then minimize it, where the constant M is much larger than the magnitude of ordinary edge weights.
[0039] Equation (5): = / max( , )
[0040]
[0041] Overlap ratio = / max( , ).
[0042] Equation (6): = w e
[0043]
[0044] Congestion Indicators = Alternatively, congestion label counting can be used. = .
[0045] Equation (7): ← ·
[0046]
[0047] ← · The segmentation rule is: if < or > Then f = <1; if ≥ and ≤ Then f = >1; otherwise f=1.
[0048] Equation (8): ≤ 1
[0049]
[0050] For any =1 shared edge and any time slot s, ≤1; In case of conflict, high-priority preemption, low-priority delay or reduction of concurrency and backpressure.
[0051] Beneficial effects:
[0052] This invention uses a communication cost graph with edge weights and congestion labels as a unified online state source, and explicitly downgrades and reuses it when telemetry expires / is missing. This ensures that subsequent structure selection, grouping, bucketing, and multi-task coordination are all based on verifiable network states, overcoming the problem that fixed algorithms and static grouping are unaware of congestion. Compared to offline synthesis routes such as TE-CCL / TACCL, this method outputs runtime scheduling decisions and a degradation state machine, which can be superimposed on the candidate set of the synthesis library. Compared to runtime continuous graph profiling methods such as AdapCC, this method uses... The hard feasible region ∈{0,1} and the spatial degradation state machine are the central hubs, and with Hard constraints trigger reprogramming rather than just full graph re-optimization of continuous edge weights.
[0053] This invention constructs a hard feasible domain in the cross-domain stage. And forcibly excluded When an edge equal to 1 has no feasible solution, it is downgraded to a safe loop or current limiting, forming a failable closed loop, which is different from solutions that only perform soft "avoidance" or simply change the algorithm name without hard constraints; the entry / exit and internal constraints of spatial domain degradation and telemetry degradation are as follows: Figure 7 As shown; dual-threshold hysteresis further suppresses The shaking caused the structure to flip frequently.
[0054] This invention enables grouping and structure selection to form a joint closed loop: first, initial grouping constraints are obtained from the cost graph, and then the results are calculated under these constraints. Then, perform another grouping and refinement based on the selected structure and reinject it into the next cycle, prohibiting static division based solely on physical racks. In congested or lightly loaded neighborhood scenarios within a rack, the joint closed loop shortens the cross-domain critical path.
[0055] This invention places the particle size of the different buckets simultaneously on the overlap ratio. Congestion Integral Under the dual feedback mechanism, and by formalizing f in a segmented rule format, it avoids excessive amplification of bucketing on congested paths due to single-factor adjustment based solely on computation / communication time; in telemetry degradation state, it prohibits... Enlarging the buckets aligns with the failure narrative of the previous effective cost diagram.
[0056] This invention only applies to multi-tasking scenarios. =1 Shared edge is reserved for mutual exclusion time slots. In case of conflict, it will preempt and backpressure to modify the granularity of the bucket. Concurrent grouping, which differs from Cassini phase offset, Crux priority routing, and MLCC transmission layer rate interleaving routes in terms of control object and back pressure landing point; this orchestration layer mechanism can be orthogonally superimposed with RoCE / ECN and intra-switch aggregation, reducing long-tail uncertainty caused by disordered contention. Attached Figure Description
[0057] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the following description of the embodiments are briefly introduced. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0058] Figure 1 A schematic diagram of the overall process of the intelligent computing center network congestion awareness collective communication scheduling method provided in an embodiment of the present invention;
[0059] Figure 2 A schematic diagram of the structure of the intelligent computing center network congestion awareness aggregate communication scheduling system provided in an embodiment of the present invention;
[0060] Figure 3 This is a schematic diagram of the congestion labeling communication cost graph and the dual-threshold hysteresis in an embodiment of the present invention;
[0061] Figure 4 This is a schematic diagram of intra-domain reduction, cross-domain hard feasible domain selection, and grouping-structure joint closed loop in an embodiment of the present invention;
[0062] Figure 5 This is a schematic diagram of overlapping calculations and dual feedback adjustment in bucket communication according to an embodiment of the present invention;
[0063] Figure 6 This is a schematic diagram illustrating the mutual exclusion of multi-task peak-shifting windows and congestion edges in an embodiment of the present invention.
[0064] Figure 7 This is a schematic diagram of the normal scheduling and degradation state machine (entry / exit and degradation internal constraints) in an embodiment of the present invention. Detailed Implementation
[0065] The present application will be described below with reference to specific embodiments:
[0066] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0067] Example 1:
[0068] This embodiment mainly describes a congestion-aware aggregated communication scheduling method for intelligent computing center networks. See also... Figure 1The method outlined is shown below, and includes the following steps. Detailed steps are described below:
[0069] Step S1: Periodically collect queue depth, bandwidth utilization of each link, and end-to-end latency metrics for each switch in the intelligent computing center network, and construct a communication cost graph based on the cluster physical topology. Where V is the set of compute nodes and switch vertices, and E is the set of physical link edges; for each edge e∈E, the edge weight is calculated by weighting the normalized queue depth, bandwidth utilization, and end-to-end delay. And generate congestion labels according to preset rules. ∈{0,1}, where the congested edge corresponds to =1; The communication cost map is refreshed incrementally with the acquisition cycle; When the telemetry index expires or is missing, the previous valid cost map is used and the telemetry degradation state is set;
[0070] In this step, the telemetry acquisition period is preferably on the order of hundreds of milliseconds to seconds, for example, 500 ms. (This is feasible.) , , The mapping example is as follows: Normalized by the switch egress / ingress queue depth based on the port buffer limit or recent quantile; It can be weighted and normalized by one or more of the following: link bandwidth utilization, ECN tagging rate, CNP arrival frequency, link occupancy rate reported by INT, or the gap between network card throughput and rated bandwidth. Normalized by the round-trip time of the probe message or its sliding window mean / quantile. This step only consumes the above signals to construct the cost graph and does not rewrite the rate update logic of the transport layer congestion control loop. Combined with the pre-acquired physical topology (compute nodes, access switches, aggregation / core switches, and links), a... For each edge, , , Normalize to [0,1] respectively, and then calculate according to equation (1). Generate according to formula (2) ∈{0,1}. For example, take... =0.4 =0.3 =0.3; Simplified threshold =0.75, or hysteresis threshold =0.80, =0.65 (within the domain threshold) Edge cutting threshold (Distinguishing between different components). Each acquisition cycle only incrementally refreshes edges where telemetry has changed; if any component of an edge is missing or its timestamp expires (e.g., after two acquisition cycles), the previous valid timestamp for that edge is used. , The global or edge state is set to telemetry downgrade. When edge additions, deletions, or node hot migrations occur in the physical topology, only the affected edges and their adjacent cut edges are locally refreshed and recalculated. ,and Both the overturning and structural replanning are affected / And a minimum hold time constraint to avoid flip-over oscillations. See also Figure 3 .
[0071] Step S2: Before executing the aggregated communication task, identify high-bandwidth domains and perform local reduction on participating nodes within the domain to compress the amount of cross-domain data transmission; for the cross-domain transmission stage, first generate initial grouping constraints from the congested connectivity components and high-weight cut edges in the communication cost graph, and then calculate the total structural cost of each candidate communication structure such as ring, tree, and hierarchical structures under these constraints. and forcibly excluded An edge with a value of 1 is in the feasible region. Solve the above Minimal communication structure ;like If the space is empty, it enters the airspace degradation state, adopts a safe ring structure that avoids congestion edges or performs flow limiting, and records the degradation state.
[0072] In this step, when a collective communication task (such as gradient global reduction in a certain iteration) needs to be executed, high-bandwidth domains are identified based on the physical topology hierarchy and edge weight distribution within the cost graph domain: for example, under the same access switch, and within the domain, each edge... All are below the intra-domain threshold =0.50 GPU node set; perform summation or average local reduction on the nodes within the subset to compress the intra-domain gradients into representative data to reduce the number of bytes across domains; local reduction must be performed in the cross-domain structure selection solution. Previously completed, candidate set After =1 Hard exclusion fallin Output can be done later .
[0073] For data that still needs to cross domains, strictly follow this order: First, according to the cost graph... =1 edge-induced congested connected components and high-weight cut edges generate initial grouping constraints, and then under these constraints, candidate structures such as rings, trees, and hierarchies are applied. Calculate according to formula (3) and according to formula (4) Solve above Hard requirements Not included An edge with a weight of 1 can also be equivalently written as... + After minimization, the example prohibits the constant M = 1 × 10² (much larger than the normalized edge weights). Candidate set This can be taken from the union of the output of offline topology-aware synthesis libraries such as TE-CCL / TACCL / MSCCL and regular ring / tree / hierarchical structures; this step only performs runtime analysis. Hard pruning and degradation do not replace the offline synthesizer itself. Example of differentiation by primitives: AllReduce prioritizes minimizing ∑ on rings / trees / hierarchies. Reduce-Scatter + AllGather can accumulate edge weights in stages; AlltoAll can improve the effective weight of many-to-one convergence sensitive convergence edges, but hard exclusion is still required first. =1 edge. If a feasible structure exists, choose... The smallest one as Execute; if If the space is empty, it enters a degraded airspace state: Construct a safe ring on the edge set where f = 0 (reduce the degree of participation if necessary), or limit cross-domain traffic (e.g., halve the number of concurrent flows or lengthen the time slots), and broadcast a degradation flag to the bucketing and multi-tasking modules; during degradation, f ≤ 1, and preemption of cross-domain concurrent bandwidth for upgrades is prohibited; when... / Non-empty reappears after lag And the critical path =When the count is cleared to zero or below the recovery count and the minimum hold time is met, exit the airspace degradation and allow re-enumeration. . Hard constraint-triggered reprogramming takes precedence over full graph re-optimization caused by small fluctuations in continuous edge weights, distinguishing it from graph modification strategies driven solely by continuous metrics. See also Figure 4 For details on entering / exiting the degraded state machine and the internal constraints of the degraded state machine, please refer to [link / reference needed]. Figure 7 Therefore, the structural selection is not a "soft preference for low-weight paths", but rather a failure degradation state machine with congestion labels as hard constraints and explicit entry / exit conditions.
[0074] Step S3: Select the communication structure in step S2 Next, the packet boundaries are refined based on the communication cost graph: Within the rack, the boundaries are formed by... When the congestion-connected component induced by the edge with weight = 1 is split into multiple subgroups, the lightly loaded neighborhood can be merged when there are no high-weight cut edges and the total structural cost does not increase after merging; the refined grouping result is back-injected as the input constraint for the candidate structure enumeration in step S2 of the next scheduling cycle, forming a group-structure joint closed loop; it is forbidden to divide the group boundary only according to the physical rack static division.
[0075] In this step, the group boundaries are no longer statically divided by physical rack name, but instead form a joint closed loop with step S2. Once selected, if a certain rack contains an issue caused by… =1 edge-induced congested connected components, then the rack is split into multiple subgroups to prevent the congested rack from being treated as a single giant node and concentrated in the cross-domain phase. For adjacent lightly loaded neighborhoods, if the connecting edges are not high-weight cut edges (e.g. Below the edge cutting threshold =0.45) and after merging, in the current If the total structural cost does not increase in the sense of the time, temporary merging is allowed. The refined grouping results are fed back into the candidate enumeration constraints of step S2 of the next scheduling cycle through the grouping-structure joint closed-loop interface, so that "what structure to choose" and "how to group" are driven by the cost map and iteratively interlocked, and the two are prohibited from being disconnected or only statically divided according to the rack.
[0076] Step S4: Divide the gradient data to be synchronized into buckets according to model layers. For the buckets corresponding to layers that have completed gradient calculation during backpropagation, immediately initiate aggregate communication and run them in parallel with the calculation process of layers that have not yet completed gradient calculation. The bucketing granularity is based on the overlap ratio. The congestion integral of the scheduled edges on the cost graph Joint regulation: when Reduce the number of bins when the value is too low; when When the current is too high, reduce the number of buckets or limit the flow; when Approaching the upper limit and When the value is too low, the number of buckets can be increased to reduce scheduling overhead and form a double feedback closed loop.
[0077] In this step, the gradients to be synchronized are divided into buckets according to the model layers. Backpropagation proceeds from the output layer to the input layer. When all the gradients in a bucket are ready, the collective communication of that bucket is immediately triggered (using the structure and grouping determined in steps S2 and S3), which runs in parallel with the computation of the layers that have not yet been completed, forming a communication-computation overlap.
[0078] Overlap ratio Calculated according to formula (5), the target interval is shown in the example [ , = [0.80, 0.95]. Congestion integral. Calculated according to formula (6); examples of the lower and upper boundaries of the warning are as follows: = The integral corresponding to the average edge weight of the path of 0.35 = The integral corresponding to the average edge weight of the path, 0.70. Update the bucket granularity according to equation (7). If measured in practice =0.72< Then let =0.8 Reduce the number of buckets to increase the trigger frequency; if > ,even though It's acceptable, but we should still reduce the number of data buckets or limit the communication flow to avoid large bursts accumulating at the congestion edge; if It has reached 0.96≥ and ≤ Then let =1.15 Increase the number of buckets to reduce scheduling overhead. When in a telemetry degradation state or airspace degradation state, it is prohibited to press... Increase the number of buckets, forcing f ≤ 1 (a conservative approach), until the corresponding downgraded state exits. See also Figure 5 This dual-feedback closed loop is the key element that distinguishes this step from "adjusting buckets solely based on bandwidth and computation time".
[0079] Step S5: When multiple training tasks are undertaken on the same physical cluster, staggered communication windows are allocated according to the priority and iteration rhythm of each task; Shared edges with a value of 1 are reserved for mutual exclusion time slots; when there is a window conflict, high-priority preemption occurs, while low-priority delay or reduction of concurrent groups and backpressure.
[0080] In this step, when the cluster runs multiple training tasks simultaneously or alternately, priority labels are read from the job scheduling system. Based on the recent average iteration time of each task, the iteration rhythm is estimated, and staggered windows are allocated. Mutual exclusion slot reservations only apply to the cost graph. Shared edges with a value of 1: According to equation (8), at most one task should occupy this edge in the same time slot; Edges with a value of 0 are not subject to mutual exclusion. If pre-allocation violates equation (8), then higher priority tasks preempt conflicting time slots; the preempted lower priority tasks are deferred to the next available window and back-coupled to step S4: Reduce its bin size Alternatively, it can freeze and amplify, reduce the number of concurrent groups, or lengthen the bucket trigger interval. Tasks in telemetry or spatial degradation states are prohibited from preemptively upgrading cross-domain concurrent bandwidth. This mechanism is used at the aggregated communication orchestration layer, controlling congestion-side time slots and bucket parameters, and can be interleaved or orthogonally superimposed with phase offset-type peak shifting or transport layer rate scaling. See also Figure 6 .
[0081] Example 2:
[0082] See Figure 2 This embodiment introduces a congestion-aware aggregated communication scheduling system for intelligent computing center networks, used to execute the method described in Embodiment 1. The system mainly includes the following modules.
[0083] The congestion labeling cost graph module is used to perform step S1, periodically collect telemetry data, calculate edge weights and congestion labels (including optional double threshold hysteresis), incrementally refresh the cost graph, and set the telemetry data to a degraded state when telemetry data expires or is missing.
[0084] The intra-domain reduction and cross-domain structure selection module is used to execute step S2, completing high-bandwidth domain local reduction, calculation of the total cost of candidate structures based on initial grouping constraints, and hard exclusion (or prohibition cost) of congested edges. )as well as Airspace degradation state machine when empty (entry / exit conditions, broadcast degradation flags to buckets and multitasking; see also) Figure 7 ).
[0085] The cost graph-driven dynamic grouping module is used to execute step S3, in The above is based on the congested connected components and high-weight cut edges for grouping and refining, and the results are injected back into the next cycle structure selection through the joint closed-loop interface.
[0086] The bucketing overlap and dual feedback scheduling module is used to execute step S4, which triggers communication by layer-by-layer bucketing—calculates overlap, and based on... and Update the granularity of the buckets according to the segmentation rules.
[0087] The multi-task peak-shifting and mutual exclusion reservation module is used to execute step S5, allocate peak-shifting windows, and only apply to... =1 side maintains the time slot occupancy table, and outputs preemption and backpressure to the bucket granularity when conflicts occur. / Signals for concurrent groups.
[0088] In terms of deployment, the cost graph module can be deployed on the same machine as the network telemetry collector or in a bypass configuration; structure selection, dynamic grouping, and bucket scheduling can be embedded into the communication backend of the aggregate communication library or training framework; the multi-task coordination module can interface with the cluster job scheduler. Modules exchange cost graph snapshots, degradation status, grouping schemes, and backpressure signals through shared memory, message queues, or RPC.
[0089] Example 3:
[0090] This embodiment provides a simplified numerical example of cost icon annotation, cross-domain structural hard constraint selection, grouping-structure joint closed loop, bucketed dual feedback, and multi-task mutual exclusion reservation (example parameters, which do not constitute a limitation on the scope of protection).
[0091] Scenario 1 (Cost Graph and Congestion Labels): Let the normalized telemetry values for links e1, e2, and e3 be (Q, U, D) = (0.80, 0.70, 0.60), (0.30, 0.40, 0.35), and (0.90, 0.85, 0.80), respectively. =0.4 =0.3 =0.3, then =0.4×0.80+0.3×0.70+0.3×0.60=0.32+0.21+0.18=0.71, =0.4×0.30+0.3×0.40+0.3×0.35=0.12+0.12+0.105=0.345, = 0.4 × 0.90 + 0.3 × 0.85 + 0.3 × 0.80 = 0.36 + 0.255 + 0.24 = 0.855. If a simplified threshold is used... =0.75, then = (0.71≥0.75)=0、 =0、 = (0.855 ≥ 0.75) = 1. If hysteresis is used... =0.80, =0.65 and the previous period =0, therefore 0.65 < 0.71 < 0.80, Keeping it at 0 can suppress jitter near the threshold. If telemetry for e3 is missing in the next cycle, then () will be used. , = (0.855, 1) and set the telemetry downgrade state.
[0092] Scenario 2 (Hard exclusion and downgrading of cross-domain structures): Candidate ring Occupying {e1,e2,e3}, because it contains =1 for e3 is forcibly excluded; candidate tree Occupying {e1,e2}, C = 0.71 + 0.345 = 1.055; Candidate stratification Occupy {e2} and another line =0 edge e4 (w = 0.40), C = 0.345 + 0.40 = 0.745. Therefore, choose If all candidates inevitably result in e3, then If empty, proceed with airspace degradation: Construct a safety ring on {e1,e2,e4}, broadcast the degradation flag, and force bucketing f≤1; (Wait for...) After delay, the value was cleared and After restoring non-empty values, exit the downgrade and then enumerate again.
[0093] Scenario 3 (Grouping-Structured Joint Closed Loop): This cycle begins with... =1 connected components and high-weight cut edges are used to obtain initial grouping constraints, and based on this, enumeration is performed. = Subsequently, an issue appeared inside rack R1. =1 connected component, split R1 into subgroups G1a and G1b; the edge weights between adjacent racks R2 and R3 are all < =0.45, after merging into G23 Under this meaning, C does not increase, and the merge is adopted. The refined result is fed back to the next cycle step S2, and enumeration is only performed on the updated inter-group links. .
[0094] Scenario 4 (Dual Feedback in Bucket): Retrieval =0.80, =0.95, The average edge weight of the corresponding path is 0.35. Corresponding to 0.70. Initial. =4 layers / bucket, measured =0.72< and > ,Pick =0.75, Update ←3; then Rising to 0.91 (still < )and Falling back to The following conditions were not met simultaneously. ≥ and ≤ This beat f=1 remains =3; then measured =0.96≥ and ≤ ,Pick =1.2, ←3.6≈4, reducing scheduling overhead. If only... Adjustment and neglect If congestion persists, the number of data buckets may be incorrectly increased, exacerbating the outbreak. If the current cycle is in a telemetry downgrade or airspace downgrade state, then even if... ≥ It is also forbidden to press Increase the size and force f ≤ 1.
[0095] Scenario 5 (Mutual Exclusion Reservation and Preemption Backpressure): Task and shared =1 edge e3. The time slot occupancy table has been registered in time slot s5. ; Upon arrival, s5 will be seized due to violation of equation (8). Delay to step S8, and back-press to step S4: divide the particles into smaller barrels. according to = 0.8 reduction, concurrent grouping reduced from 4 to 2, bucket trigger interval lengthened, until e3 Reset or stagger the windows. Edges with a value of 0 are not added to the mutex list. See also Figure 6 The illustrated values are shown in the figure.
[0096] Scenario 6 (Topology jitter and) Trigger Replan): Set the current Occupying {e2, e4}. Link e4 jitter during a certain period causes... Rising to 0.82 and after Place =1, because Hard constraint falls into Occupied set, immediately after the update The above is an enumeration; if only exist and If there are slight fluctuations between them, then maintain... and Unchanged. When adding or deleting edges, only the affected edges and adjacent cut edges are locally refreshed, and the minimum retention time is followed to avoid structural flipping and oscillation.
[0097] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other.
Claims
1. A congestion-aware aggregated communication scheduling method for intelligent computing center networks, characterized in that, Includes the following steps: Step S1: Periodically collect queue depth, bandwidth utilization of each link, and end-to-end latency metrics of each switch in the intelligent computing center network, and construct a communication cost graph based on the cluster physical topology. , where V is the set of compute nodes and switch vertices, and E is the set of physical link edges; For each edge e∈E, the edge weight is calculated by weighting it with the normalized queue depth, bandwidth utilization, and end-to-end delay. And generate congestion labels according to preset rules. ∈{0,1}, where the congested edge corresponds to =1; The communication cost map is refreshed incrementally with the collection cycle; When telemetry indicators expire or are missing, the previous effective cost map is used and the telemetry degradation status is set. Step S2: Before executing the aggregated communication task, identify high-bandwidth domains and perform local reduction on participating nodes within the domain to compress the amount of cross-domain data transmission; For the cross-domain transmission phase, initial packet constraints are first generated from the congested connectivity components and high-weight cut edges in the communication cost graph. Then, under these constraints, the total structural cost of each candidate communication structure, such as ring, tree, and hierarchical structures, is calculated. and forcibly excluded An edge with a value of 1 is in the feasible region. Solve the above Minimal communication structure ; like If the space is empty, it enters the airspace degradation state, adopts a safe ring structure that avoids congestion edges or performs flow limiting, and records the degradation state. Step S3: Select the communication structure in step S2 Next, the packet boundaries are refined based on the communication cost graph: Within the rack, the boundaries are formed by... When congestion is induced by an edge with weight = 1, the component is split into multiple subgroups. Lightly loaded neighborhoods can be merged when there are no high-weight cut edges and the total structural cost does not increase after merging. The refined grouping results are fed back into the input constraints for the candidate structure enumeration in step S2 of the next scheduling cycle, forming a grouping-structure joint closed loop; Group boundaries must not be defined solely based on physical racks; Step S4: Divide the gradient data to be synchronized into buckets according to model layers. For the buckets corresponding to the layers that have completed gradient calculation in backpropagation, initiate set communication immediately and run them in parallel with the calculation process of the layers that have not yet completed gradient calculation. The particle size of the bins is determined by the overlap ratio. Congestion integral of scheduled edges on the cost graph Joint regulation: when Reduce the number of bins when the value is too low; when If the flow rate is too high, reduce the number of buckets or limit the flow. when Approaching the upper limit and When the value is too low, the number of buckets can be increased to reduce scheduling overhead and form a double feedback closed loop. Step S5: When multiple training tasks are undertaken in the same physical cluster, staggered communication windows are allocated according to the priority and iteration rhythm of each task; right Shared edges with a value of 1 are reserved for mutual exclusion time slots; When there is a window conflict, the high-priority window preempts the low-priority window, while the low-priority window delays or reduces the number of concurrent groups and applies back pressure.
2. The method according to claim 1, characterized in that, The edge weights described in step S1 are calculated as follows: = + + Calculation, where , , These represent the normalized queue depth, bandwidth utilization, and end-to-end latency of edge e, respectively. , , The weight coefficients are non-negative and + + =1; The It is primarily obtained by normalizing the depth of the switch's egress or ingress queues; The It can be obtained by normalizing one or more of the following: link bandwidth utilization, explicit congestion notification ECN marking rate, congestion notification message CNP arrival frequency, in-band network telemetry (INT) link occupancy rate, or the gap between network interface card throughput and rated bandwidth. The It can be obtained by normalizing the end-to-end detection delay, round-trip delay, or their sliding window statistics; This step consumes the above telemetry and congestion signals to construct a cost graph without rewriting the transport layer congestion control loop itself. When any component expires or is missing, the previous valid component on that edge is used. , Concurrently set the telemetry to a degraded state; The condition for exiting the telemetry degradation state is: when telemetry becomes valid again, the previous valid edge is no longer used. , And update according to the current telemetry data; The congestion label is preferably updated using a dual-threshold hysteresis method: = ( ≥ ), For a single congestion threshold, That is when ≥ Time setting =1, when ≤ Time setting =0, when < < Maintain the previous cycle Unchanged, among which , Preset upper and lower thresholds and > ; The incremental refresh includes updating only the edge weights and congestion labels that have changed in telemetry within the current period.
3. The method according to claim 2, characterized in that, The total structural cost in step S2 is defined as follows: = ,in For candidate communication structures, for The set of edges occupied; The feasible region ={ ∈ ∣∀e∈ , =0}, = ;in A set of candidate communication structures; Hard exclusion means prohibiting the use of any of them. An edge with a value of 1, or equivalently, an edge congested with a prohibition cost. Minimize after incorporating edge weights, where M is a prohibition constant that is much larger than the magnitude of ordinary edge weights; The safety ring structure is in A ring communication topology constructed on the edge set where = 0; The rate limiting is implemented to reduce cross-domain concurrent traffic or lengthen communication time slots; when When the airspace is empty, it enters the airspace degradation state: it adopts the aforementioned safety ring structure or flow limiting execution, and broadcasts the degradation flag to the bucket and multi-task modules; The condition for exiting the airspace degradation state is: there is a non-empty area after the double threshold hysteresis update. And on the critical path =1 side has been cleared or is below the preset recovery count; Candidate set It can be derived from the union of structures and rules such as rings, trees, and hierarchical structures provided by the offline topology-aware synthesis library; This step is based on the runtime... right Perform hard feasible region pruning and output Rather than a one-time offline adjustment, it is a downgraded state. Spend .
4. The method according to claim 3, characterized in that, Candidate set in step S2 and total structural cost Differentiated by the type of set communication primitives: reduction-broadcast primitives for AllReduce or equivalents. according to Sum the edge weights of occupied edges and prioritize enumerating candidates for circular, tree, and hierarchical structures; For the combination of Reduce-Scatter and AllGather pipelines, candidate It can include phased heterogeneous components and accumulate the edge weights of each phase separately; For AlltoAll primitives, their sensitivity to many-to-one convergence edges can be improved. Effective weights or priority hard exclusion in =1 and is located at the convergence point, then in Solve above ; The above differentiation remains unchanged. =1 Hard exclusion and spatial degradation state machine.
5. The method according to claim 3, characterized in that, When the physical topology undergoes edge additions or deletions, node hot migrations, or link jitter, the communication cost graph only performs local incremental updates and recalculations on the affected edges and their adjacent cut edges. ; Grouping constraints and The replan is by Hard constraint trigger: When a new constraint is added =1 edge falls into the current Occupied set or The graph is re-enumerated immediately when it changes from non-empty to empty, rather than being re-optimized across the entire graph just because of small fluctuations in the weights of consecutive edges. Re-planning and Updates are subject to / Hysteresis constraints are applied, and a minimum hold time is set to avoid structural overturning oscillations.
6. The method according to claim 1, characterized in that, The initial grouping constraints mentioned in step S2 are determined solely from the cost graph. The congested connected components induced by edges with weights equal to 1, and the high-weight cut edges connecting different group candidates with edge weights exceeding a preset cut edge threshold, are determined without relying on the already selected final communication structure. ; The refining described in step S3 includes: splitting the rack into multiple subgroups when there is a congested connectivity component within the rack; When there are no high-weight cut edges between adjacent lightly loaded neighborhoods and they are merged, in the current... Merge if the total structural cost does not increase in the sense of the present; The refined result serves only as an input constraint for the candidate structure enumeration in step S2 of the next scheduling cycle. It is not used to rewrite the selected structure within the cycle. Together with step S2, they form a grouping-structure joint closed loop, rather than a unidirectional static partition.
7. The method according to claim 1, characterized in that, The overlap ratio mentioned in step S4 Defined as = / max( , ),in The overlap time between communication and parallel computing, For the execution duration of the communication task, This refers to the duration of parallel computation. The congestion integral Defined as = or = Where P is the set of edges traversed by the currently scheduled communication path; Piling size according to ← · Updated, the segmentation rule is: if < or > Then f = <1; like ≥ and ≤ Then f = >1; Otherwise, f = 1; , For the lower and upper bounds of the overlapping targets, , For congestion warnings at lower and upper limits, , The preset adjustment coefficient; When in a telemetry downgraded state or airspace downgraded state, it is prohibited to press... The buckets are enlarged, and a conservative adjustment is forced to make f≤1 until the corresponding downgraded state exits.
8. The method according to claim 7, characterized in that, The mutual exclusion time slot reservation mentioned in step S5 only applies to Establish a time slot occupancy table for shared edges with a value of 1. Edges with a value of 0 are not subject to mutual exclusion enforcement. At most one set of tasks is allowed to occupy the congestion edge in the same time slot, i.e., for ∀e: =1 and ∀time slot s satisfy ≤1, where ∈{0,1} indicates whether task t occupies edge e in time slot s, where T is the set of concurrent tasks; When pre-allocation would violate the above mutual exclusion constraints, high-priority tasks would preempt conflicting time slots already occupied by low-priority tasks. The backpressure acts on step S4: notifying the preempted or delayed low-priority tasks to modify the bucket granularity. (according to (Shrink or freeze the amplification), reduce the number of concurrent groups or reduce the frequency of bucketing triggering, so that the multi-task mutual exclusion decision and the single-task bucketing dual feedback form a closed loop; Tasks in a state of airspace degradation or telemetry degradation are prohibited from preemptively occupying cross-domain concurrent bandwidth for upgrades.
9. The method according to any one of claims 1 to 8, characterized in that, The local reduction described in step S2 requires cross-domain structure selection for solution. This was completed earlier so that the cross-domain phase only transmits the reduced representative data from within the domain; The local reduction employs summation or average aggregation; The high-bandwidth domain is a set of nodes identified based on the physical topology hierarchy and the distribution of edge weights within the domain in the communication cost graph. Specifically, it is a set of nodes that are mounted on the same access switch or in the same rack and are high-speed interconnected, with all edge weights within the domain being lower than a preset threshold within the domain. The candidate communication structures include at least two of the following: ring structure, tree structure, and hierarchical structure; candidate set Must pass =1 Hard exclusion from falling into the feasible region Output can be done later ; when When the airspace is empty, it enters a degraded airspace state, adopts a safe ring structure that avoids congestion edges or performs flow limiting, and records and broadcasts the degraded state.
10. A congestion-aware aggregated communication scheduling system for intelligent computing center networks, used to execute the method as described in any one of claims 1 to 9, characterized in that, include: The congestion labeling cost graph module is used to execute step S1; The intra-domain reduction and cross-domain structure selection module is used to execute step S2; The cost graph-driven dynamic grouping module is used to execute step S3; The bucket overlap and dual feedback scheduling module is used to execute step S4; The multi-task peak-shifting and mutual exclusion reservation module is used to execute step S5; The congestion labeling cost graph module outputs a communication cost graph with edge weights and congestion labels, as well as telemetry degradation status, to the other modules. The intra-domain reduction and cross-domain structure selection module outputs... Or, the airspace is in a degraded state, and a degrade flag is broadcast to the bucket and multi-task modules; A joint closed-loop interface for grouping and structure is established between the intra-domain reduction and cross-domain structure selection module and the cost graph-driven dynamic grouping module: the former generates initial grouping constraints from the cost graph and outputs them. The latter uses the refining results as input constraints for the next cycle. The overlap ratio between the bucketing overlap and the dual feedback scheduling module reception Congestion Integral And write back the granularity of the buckets; The multi-task peak shaving and mutual exclusion reservation module maintains the congestion edge time slot occupancy table and outputs the preemption backpressure signal.
Citation Information
Patent Citations
DE260027A
