Intelligent hybrid parallel synchronization method for large-scale distributed simulation

Abstract

The application belongs to the technical field of parallel discrete event simulation and network system simulation, and discloses an intelligent hybrid parallel synchronization method for large-scale distributed simulation, which collects key operation indexes such as neighborhood timestamp lower bound, next event time, round-trip delay quantile value, cross-partition event output rate and virtual time jitter, evaluates system load and communication state in real time, judges the best time to enter optimism or rollback conservation based on the collected key operation indexes, prevents frequent switching by using hysteresis mechanism and minimum residence time, sets time-bounded gates at the mode switching boundary, only allows events within the safe range to propagate, temporarily stores events exceeding the range and sends them in batches when the conditions are met, and thus guarantees causal consistency. The application takes into account real-time performance and robustness, shortens the global empty waiting time, facilitates real-time online acceleration and stable submission in a simulation platform and a message passing interface, and has both performance and interpretability.

Description

Technical Field

[0001] This invention belongs to the field of parallel discrete event simulation and network system simulation, specifically relating to an intelligent hybrid parallel synchronization method for large-scale distributed simulation. Background Technology

[0002] ns-3, as a mainstream open-source simulation platform for network system research and teaching, is widely used for protocol and architecture evaluation in cloud data centers, backbone networks, 5G / 6G, and large-scale model training networks. With the increasing demands for simulation scale and accuracy, single-machine simulation can no longer meet the computational and memory requirements of large-scale topologies and high-traffic concurrency scenarios. Distributed PDES relying on Message Passing Interface (MPI) has become an inevitable choice. Existing parallel discrete event simulation (PDES) synchronization mechanisms are mainly divided into two categories: one is conservative synchronization based on empty messages or time windows, relying on safety boundaries such as lookahead (i.e., the time ahead of which an event can be safely advanced) to avoid causal violations, possessing interpretability and stability. However, when the lookahead is small, the cross-partition event interaction density is high, or the network latency jitter is large, global waiting and throughput degradation are prone to occur. The other is optimistic synchronization based on Time Warp (event rollback mechanism), which supports rollback through state preservation and anti-messages, theoretically achieving higher parallelism. However, when cross-partition events are dense or the network is unstable, it is prone to triggering rollback storms, leading to a significant increase in peak memory usage and output submission latency.

[0003] However, current distributed simulation implementations of ns-3 mostly employ fixed, conservative scheduling (such as neighborhood-based lower bound advancement or time window barriers), and do not yet support optimistic synchronization mechanisms. Therefore, the potential of optimistic synchronization in improving parallel performance cannot be fully utilized. Furthermore, although optimistic synchronization boasts higher theoretical parallelism, in real-world large-scale distributed scenarios, its rollback and resource consumption issues are difficult to control due to limitations in event distribution and network conditions, making it difficult to directly apply in complex simulation environments. Therefore, there is an urgent need to design a hybrid synchronization method that can dynamically switch synchronization strategies based on actual operating conditions to balance system throughput and stability.

[0004] Existing hybrid approaches also face two main challenges: First, the causal protection mechanism at the boundary between optimistic and conservative logical processes (LPs) is insufficient, easily leading to premature submission of cross-boundary events and contaminating the conservative timeline. Second, the lack of measurable and implementable switching criteria makes it difficult to balance throughput and stability under different loads and network conditions. With the introduction of ultra-large-scale topologies, bursty services, and heterogeneous clusters (different machine types / network layers), these problems become more pronounced, urgently requiring a feasible and adaptive optimistic-conservative hybrid synchronization method in ns-3 environments to achieve performance optimization under the premise of causal consistency. Summary of the Invention

[0005] Faced with the continuous growth of simulation scale, time-varying density of cross-partition event interactions, and network jitter caused by heterogeneous clusters, optimizing the parallel performance of ns-3 distributed simulation to the best while ensuring causal correctness has become the core challenge in the field of parallel discrete event simulation (PDES). To solve the above technical problems, this application proposes an intelligent hybrid parallel synchronization method for large-scale distributed simulation, which aims to overcome the problems of lookahead instability, network state fluctuations, and rollback overhead and global null caused by dense cross-logical process (LP) events.

[0006] To achieve the above objectives, this application employs the following technical solution:

[0007] This application discloses an intelligent hybrid parallel synchronization method for large-scale distributed simulation. The intelligent hybrid parallel synchronization method includes a collection communication module, a utility evaluation and judgment module, and a gating and buffering module. Specifically, the intelligent hybrid parallel synchronization method includes the following steps:

[0008] Step 1: Collect observations during distributed simulation runtime online and construct a lightweight data observation set. The observations include: the lower bound of the neighborhood timestamp (LBTS) and the local next event timestamp. High percentile of round-trip latency, MPI send queue depth, retry and retransmission count, and cross-partition event output rate. Local virtual time jitter And statistics on recent rollback density ;

[0009] Step 2: Based on the observations collected in Step 1, the ensemble communication module uses a persistent ensemble communication mechanism to determine the local minimum candidate time of the parallel simulation units LP participating in the synchronization. The reduction process yields the Global Virtual Time (GVT), which is used to characterize the system's safe submission front, thereby forming a stable local minimum candidate time. This provides reliable data support for adaptive decision-making in subsequent synchronization strategies;

[0010] Step 3: Based on the utility evaluation and judgment module, dynamically score each optimistic and conservative parallel simulation unit (LP) from the time dimension, network dimension, and simulation dimension. That is, periodically substitute the observations from Step 1 into the utility evaluation and judgment module to calculate the utility value. Based on this, a mode control signal is generated to switch between optimistic and conservative modes;

[0011] Step 4, Gating and Buffering Mechanism: The gating and buffering module sets a time-bounded gate at the boundary between optimistic and conservative modes, and determines the corresponding gate time (gate_time) for each neighboring parallel simulation unit (LP). Only cross-parallel simulation unit LP events with event timestamps not exceeding the gate time (gate_time) are allowed to be sent out through the message passing layer (such as the MPI communication library) and delivered to the input event queue of the target parallel simulation unit LP. Cross-parallel simulation unit LP events with event timestamps exceeding the gate time (gate_time) are written to the external temporary storage buffer (GateBuffer). When the gate time (gate_time) advances with the global virtual time (GVT) or the lower bound of the neighboring timestamp (LBTS), it is flushed in batches according to neighbors. When switching from optimistic mode to conservative mode, rollback control is performed on the events that have been sent out and whose event timestamps exceed the gate time (gate_time).

[0012] Step 5: When in optimistic mode, enable differential state log overhead. Anti-message overhead The optimistic runtime component registers and confirms cross-partition outgoing events. When the rollback conservative mode is triggered, it freezes the outgoing events of cross-parallel simulation unit LP events where the event timestamp ts exceeds the gate time gate_time. It sends a counter-message to outgoing records in the registration and waits for confirmation. Then it shuts down the optimistic runtime component.

[0013] Step 6: Under the causal consistency boundary jointly defined by the global virtual time (GVT) in Step 2 and the time-bounded gate (gate) and its gate time (gate_time) in Step 4, based on the adaptive switching strategy of the utility evaluation and judgment module described in Step 3, online advancement and state reclamation are executed. A collaborative mechanism between the optimistic segment execution queue (OptExecQueue) and the external temporary storage buffer (GateBuffer) is constructed. The global virtual time (GVT) is periodically calculated through persistent set communication, triggering fossil reclamation and event output submission. For cross-partition external events and reverse message overhead... It adopts batch packaging and derived data type transmission methods to reduce message startup latency and CPU usage, and ensures communication progress through progress threads or periodic polling mechanisms, realizing online switching and stable submission. Thus, under the constraint of causal consistency, it achieves online adaptive switching and stable parallel progress between optimistic and conservative modes.

[0014] A further improvement of this application is that step 1 specifically includes the following steps:

[0015] Step 1.1: Collect time-related metrics: Obtain the lower bound of the neighborhood timestamp (LBTS) and the local next event timestamp. By using the lower bound of the neighbor timestamp and the local next event timestamp Calculate the safety margin;

[0016] Step 1.2: Collect network metrics: online statistical estimation of the high quantile value of round-trip time (RTT), message passing interface (MPI) sending queue depth, and retry and retransmission counts;

[0017] Step 1.3: Collect simulation metrics: Calculate the cross-partition event output rate. Local virtual time jitter And statistics on recent rollback density .

[0018] A further improvement in this application is that step 2 specifically involves: initializing the global virtual time calculation channel: establishing a global virtual time (GVT) reduction channel based on a collective communication mechanism, such as MPI-based persistent Allreduce, with each optimistic and conservative parallel simulation unit (LP) periodically reporting its local minimum candidate time. The local minimum candidate time It is obtained by taking the minimum value among local virtual time, ready queue head time, minimum receive buffer time, and minimum time for unacknowledged outgoing events. =min{local virtual time, ready queue head time, minimum receive buffer time, minimum unacknowledged outgoing time}, to maintain the minimum in-transit event timestamp ts, each parallel simulation unit (LP) maintains an outgoing queue (Outbox) for unacknowledged outgoing events. When sending an event across parallel simulation units (LPs), a record containing the event identifier (eventId) and event timestamp ts is registered in the Outbox. After receiving confirmation from the other end, the registration record corresponding to the event identifier (eventId) is deleted. This is achieved by optimizing the local minimum candidate time of each parallel simulation unit (LP). Minimum reduction is performed to obtain the global virtual time GVT, where the global virtual time GVT is the local minimum candidate time reported by each parallel simulation unit (LP). The global minimum value, i.e. .

[0019] A further improvement in this application is that, in step 3, from a time dimension, the lower bound of the neighboring timestamp and the local next event timestamp are used. Calculate safety margin From a network perspective, network jitter and congestion are measured using the high quantile of the estimated round-trip time (RTT). Combined with the message passing interface (MPI) send queue depth, retry and retransmission counts, a comprehensive congestion indicator is obtained. From a simulation perspective: based on cross-partition event output rate As an estimate of interaction strength, local virtual time jitter is used. As a local virtual time LVT jitter, the recent rollback density is statistically analyzed. The activity intensity is used as the input to the risk function and the switching criterion. The specific steps in step 3 include the following:

[0020] Step 3.1: Construct a monotonic risk function : Regarding safety margin Decreasing, relative Incremental, using logical functions or piecewise linear approximations, with a safety margin in the function. Round-trip delay High quantile values, cross-partition event output rate and local virtual time jitter The weighting coefficients are updated online adaptively using exponential smoothing based on recent observations.

[0021] Step 3.2: Evaluate the workload gain that can be released by early execution. Estimate rollback costs ,in, This is due to the overhead of differential state logging. For anti-message overhead, Upper bounds are set for replay time and memory overhead, and based on recent rollback density. The water levels in the optimistic section execution queue and the external temporary buffer queue, i.e., the current occupancy level of the queues, determine the rollback cost adjustment coefficient, and adjust the rollback cost accordingly. Make the appropriate enlargement or reduction;

[0022] Step 3.3, Utility Function and Hysteresis Decision: Based on Workload Gain Calculate expected utility = Using an upper threshold with hysteresis Lower threshold Minimum Dwell Time Decision: An upper threshold θ↑ and a lower threshold θ↓ are set for the expected utility value U of each parallel simulation unit LP. When the expected utility value U of a parallel simulation unit LP is higher than the upper threshold θ↑ for K consecutive cycles, a mode control signal to switch to optimistic mode is sent to the parallel simulation unit LP, and the running mode flag is set to optimistic mode. When the expected utility value U of the parallel simulation unit LP is lower than the lower threshold θ↓ for K consecutive cycles, a mode control signal to switch to conservative mode is sent to the parallel simulation unit LP, and the running mode flag is set to conservative mode. The default startup mode is conservative mode, and θ↑ > θ↓ to prevent mode jitter.

[0023] A further improvement of this application is that step 4 specifically includes the following steps:

[0024] Step 4.1: The gating and buffering module sets a time-bounded gate at the boundary between optimistic and conservative modes. Each target parallel simulation unit LP is denoted as target LP and identified as dst, with the gate time... Defined as:

[0025] }

[0026] in, This indicates the lower bound of the acceptable timeframe for the other party's external declaration. This is a safety margin compensation item;

[0027] Step 4.2: Only release events whose timestamp ts does not exceed the gate time. Cross-partition events and out-of-bounds events enter the external temporary buffer GateBuffer to wait. When the global virtual time (GVT) or the lower bound of a neighbor is raised, events in the external temporary buffer GateBuffer that satisfy the condition that the event timestamp ts ≤ gate time are considered. The event is flushed in batches by neighbor.

[0028] A further improvement in this application is that, in step 4, the GateBuffer is managed by bucketing according to the neighbor dimension: when the Global Virtual Time (GVT) or the lower bound of the neighbor's timestamp (LBTS) is raised, causing the gate to move forward, the temporary events that meet the gate conditions are flushed in batches by bucket, and conventional non-contiguous memory batch packaging methods such as MPI derived data types or POSIX iovec, or equivalent implementations, are sent at once.

[0029] A further improvement of this application is that step 6 specifically includes the following steps:

[0030] Step 6.1: Construct the optimistic segment execution queue: Under the control of the obtained global virtual time GVT and the gate time determined in step 4, establish the optimistic segment execution queue OptExecQueue, the external temporary storage buffer queue GateBuffer, the neighbor-based bucketing, and the pending confirmation outbound queue Outbox. Driven by the online observation set obtained in step 1 and the mode control signal output in step 3, organize and adjust the optimistic segment execution queue OptExecQueue, the external temporary storage buffer queue GateBuffer, and the pending confirmation outbound queue Outbox to provide a foundation for subsequent gating and batch transmission.

[0031] Step 6.2: Periodically calculate the Global Virtual Time (GVT) via persistent set communication: Based on the optimistic segment execution queue constructed in Step 6.1, each A persistent global minimum reduction operation based on the message passing interface MPI is periodically triggered, using the local minimum candidate time periodically reported by each parallel simulation unit (LP). Obtain the global virtual time (GVT). , The trigger period for calculating the Global Virtual Time (GVT);

[0032] Step 6.3, Fossil Recycling and Submission: Based on the latest Global Virtual Time (GVT) obtained in Step 6.2, when the event timestamp ts is less than the GVT, the corresponding differential state log overhead is... Message records are deleted or marked as reusable to safely reclaim and release occupied storage space, and externally visible output is committed less than the Global Virtual Time (GVT).

[0033] Step 6.4, Batch Packaging and Derivation of Data Types: This step addresses the cross-partition outgoing events and reverse message overhead in the optimistic segment execution queue OptExecQueue, the external temporary buffer queue GateBuffer, and the pending confirmation outgoing queue Outbox described in Step 6.1. It adopts a batch packaging method, performs bucket aggregation based on the neighbor dimension, and uses the derived data type of the message passing interface MPI for one-time transmission;

[0034] Step 6.5, Communication Progress and Schedule Assurance: To ensure the progress of communication during Step 6.2 Global Virtual Time (GVT), Step 6.3 fossil recycling, and Step 6.4 batch packaging and transmission, an independent progress thread or periodic... Polling is performed to ensure that communication can still proceed in computationally intensive sections; high water level thresholds are set for the GateBuffer and the Outbox queue to be confirmed, and if they are exceeded, priority flushing or temporary degradation is triggered.

[0035] Step 6.6, Parameter Adaptation: During the above operation, based on the high quantile value of the estimated round-trip time (RTT)... Safety margin compensation item Adaptive adjustments are made; the upper threshold θ↑ and lower threshold θ↓ are dynamically updated based on the mean and variance of historical expected utility U; and the recent rollback density is considered. The minimum residence time is adjusted in relation to the water level of the Gate Buffer outside the gate, so as to improve the flexibility of mode switching while ensuring stability.

[0036] Step 6.7, Hierarchical Gate Management in Multi-Layer Networks: In a multi-layer network scenario, based on the gating buffer mechanism and the Global Virtual Time (GVT) and message processing results obtained in steps 6.2-6.4, gates are set hierarchically, i.e., gate_time_intra ≥ gate_time_inter. Here, gate_time_intra refers to the gate time calculated for neighbors within the same layer and near-end neighbors, and gate_time_inter refers to the gate time calculated for neighbors across layers and far-end neighbors. Through the hierarchical gate time and the global virtual time (GVT) working together, relevant events are only made visible and submitted when the event timestamp ts is less than or equal to the corresponding gate and the event timestamp ts is less than or equal to the global virtual time (GVT). This achieves globally causally consistent synchronous parallel progress in a multi-layer network. A further improvement of this application is that in step 6.5, communication progress and schedule assurance include independent progress threads or periodic... Polling is used to configure high water level thresholds for the external buffer and external display. If the high water level threshold is exceeded, priority flushing or temporary degradation is triggered.

[0037] Further improvements in this application are as follows: in conservative mode, neighborhood lower bound advancement and empty messages are used to maintain causal order; in optimistic mode, differential state logs and reverse messages are used to support rollback; and during the switching between the two modes, the conservative side does not see undetermined events.

[0038] The beneficial effects of this application are:

[0039] This application establishes an observation system based on a small number of indicators (S, p90 RTT, λ_out, J, GVT), taking into account both real-time performance and robustness, to provide accurate decision input for hybrid propulsion.

[0040] This application only activates the optimistic mode when the expected return is positive, resulting in overall throughput superior to a single strategy. Based on the utility-risk criterion, it activates the optimistic mode only when the expected return is positive, and uses a time-bounded gate to make the causal boundary explicit. In the weakly coupled region, it approaches the acceleration upper bound of pure optimism, and automatically de-conservatives in the strongly coupled / jittering region, significantly reducing rollback density and memory peak, and shortening global wait time.

[0041] The persistent collection communication, event batching, and derived types of this application result in low operating overhead and minimal modifications; the switching basis is traceable (U and threshold trajectory) and the gate boundary is auditable, which facilitates real-time online acceleration and stable submission in simulation platforms and message passing interfaces (MPI), combining performance and interpretability. Attached Figure Description

[0042] Figure 1 This is a framework diagram of the intelligent hybrid parallel synchronization method of this application.

[0043] Figure 2 This is a schematic diagram illustrating the changing trends of parallel speedup ratio under different synchronization mechanisms and different coupling strengths.

[0044] Figure 3 This is a diagram comparing the P95 commit latency of different synchronization mechanisms in weak, medium, and strong coupling scenarios.

[0045] Figure 4 This is a schematic diagram comparing the rollback density of different synchronization mechanisms in weak, medium, and strong coupling scenarios.

[0046] Figure 5 This is a flowchart illustrating the process of this application. Detailed Implementation

[0047] The embodiments of the present invention will be disclosed below with reference to the drawings. For clarity, many practical details will be described in the following description. However, it should be understood that these practical details are not intended to limit the invention. That is, in some embodiments of the invention, these practical details are not essential.

[0048] like Figure 1 and Figure 5 As shown, this application presents an intelligent hybrid parallel synchronization method for large-scale distributed simulation. The intelligent hybrid parallel synchronization method includes a collection communication module, a utility evaluation and judgment module, and a gating and buffering module. Specifically, the intelligent hybrid parallel synchronization method includes the following steps:

[0049] Step 1: Collect observations during distributed simulation runtime online and construct a lightweight data observation set, specifically including the following steps:

[0050] Step 1.1: Collect time-related metrics: Obtain the lower bound of the neighborhood timestamp (LBTS) and the local next event timestamp. By using the lower bound of the neighbor timestamp and the local next event timestamp Calculate the safety margin; the safety margin is defined as the difference between the lower bound of the neighborhood timestamp (LBTS) and the local next event time, used as the core time metric for entering optimistic mode. The LBTS is maintained by a neighbor null message announcement or a conservative thruster. As the time dimension input for the risk function and utility function in step 3;

[0051] Step 1.2: Collect network metrics: online statistical estimation of the high quantile value of round-trip time (RTT), message passing interface (MPI) sending queue depth, and retry and retransmission counts; in this embodiment, the high quantile value of the RTT is selected as the 90th percentile, and the online estimation adopts the sliding window quantile or streaming quantile algorithm to achieve an approximate constant time update.

[0052] Step 1.3: Collect simulation metrics: Calculate the cross-partition event output rate. Local virtual time jitter And statistics on recent rollback density The aforementioned statistics on recent rollback density Minimum dwell time / threshold adaptation for step 3.6 and parameter adaptation for step 6.6.

[0053] Step 2: Based on the observations collected in Step 1, the ensemble communication module uses a persistent ensemble communication mechanism to determine the local minimum candidate time of the parallel simulation units LP participating in the synchronization. The reduction process yields the Global Virtual Time (GVT), which is used to characterize the system's safe submission front, thereby forming a stable local minimum candidate time. This provides reliable data support for adaptive decision-making in subsequent synchronization strategies;

[0054] Step 2 specifically involves: Initializing the global virtual time calculation channel: Establishing a global virtual time (GVT) reduction channel based on a collective communication mechanism such as MPI-based persistent Allreduce, with each optimistic and conservative parallel simulation unit (LP) periodically reporting its local minimum candidate time. The local minimum candidate time It is obtained by taking the minimum value among local virtual time, ready queue head time, minimum receive buffer time, and minimum time for unacknowledged outgoing events. =min{local virtual time, ready queue head time, minimum receive buffer time, minimum unacknowledged outgoing time}, to maintain the minimum in-transit event timestamp ts, each parallel simulation unit (LP) maintains an outgoing queue (Outbox) for unacknowledged outgoing events. When sending an event across parallel simulation units (LPs), a record containing the event identifier (eventId) and event timestamp ts is registered in the Outbox. After receiving confirmation from the other end, the registration record corresponding to the event identifier (eventId) is deleted. This is achieved by optimizing the local minimum candidate time of each parallel simulation unit (LP). Minimum reduction is performed to obtain the global virtual time GVT, where the global virtual time GVT is the local minimum candidate time reported by each parallel simulation unit (LP). The global minimum value, i.e. Step 3: Based on the utility evaluation and judgment module, dynamically score each optimistic and conservative parallel simulation unit (LP) from the time dimension, network dimension, and simulation dimension. That is, periodically substitute the observations from Step 1 into the preset utility function to calculate the utility value. Based on this, a mode control signal is generated to switch between optimistic and conservative modes; the utility evaluation function is used to comprehensively characterize the workload gain that can be released by early execution and the potential rollback risks and costs, to obtain the expected utility value U.

[0055] In step 3, the time dimension is based on a safety margin. As the core metric, the spatial dimension considers the event dependencies between neighborhood parallel simulation units (LPs), while the simulation dimension combines network and distributed simulation runtime observations to comprehensively evaluate the operational status of the distributed simulation environment. From a temporal perspective, this is achieved through the lower bound of the neighborhood timestamp and the local next event timestamp. Calculate safety margin The larger the security margin, the less likely it is to be affected by a remote event timestamp being earlier than the local next event timestamp. The cross-parallel simulation unit LP input event breakdown; from the network dimension, the high quantile value of the estimated round-trip time (RTT) is used to measure network jitter and congestion, and combined with the message passing interface (MPI) send queue depth, retry and retransmission counts, a comprehensive congestion indicator is obtained. From a simulation perspective: based on cross-partition event output rate As an estimate of interaction strength, local virtual time jitter is used. As a local virtual time LVT jitter, the recent rollback density is statistically analyzed. The activity intensity reflects a comprehensive measure of the simulation load activity and the degree of cross-regional interaction. As an input to the risk function and switching criterion, specific step 3 includes the following steps:

[0056] Step 3.1: Construct a monotonic risk function : Regarding safety margin Decreasing, relative Incremental, using logical functions or piecewise linear approximations, with a safety margin in the function. Round-trip delay High quantile values, cross-partition event output rate and local virtual time jitter The weighting coefficients are updated online adaptively using exponential smoothing based on recent observations; the monotonic risk function... It is implemented using a logistic function or piecewise linear approximation, and the parameters are adaptively updated through exponential smoothing based on online observations;

[0057] Step 3.2: Evaluate the workload gain that can be released by early execution. The workload gain ΔW is estimated based on the safety margin S and the event processing rate of the parallel simulation unit LP. Calculate, where, Estimate the rollback cost based on the average number of events processed by the parallel simulation unit LP per unit simulation time. ,in, This is due to the overhead of differential state logging. For anti-message overhead, Upper bounds are set for replay time and memory overhead, and based on recent rollback density. The water levels in the optimistic section execution queue and the external temporary buffer queue, i.e., the current occupancy level of the queues, determine the rollback cost adjustment coefficient, and adjust the rollback cost accordingly. Make the appropriate enlargement or reduction;

[0058] Step 3.3, Utility Function and Hysteresis Decision: Based on Workload Gain Calculate expected utility = Using an upper threshold with hysteresis Lower threshold Minimum Dwell Time Decision: An upper threshold θ↑ and a lower threshold θ↓ are set for the expected utility value U of each parallel simulation unit LP. When the expected utility value U of a parallel simulation unit LP is higher than the upper threshold θ↑ for K consecutive cycles, a mode control signal to switch to optimistic mode is sent to that parallel simulation unit LP, and the running mode flag is set to optimistic mode. When the expected utility value U of the parallel simulation unit LP is lower than the lower threshold θ↓ for K consecutive cycles, a mode control signal to switch to conservative mode is sent to that parallel simulation unit LP, and the running mode flag is set to conservative mode. The default startup is conservative mode, and θ↑ > θ↓ to prevent mode jitter. In this embodiment, the lower threshold is used. = 0 (U≤0 means "not worthwhile"); take =Expected utility value over a recent period The average value and a small margin such as +10%; K is 3–5.

[0059] Step 4, Gating and Buffering Mechanism: The gating and buffering module sets a time-bounded gate (gate) at the boundary between optimistic and conservative modes, and determines a corresponding gate time (gate_time) for each neighboring parallel simulation unit (LP). Only cross-parallel simulation unit LP events with event timestamps (ts) not exceeding the gate time (gate_time) are allowed to be sent out via a message passing layer such as the MPI communication library and delivered to the input event queue of the target parallel simulation unit LP. Cross-parallel simulation unit LP events with event timestamps exceeding the gate time (gate_time) are written to the external temporary storage buffer (GateBuffer). As the gate time (gate_time) advances with the global virtual time (GVT) or the lower bound of the neighboring timestamp (LBTS), it is flushed in batches according to neighbors. When switching from optimistic mode to conservative mode, rollback control is performed on events that have been sent out and whose event timestamps exceed the gate time (gate_time). Specifically, this includes the following steps:

[0060] Step 4.1: The gating and buffering module sets a time bounded gate at the boundary between optimistic and conservative modes, and applies this gate to each target parallel simulation unit LP. Let LP be the target, dst be the identifier, and gate time be specified. Defined as:

[0061] }

[0062] in, This indicates the acceptable lower bound of the time declared by the peer, when the target parallel simulation unit LP is in conservative mode. The lower bound of the time for the target parallel simulation unit LP is obtained based on a conservative synchronization algorithm; when the target parallel simulation unit LP is in optimistic mode... Take the lower bound of the submission or acceptance time of the target parallel simulation unit LP, for example, a lower bound determined based on the global virtual time GVT or the internal submission water level. This is a safety margin compensation item;

[0063] Step 4.2: Only release events whose timestamp ts does not exceed the gate time. Cross-partition events and out-of-bounds events enter the external temporary buffer GateBuffer to wait. When the global virtual time (GVT) or the lower bound of a neighbor is raised, events in the external temporary buffer GateBuffer that satisfy the condition that the event timestamp ts ≤ gate time are considered. The event is flushed in batches by neighbor.

[0064] Gate time according to (neighbor) = max(GVT, LBTS_neighbor) Confirmed; Safety Compensation Item The result is obtained adaptively based on the minimum link latency or by multiplying the round-trip latency by a scaling factor based on the highest quantile.

[0065] In this step, the GateBuffer is managed by bucketing according to the neighbor dimension: when the Global Virtual Time (GVT) or the lower bound of the neighbor's timestamp (LBTS) is raised, causing the gate to move forward, the temporary events that meet the gate conditions are flushed in batches by bucket, and sent all at once using conventional non-contiguous memory batch packaging methods (such as MPI derived data types or POSIX iovec, or equivalent implementations).

[0066] Step 5: When in optimistic mode, enable differential state log overhead. Anti-message overhead The optimistic runtime component registers and confirms cross-partition outgoing events. When a conservative rollback mode is triggered, it freezes outgoing events across parallel simulation units (LPs) where the event timestamp ts exceeds the gate time (gate_time). It sends a counter-message to outgoing records in the registration and waits for confirmation, then shuts down the optimistic runtime component. Cross-partition outgoing events register event identifiers and timestamps upon transmission, and delete them after confirmation by the receiving end. When downgrading to conservative mode, it sends a counter-message to records whose timestamps exceed the gate time and waits for confirmation to eliminate the impact of outgoing events. In step 5, a rollback budget and commit priority are set: when the rollback frequency or span exceeds the budget, optimistic advancement is restricted and downgraded to conservative window advancement; the commit priority prioritizes safety events smaller than GVT.

[0067] Step 6: Under the causal consistency boundary jointly defined by the global virtual time (GVT) in Step 2 and the time-bounded gate (gate) and its gate time (gate_time) in Step 4, based on the adaptive switching strategy of the utility evaluation and judgment module described in Step 3, online advancement and state reclamation are executed. A collaborative mechanism between the optimistic segment execution queue (OptExecQueue) and the external temporary storage buffer (GateBuffer) is constructed. The global virtual time (GVT) is periodically calculated through persistent set communication, triggering fossil reclamation and event output submission. For cross-partition external events and reverse message overhead... This approach employs batch packaging and derived data type transmission to reduce message startup latency and CPU usage. It also ensures communication progress through a progress thread or periodic polling mechanism, enabling online switching and stable commits. This allows for online adaptive switching and stable parallel progress between optimistic and conservative modes under causal consistency constraints. Specifically, it includes the following steps:

[0068] Step 6.1: Construct the optimistic segment execution queue: Under the control of the obtained global virtual time GVT and the gate time determined in step 4, establish the optimistic segment execution queue OptExecQueue, the external temporary storage buffer queue GateBuffer, the neighbor-based bucketing, and the pending confirmation outbound queue Outbox. Driven by the online observation set obtained in step 1 and the mode control signal output in step 3, organize and adjust the optimistic segment execution queue OptExecQueue, the external temporary storage buffer queue GateBuffer, and the pending confirmation outbound queue Outbox to provide a foundation for subsequent gating and batch transmission.

[0069] Step 6.2: Periodically calculate the Global Virtual Time (GVT) via persistent set communication: Based on the optimistic segment execution queue constructed in Step 6.1, each A persistent global minimum reduction operation based on the message passing interface MPI is periodically triggered, using the local minimum candidate time periodically reported by each parallel simulation unit (LP). Obtain the global virtual time (GVT). , The trigger period for global virtual time (GVT) calculation is used to control the reduction frequency and can be taken as the wall clock time interval or the event batch number threshold. The decision period of the utility criterion can be set differently from the GVT reduction period to improve switching sensitivity and maintain recovery stability in computationally intensive sections.

[0070] Step 6.3, Fossil Recycling and Submission: Based on the latest Global Virtual Time (GVT) obtained in Step 6.2, when the event timestamp ts is less than the GVT, the corresponding differential state log overhead is... Message records are deleted or marked as reusable to safely reclaim and release occupied storage space, and externally visible outputs less than the Global Virtual Time (GVT) are committed to ensure reproducible results;

[0071] Step 6.4, Batch Packaging and Derivation of Data Types: This step addresses the cross-partition outgoing events and reverse message overhead in the optimistic segment execution queue OptExecQueue, the external temporary buffer queue GateBuffer, and the pending confirmation outgoing queue Outbox described in Step 6.1. It adopts a batch packaging method, performs bucket aggregation based on the neighbor dimension, and uses the derived data type of the message passing interface MPI for one-time transmission;

[0072] Step 6.5, Communication Progress and Schedule Assurance: To ensure the progress of communication during Step 6.2 Global Virtual Time (GVT), Step 6.3 fossil recycling, and Step 6.4 batch packaging and transmission, an independent progress thread or periodic... Polling is performed to ensure that communication can still proceed in computationally intensive sections; high water level thresholds are set for the GateBuffer and Outbox queues to be acknowledged, and if they are exceeded, priority flushing or temporary degradation is triggered; the publication is organized using a lock-free ring or fragmented hash table; the event identifier is composed of the sender identifier and a monotonic count to facilitate the rapid location of reverse messages and acknowledgments.

[0073] Step 6.6, Parameter Adaptation: During the above operation, based on the high quantile value of the estimated round-trip time (RTT)... Safety margin compensation item Adaptive adjustments are made; the upper threshold θ↑ and lower threshold θ↓ are dynamically updated based on the mean and variance of historical expected utility U; and the recent rollback density is considered. The minimum residence time is adjusted in relation to the water level of the Gate Buffer outside the gate, so as to improve the flexibility of mode switching while ensuring stability.

[0074] Step 6.7, Hierarchical Gate Management in Multi-Layer Networks: In a multi-layer network scenario, based on the gated buffer mechanism described in Step 4 and the Global Virtual Time (GVT) and message processing results obtained in Steps 6.2 to 6.4, gate times are set according to network layers, i.e., gate_time_intra ≥ gate_time_inter, where gate_time_intra refers to the gate time calculated for neighbors in the same layer and near-end neighbors, and gate_time_inter refers to the gate time calculated for neighbors across layers and far-end neighbors. Through the hierarchical gate time and the global virtual time (GVT) working together, relevant events are only made visible and submitted when the event timestamp ts is less than or equal to the corresponding gate time and the event timestamp ts is less than or equal to the global virtual time (GVT), thereby achieving globally causal consistent synchronous parallel progress in a multi-layer network.

[0075] In a multi-layer network, gate times are set hierarchically according to communication distance, allowing intra-layer / near-end neighbors and cross-layer / far-end neighbors to use different gate times. This enables fine-grained control over when events become visible to the outside world when working in conjunction with GVT, gating buffers, and commit policies. While ensuring global causal consistency, this avoids excessively constraining near-end communication with a single gate, thereby improving overall parallelism and efficiency.

[0076] In conservative mode, neighborhood lower bound advancement and empty messages are used to maintain causal order; in optimistic mode, differential state logs and reverse messages are used to support rollback; during the switching between the two modes, the conservative side is not allowed to see undetermined events.

[0077] To evaluate the acceleration and stability of this application in ns-3 distributed simulation, representative network topologies and service loads were selected and compared with conservative synchronization (neighborhood lower bound / empty message), barrier time window, and pure optimistic time warp under the same partition mapping and traffic model. All results were verified by the causal order of serial ns-3 and key intervals were verified by deterministic replay to ensure repeatability.

[0078] The experiments were conducted on an x86_64 server (32-core CPU, 128GB memory, Linux / OpenMPI 4.x), and the code was implemented in C++ based on ns-3.37. The intelligent hybrid parallel synchronization method GateHybrid proposed in this application uses conservative default parameters: The threshold is set to 0.7 times the 90th percentile of RTT. GVT is persisted with a 500μs cycle, and the switching hysteresis window K=3. The threshold is adaptively estimated from historical utility. Evaluation metrics focus on relative speedup, commit tail latency (P95), rollback density and span, peak memory usage, and MPI startup overhead.

[0079] The results show that in scenarios with latency jitter, in the weak coupling region (≈0.05–0.20), the proposed intelligent hybrid parallel synchronization method, GateHybrid, falls between conservative and purely optimistic performance. In the medium coupling region (≈0.40–0.60), GateHybrid achieves the best or tied-best relative serial speedup at most sampling points, demonstrating a relative advantage. Figure 2 As shown; the submission delay for P95 is reduced by approximately 6%–35% compared to a purely optimistic approach, such as… Figure 3 As shown, in sections with strong coupling and sudden congestion, the method automatically maintains or reverts to a conservative mode, with throughput approaching the conservative baseline, while avoiding the rollback storms that are prone to occur with purely optimistic approaches. In terms of resources and stability, the time-bounded gate effectively intercepts uncertain events, reducing rollback density by approximately 75%–91% compared to purely optimistic approaches. Figure 4 As shown, the combined peak memory usage of the state differential log and messages in transit decreased by approximately 25%-35%, and no instances of the conservative side receiving uncertain events were observed. Regarding communication, batch flushing and derived type packaging reduced the number of message initiations, resulting in a decrease of approximately 15% in the number of MPI messages per unit time compared to a purely optimistic approach. The control flow overhead from gating, outboxes, and their explicit acknowledgment messages (ACKs) was manageable, with additional CPU usage within 3%.

[0080] As can be seen from the above results, under the premise of ensuring causal consistency, this application can approach the parallelism of pure optimism in suitable sections and maintain a conservative deterministic boundary in unfavorable sections. The overall performance is more stable than a single conservative or pure optimism strategy. Moreover, the switching trajectory and gate advancement have good interpretability and reproducibility, meeting the comprehensive requirements of engineering applications for throughput, tail delay and resource upper bound.

[0081] This application possesses the characteristics of high efficiency, stability, and interpretability, making it suitable for large-scale ns-3 distributed simulation scenarios.

[0082] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. An intelligent hybrid parallel synchronization method for large-scale distributed simulation, characterized in that: The intelligent hybrid parallel synchronization method is realized by an intelligent hybrid parallel synchronization device, which comprises a collection communication module, an utility evaluation and judgment module, and a gating and buffering module. Step 1, collecting distributed simulation runtime observations online and constructing a light data observation set, the observations including: neighborhood lower bound of time stamp LBTS, local event time stamp , high quantile value of round trip latency, message passing interface MPI send queue depth, retry and retransmission count, cross partition event output rate , local virtual time jitter and statistical recent rollback density ; Step 2, based on the observation collected in Step 1, the set communication module obtains the local minimum candidate time of the parallel simulation unit LP participating in synchronization reduction, and obtains the global virtual time GVT; Step 3, based on the utility evaluation and decision module, dynamically scoring each optimistic and conservative parallel simulation unit LP from the time dimension, network dimension and simulation dimension, i.e. periodically substituting the observation quantity of step 1 into the utility evaluation and decision module to calculate the utility value And accordingly generating a mode control signal to realize the switching between the optimistic mode and the conservative mode; Step 4, gating and buffering mechanism: the gating and buffering module sets a time-bounded gate at the mode boundary between the optimistic mode and the conservative mode, and determines the gate time gate_time for each neighbor parallel simulation unit LP, and only allows the cross-parallel simulation unit LP event with an event timestamp ts not exceeding the gate time gate_time to be sent out and delivered to the input event queue of the target parallel simulation unit LP, writes the cross-parallel simulation unit LP event with an event timestamp ts exceeding the gate time gate_time into the gate-out temporary buffer GateBuffer, and flushes the events in batches according to the neighbors when the gate time gate_time is advanced with the global virtual time GVT or the neighbor timestamp lower bound LBTS, and performs rollback control on the events with an event timestamp ts exceeding the gate time gate_time which have been sent out when the optimistic mode is switched to the conservative mode; Step 5, when in the optimistic mode, start the differential state log overhead , the anti-message overhead of the optimistic runtime component, and register and confirm the cross-partition outgoing events, freeze the new cross parallel simulation unit LP event outgoing when the event time stamp ts exceeds the gate time gate_time, send the anti-message to the out-of-bound record in the registration and wait for the confirmation, and then close the optimistic runtime component; Step 6, under the causally consistent boundary defined by the global virtual time GVT of step 2 and the time-bounded gate of step 4 and its gate time gate_time, based on the adaptive switching of the utility evaluation and decision module described in step 3, perform online advancement and state recovery, construct a cooperative mechanism of the optimistic section execution queue OptExecQueue and the gate buffer GateBuffer, periodically calculate the global virtual time GVT, trigger fossil recovery and event output submission, and for cross-partition outgoing events and anti-message overhead , adopt batch packaging and derived data type transmission method to reduce message start-up delay and CPU occupancy, and guarantee communication progressiveness, realize online switching and stable submission, and realize online adaptive switching and stable parallel advancement of optimistic mode and conservative mode under the constraint of causal consistency. The step 2 is specifically: initializing a global virtual time calculation channel: establishing a global virtual time GVT reduction channel based on a collective communication mechanism, and periodically reporting a local minimum candidate time by each optimistic and conservative parallel simulation unit LP , wherein the local minimum candidate time is obtained by taking the minimum value among a local virtual time, a head-of-line queue time, a minimum time of a receiving buffer, and a minimum time of unconfirmed outgoing events, that is, =min{local virtual time, head-of-line queue time, minimum time of receiving buffer, minimum time of unconfirmed outgoing events}, for maintaining an in-transit minimum event timestamp ts, each parallel simulation unit LP maintains an outgoing table to-be-confirmed outgoing queue Outbox, registers a record containing an event identifier eventId and an event timestamp ts in the Outbox when sending an event across the parallel simulation unit LP, deletes the registered record corresponding to the event identifier eventId after receiving a confirmation from a peer end, and performs minimum value reduction on the local minimum candidate time of each parallel simulation unit LP to obtain the global virtual time GVT, wherein the global virtual time GVT is a global minimum value of the local minimum candidate time reported by each parallel simulation unit LP, that is, .

2. The intelligent hybrid parallel synchronous approach for large scale distributed simulation according to claim 1, wherein: The step 1 specifically comprises the following steps: Step 1.1, collecting time class index: obtaining a neighborhood timestamp lower bound LBTS and a local synchronization event timestamp , calculating a safety margin by the neighborhood timestamp lower bound and the local synchronization event timestamp ; Step 1.2, collecting network indicators: online statistical estimation of high quantile values of round-trip time RTT, message passing interface MPI sending queue depth, retry and retransmission count; Step 1.

3. Collecting simulation class metrics: count cross-partition event output rate Local virtual time jitter And count recent rollback density .

3. The intelligent hybrid parallel synchronous approach for large scale distributed simulation according to claim 1, wherein: In step 3, from the time dimension, the lower bound of the neighborhood timestamp and the local timestamp of the event are used to estimate the time difference between the two Calculate the safety margin From the network dimension, measure the network jitter and congestion by estimating the high quantile value of the round-trip time RTT, combined with the message passing interface MPI sending queue depth, retry and retransmission count, to obtain the comprehensive congestion indicator From the simulation dimension: the cross-partition event output rate As an estimate of the interaction intensity, the local virtual time jitter As the local virtual time LVT jitter, the recent rollback density is counted The activity intensity is used as the input of the risk function and the switching criterion, and the specific step 3 includes the following steps: Step 3.1: Construct a monotonic risk function : Regarding safety margin Decreasing, relative Incremental, using logical functions or piecewise linear approximations, with a safety margin in the function. Round-trip delay High quantile values, cross-partition event output rate and local virtual time jitter The weighting coefficients are updated online adaptively using exponential smoothing based on the observation results; Step 3.2, evaluating the workload gain that can be released by early execution , estimating rollback cost wherein, is the differential state log overhead, is the anti-message overhead, is the time and memory overhead upper bound of replay, and is determined according to the recent rollback density and the water level of the optimistic section execution queue and the gate-off staging buffer queue to determine the rollback cost adjustment coefficient, which amplifies or reduces the rollback cost ; Step 3.3, Utility function and hysteresis decision: based on workload gain Calculate the expected utility value = , using upper threshold with hysteresis , lower threshold and minimum dwell time decision: set an upper threshold θ↑ and a lower threshold θ↓ for the expected utility value U of each parallel simulation unit LP, when the expected utility value U of a parallel simulation unit LP is higher than the upper threshold θ↑ for K consecutive periods, send a mode control signal to the parallel simulation unit LP to switch to the optimistic mode, and set the running mode flag to the optimistic mode; when the expected utility value U of the parallel simulation unit LP is lower than the lower threshold θ↓ for K consecutive periods, send a mode control signal to the parallel simulation unit LP to switch to the conservative mode, and set the running mode flag to the conservative mode. The default is to start in the conservative mode, and θ↑ > θ↓ to prevent mode chattering.

4. The intelligent hybrid parallel synchronous approach for large scale distributed simulation according to claim 1, wherein: The step 4 specifically comprises the following steps: Step 4.

1. The gating and buffering module sets a time-bounded gate at the boundary of the optimistic and conservative modes, for each target parallel simulation unit LP, gate time is defined as: } wherein, represents the lower bound of the acceptable time to the peer, is a safety margin compensation term; Step 4.2, only release event timestamp ts no more than gate time cross-partition event, super-boundary event enters gate-out staging buffer GateBuffer, when global virtual time GVT or neighbor lower bound is promoted, flush events in gate-out staging buffer GateBuffer that satisfy event timestamp ts ≤ gate time by neighbor batch.

5. The intelligent hybrid parallel synchronous approach for large scale distributed simulation according to claim 1, wherein: In step 4, the gate-out temporary buffer GateBuffer is managed in a bucket according to the neighbor dimension: when the gate is moved forward due to the global virtual time GVT or the neighbor timestamp lower bound LBTS, the temporarily stored events satisfying the gate condition are flushed in batches according to the bucket, and a batch packaging means of non-continuous memory is used for one-time sending.

6. The intelligent hybrid parallel synchronous approach for large scale distributed simulation according to claim 1, wherein: The step 6 specifically comprises the following steps: Step 6.1, constructing an optimistic section execution queue: under the control of the obtained global virtual time GVT and gate time gate_time, an optimistic section execution queue OptExecQueue, a gate-out temporary buffer queue GateBuffer, and a to-be-confirmed outbox queue Outbox are established; Step 6.2, periodically compute global virtual time GVT by communicating through the persistent collection: on the basis of the optimistic section execution queue built in step 6.1, each a message passing interface (MPI) based persistent global minimum reduction operation is triggered periodically to report the local minimum candidate time periodically reported by each parallel simulation unit (LP) get the global virtual time GVT, , trigger period for computing the global virtual time GVT; Step 6.3, fossil recovery and commit: based on the latest global virtual time GVT obtained in step 6.2, when the event timestamp ts is less than the global virtual time GVT, the corresponding differential state log overhead is deleted or marked as reusable, and the occupied storage space is safely recycled and released, and the externally visible output less than the global virtual time GVT is committed; Step 6.

4. Batch packing and derived data type: cross-partition outgoing event and anti-message overhead in the optimistic section execution queue OptExecQueue, the gate buffer GateBuffer, and the outbox Outbox for the optimistic section described in Step 6.1 , using batch packing, bucket aggregation according to the neighbor dimension, and one-time transmission using the derived data type of the message passing interface (MPI). Step 6.

5. Communication Progress and Progress Guarantee: To guarantee the progress of communication in step 6.2 global virtual time GVT, step 6.3 fossil collection, and step 6.4 batch packing and transmission, an independent progress thread or periodic polling is used to ensure that communication can still progress in the computation-intensive section; a high water level threshold is set for the gate buffer outside the gate and the outbox to be confirmed, and if it is exceeded, it will trigger a priority flush or temporary degradation; ​ Step 6.6, Parameter Adaptation: During the above running process, according to the high quantile value of the estimated round-trip time RTT The safety margin compensation term is adaptively adjusted; the upper threshold θ↑ and the lower threshold θ↓ are dynamically updated according to the mean and variance of the historical expected utility U; and the minimum residence time is adjusted according to the recent rollback density and the gate buffer level, so as to improve the flexibility of mode switching while ensuring stability. Step 6.7, multi-layer network hierarchical gate: in a multi-layer network scenario, on the basis of the global virtual time GVT and the message processing result obtained by the gating and buffering mechanism and steps 6.2-6.4, a gate gate_time_intra≥ gate_time_inter is set according to the level, wherein gate_time_intra refers to the gate time calculated for the same layer and the near-end neighbor, and gate_time_inter refers to the gate time calculated for the cross-layer and the far-end neighbor. Through the cooperative constraint of the hierarchical gate time and the global virtual time GVT, only when the event timestamp ts is less than or equal to the corresponding gate and the event timestamp ts is less than or equal to the global virtual time GVT, the related event is visible and submitted to the outside, so as to realize the globally causal consistent synchronization and parallel advancement under the multi-layer network.

7. The intelligent hybrid parallel synchronous approach for large-scale distributed simulation according to claim 6, wherein: In the step 6.5, the communication advance and progress assurance includes independent progress thread or periodic Polling, the gate outside buffer and the outbound table configuration high water level threshold, exceeding the high water level threshold triggers the priority flushing or temporary degradation.

8. The intelligent hybrid parallel synchronous approach for large scale distributed simulation according to claim 1, wherein: In the conservative mode, neighborhood lower bound is used to advance the causal order and empty messages are used to maintain the causal order. In the optimistic mode, differential state log and anti-messages are used to support rollback. The undetermined events are not visible to the conservative side during the mode switching.

Images