PCIe active retransmission method based on GRU neural network prediction, switch chip and system

Abstract

The application discloses a PCIe active retransmission method based on GRU neural network prediction, an exchange chip and a system, and the method comprises the following steps: collecting quality indexes of each data link and performing window aggregation to obtain a first feature vector; using a GRU recurrent neural network to infer and predict based on the previous first feature vector to obtain pre-retransmission data packets in a data packet queue to be transmitted, the retransmission risk probability of the pre-retransmission data packets being higher than that of non-pre-retransmission data packets; locking the storage position of the pre-retransmission data packets in a retransmission buffer, and constructing a preprocessing table comprising a first index of each pre-retransmission data packet, the first index containing a data pointer indicating the storage position; if the response signal corresponding to the pre-retransmission data packet is NACK, the retransmission controller determines that the pre-retransmission data packet is a target retransmission data packet and retransmits the target retransmission data packet in the retransmission buffer. The application can significantly reduce the retransmission reaction time, improve the link reliability and the throughput stability.

Description

Technical Field

[0001] This invention relates to the field of high-speed interconnect and intelligent data communication technology, and in particular to a PCIe active retransmission method, switching chip and system based on GRU neural network prediction. Background Technology

[0002] PCI Express (PCIe) is currently the most mainstream high-speed serial interconnect technology in data centers, server motherboards, AI accelerators, and high-speed storage systems. Its data link layer uses sequence number (SN), acknowledgment / negation acknowledgment (ACK / NACK) signals and retransmission mechanisms to ensure the reliability of data transmission and realize message acknowledgment and retransmission across high-speed serial links.

[0003] Traditional PCIe data retransmission processes have clear and strictly defined working steps, mainly including message sending and retransmission buffering mechanisms, link transmission and receiver verification processes, SN continuity checks, and byte block structure checks.

[0004] The message sending and retransmission buffering mechanism works as follows: when the sender is about to send a TLP (Transaction Layer Packet), it assigns it a monotonically increasing sequence number (SN) and writes the message along with the sequence number and related metadata into the retransmission buffer. The retransmission buffer is used to temporarily store all TLP data packets that have been sent but not yet acknowledged by the receiver. If an error occurs later, the message can be rescheduled from this area for retransmission.

[0005] The link transmission and receiver verification process includes encapsulating the TLP data packet at the link layer and then transmitting it to the receiver through the physical layer. After receiving the message, the receiver performs the following verification steps: CRC check (LCRC). If the CRC is correct and the sequence number is continuous, the message is considered complete and error-free; SN continuity check. If the SN is not continuous or skips numbers, the message is considered missing; byte block structure check (under 128b / 130b encoding methods, etc.).

[0006] Based on the above check results, the receiving end will return ACK (acknowledgment) or NACK (denial) to the sending end: ACK indicates that this TLP and all previously unacknowledged TLPs have been successfully received; NACK indicates that several TLPs within the current TLP or SN range have errors or are missing and need to be retransmitted.

[0007] like Figure 1As shown, the industry generally believes that the PCIe retransmission triggering mechanism mainly includes three types of situations: (1) No error in ideal link conditions: the receiver continuously returns ACKs and the sender continuously clears the confirmed TLPs in the retransmission buffer; (2) TLP error or SN discontinuity: the receiver returns NACKs for TLPs with CRC errors or SN disorder, and the sender retransmits the corresponding TLPs from the retransmission buffer accordingly; (3) ACK timeout or loss: if the sender does not receive ACKs within the specified time or the ACKs are lost, the retransmission delay timeout will trigger the sender to perform retransmission.

[0008] As can be seen, in the traditional PCIe architecture, the sending end only retransmits the corresponding packet from the replay buffer after the receiving end detects a packet checksum error and sends a NACK signal. Therefore, the entire retransmission path is entirely dependent on link event triggering, exhibiting significant passivity and unpredictability. The retransmission delay is affected by the NACK round-trip delay, the replay state machine startup time, and the buffer scheduling delay.

[0009] Therefore, although the retransmission mechanism of the PCIe standard is designed to be stable and reliable, its limitations are gradually becoming apparent as bandwidth in data centers, AI training clusters, NVMe SSDs, and high-density interconnect systems continues to increase. The main problems include: traditional mechanisms must wait for CRC failure, SN discontinuity, or ACK timeout before retransmitting. This means that the sender cannot detect link quality deterioration in advance, nor can it prepare data packets that need to be retransmitted or schedule retransmission buffers in advance, thus retransmission delay becomes an unavoidable tail delay.

[0010] Furthermore, in actual operation of high-speed links, the physical layer of PCIe high-speed links is affected by various microsecond-level short-term disturbances such as jitter spikes, instantaneous bit error spikes, DFE tap saturation, temperature drift, and power supply noise. Although these short-term anomalies may last for a very short time (microseconds), they are enough to cause packet corruption and trigger retransmission. Existing PCIe switching chips generally lack the ability to predict these transient characteristics and can only initiate retransmission after an error has occurred, a NACK has been returned, or a timer has expired. They always try to remedy the situation after an anomaly has occurred, thus introducing additional link instability and causing additional round-trip time (RTT) and retransmission delay, significantly increasing the system tail latency (P99 / P999) and further increasing the pressure on the retransmission buffer in high-traffic scenarios.

[0011] In summary, existing PCIe architecture data link solutions cannot distinguish between short-term fluctuations and long-term degradation of the link, and lack intelligent scheduling or prioritization mechanisms, making it difficult to meet the requirements of low-tail latency and high stability in high-bandwidth scenarios such as data centers, AI acceleration clusters, and NVMe SSDs.

[0012] The above background information is provided only to assist in understanding the inventive concept and technical solution of this invention. It does not necessarily belong to the prior art of this application, nor does it necessarily provide technical teaching. In the absence of clear evidence that the above information was disclosed before the filing date of this application, the above background information should not be used to evaluate the novelty and inventiveness of this application. Summary of the Invention

[0013] The purpose of this invention is to provide a PCIe active retransmission method, switching chip and system based on GRU neural network prediction, which can significantly reduce retransmission response time and improve link reliability and throughput stability.

[0014] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0015] A PCIe proactive retransmission method based on GRU neural network prediction includes the following steps:

[0016] Quality metrics of each data link are collected, and window aggregation is performed on the quality metrics at a preset inference period to obtain a first feature vector. The quality metrics include multiple of the following: bit error rate, adaptive tap coefficient, carrier frequency offset, and clock offset.

[0017] A GRU recurrent neural network is used to perform inference and prediction based on the first feature vector corresponding to one or more prior time windows to obtain the pre-retransmission data packets in the data packet queue to be transmitted. The retransmission risk probability corresponding to the pre-retransmission data packets is higher than the retransmission risk probability corresponding to the non-pre-retransmission data packets.

[0018] The storage location of the pre-retransmission data packet in the retransmission buffer is locked, and a preprocessing table is constructed. The preprocessing table includes a first index corresponding to each pre-retransmission data packet, and the first index includes a data pointer for indicating the storage location of the pre-retransmission data packet.

[0019] Monitor the acknowledgment signal corresponding to each of the pre-retransmission data packets. If the acknowledgment signal corresponding to the pre-retransmission data packet is NACK, then determine that the pre-retransmission data packet is the target retransmission data packet, and construct a fast retransmission queue. The fast retransmission queue includes a second index corresponding to each target retransmission data packet, and the second index includes the data pointer.

[0020] The retransmission controller retransmits the target retransmission data packets indicated by each second index in the fast retransmission queue, including: the retransmission controller determines the storage location of the target retransmission data packet in the retransmission buffer according to the data pointer as the target storage location, and retransmits the target retransmission data packet stored at the target storage location.

[0021] Furthermore, following any one or a combination of the aforementioned technical solutions, the pre-retransmission data packets in the queue of data packets to be transmitted are predicted using a GRU recurrent neural network in the following manner:

[0022] A GRU recurrent neural network is used to predict the probability of a link error occurring in the data link within a future target time window based on the first feature vector corresponding to one or more prior time windows.

[0023] If the probability of a link error is greater than a preset first failure rate threshold for N consecutive times, and N is not less than a preset first number threshold, then the data packets transmitted through the data link within the future target time window are determined to be pre-retransmission data packets.

[0024] Furthermore, following any one or a combination of the aforementioned technical solutions, a pre-trained GRU prediction engine is set inside the switching chip. The GRU prediction engine includes a fully connected layer and a softmax output layer. The input of the GRU prediction engine is the first feature vector, and the output of the GRU prediction engine is the probability of each data link experiencing a link error as predicted.

[0025] Furthermore, following any one or a combination of the aforementioned technical solutions, a GRU recurrent neural network is used to predict the probability of a link error occurring in the data link within a future target time window based on the first feature vector corresponding to one or more prior time windows.

[0026] If the probability of a link error occurring in the data link is greater than a preset second failure rate threshold, and the second failure rate threshold is greater than the first failure rate threshold, then the retransmission controller determines the data packet that needs to be transmitted from the data link in the preprocessing table as the target retransmission data packet.

[0027] The retransmission controller determines the data pointer corresponding to the target retransmission data packet from the fast retransmission queue as the target data pointer, and retransmits the data packet corresponding to the target data pointer in the retransmission buffer.

[0028] Furthermore, following any one or a combination of the aforementioned technical solutions, the method further includes the following steps:

[0029] The feature indicators of each data packet in the queue of data packets to be transmitted are collected and window aggregation is performed on the feature indicators at a preset inference period to obtain a second feature vector. The feature indicators include multiple of the following: packet length, packet type, address sequence pattern, and routing information.

[0030] A GRU recurrent neural network is used to perform inference and prediction based on the first feature vector and the second feature vector corresponding to one or more prior time windows to obtain the pre-retransmission data packets in the data packet queue to be transmitted. The retransmission risk probability corresponding to the pre-retransmission data packets is higher than the retransmission risk probability corresponding to the non-pre-retransmission data packets.

[0031] Furthermore, following any one or a combination of the aforementioned technical solutions, the pre-retransmission data packets in the queue of data packets to be transmitted are predicted using a GRU recurrent neural network in the following manner:

[0032] A pre-trained GRU prediction engine is set inside the switching chip. The GRU prediction engine includes a fully connected layer and a softmax output layer. The input of the GRU prediction engine is the first feature vector and the second feature vector. The output of the GRU prediction engine is the retransmission risk probability corresponding to each data packet.

[0033] If the retransmission risk probability corresponding to the data packet is greater than the preset retransmission risk probability threshold for M consecutive times and M is not less than the preset second number threshold, then the data packet is determined to be a pre-retransmission data packet.

[0034] Furthermore, following any one or a combination of the aforementioned technical solutions, if the acknowledgment signal corresponding to the pre-retransmitted data packet is ACK, then the retransmission controller deletes the first index corresponding to ACK in the preprocessing table; and / or,

[0035] If the acknowledgment signal corresponding to the target retransmitted data packet is ACK, then the retransmission controller deletes the second index corresponding to ACK from the fast retransmission queue; and / or,

[0036] If the acknowledgment signal corresponding to the target retransmitted data packet is still NACK, the retransmission controller will retransmit the target retransmitted data packet again. If the acknowledgment signal corresponding to the target retransmitted data packet is NACK two or more times consecutively, the NACK acknowledgment signal will be returned to the sending end.

[0037] Furthermore, in accordance with any or a combination of the aforementioned technical solutions, the retransmission controller is further configured to determine, when an anomaly is determined in the data link, the data packet in the data packet queue that needs to be transmitted from the data link is the target retransmission data packet;

[0038] If the data link meets one of the following conditions, then the data link is determined to be abnormal:

[0039] The change in any physical layer parameter in the data link is greater than the corresponding preset change threshold.

[0040] The jitter magnitude of the data link is greater than a preset jitter magnitude threshold;

[0041] The DFE tap corresponding to the data link is saturated.

[0042] Furthermore, in accordance with any or a combination of the aforementioned technical solutions, the retransmission controller is further configured to determine whether an equivalent link exists on the data link corresponding to the target retransmission data packet; if so, the retransmission controller controls the target retransmission data packet to be retransmitted through the equivalent link.

[0043] And / or,

[0044] For each sending end, there are equivalent first and second links. The first link is configured for data transmission in non-retransmission mode, and the second link is configured for data transmission in retransmission mode.

[0045] Furthermore, following any one or a combination of the aforementioned technical solutions, window aggregation includes calculating one or more of the following: mean, variance, quantiles, and extreme values; and / or,

[0046] It also includes performing window aggregation on the quality indicators at a preset inference cycle and normalizing the quality indicators after window aggregation to obtain a first feature vector; and / or,

[0047] The first feature vector is configured to be stored in a preset telemetry first-in-first-out buffer and / or independent of other buffers.

[0048] The first index includes a packet header, sequence number, data pointer, and packet status information corresponding to each of the pre-retransmission data packets; and / or,

[0049] The second index includes a header and a data pointer corresponding to each of the said target retransmission data packets.

[0050] According to another aspect of the present invention, a switching chip suitable for active retransmission is provided, including a telemetry acquisition module, a data processing module, a GRU prediction engine, a retransmission controller, and a retransmission buffer.

[0051] The telemetry acquisition module is configured to collect quality indicators of each data link.

[0052] The data processing module is configured to perform window aggregation on the quality indicators at a preset inference period to obtain a first feature vector. The quality indicators include multiple of the following: bit error rate, adaptive tap coefficient, carrier frequency offset, and clock offset.

[0053] The GRU prediction engine is configured to perform inference and prediction on the first feature vector corresponding to one or more prior time windows based on the GRU recurrent neural network to obtain the pre-retransmission data packets in the data packet queue to be transmitted, wherein the retransmission risk probability corresponding to the pre-retransmission data packets is higher than the retransmission risk probability corresponding to the non-pre-retransmission data packets.

[0054] The retransmission controller is configured to perform a pre-retransmission operation and a fast retransmission operation. The pre-retransmission operation includes: locking the storage location of the pre-retransmission data packet in the retransmission buffer, and constructing a preprocessing table, the preprocessing table including a first index corresponding to each pre-retransmission data packet, the first index including a data pointer for indicating the storage location of the pre-retransmission data packet, and obtaining the acknowledgment signal corresponding to each pre-retransmission data packet.

[0055] The fast retransmission operation includes: in response to determining that the acknowledgment signal corresponding to the pre-retransmission data packet is NACK, the retransmission controller determines that the pre-retransmission data packet is a target retransmission data packet and constructs a fast retransmission queue, the fast retransmission queue including a second index corresponding to each target retransmission data packet, the second index including the data pointer; the retransmission controller determines the storage location of the target retransmission data packet according to the second index and retransmits it.

[0056] According to another aspect of the present invention, a communication system suitable for active retransmission is provided, including a switching chip and a plurality of devices to be communicated, each of the devices being electrically connected to the switching chip, the switching chip performing communication between the plurality of devices based on the PCIe active retransmission method based on GRU neural network prediction as described in any one or a combination of the above technical solutions.

[0057] The beneficial effects of the technical solution provided by this invention are as follows:

[0058] a. This invention collects data link quality indicators in real time at the physical layer of the data link, constructs a first feature vector reflecting the link quality, and uses a lightweight GRU model to predict the probability of data link errors. It can identify error trends in advance before link errors occur. Through this predictive judgment, it can enter the pre-retransmission state before potential errors occur, select high-risk packets, and complete resource locking and fast retransmission preparation for high-risk packets without changing the protocol semantics. This fundamentally eliminates the latency bottleneck caused by NACK waiting in traditional retransmission mechanisms, thereby significantly reducing retransmission response time and improving link reliability and throughput stability.

[0059] b. This invention constructs a two-level structure of a preprocessing table and a fast retransmission queue, realizing the pre-indexing and queued loading of high-risk data packets, forming a fast retransmission path parallel to the retransmission buffer. The preprocessing table only stores the first index of high-risk packets and does not copy their contents; the fast retransmission queue caches the second index containing the packet header and data pointer, ensuring that in the event of a sudden change in link quality or NACK, the retransmission controller can directly schedule data from the retransmission buffer and send it quickly with zero table lookup and zero loading delay, achieving true instant retransmission. This mechanism effectively reduces the critical path delay caused by retransmission buffer access and scheduling, and significantly improves the available bandwidth and throughput of the link in high error rate scenarios.

[0060] c. This invention uses a multi-level screening method to select pre-retransmission data packets and target retransmission data packets with high retransmission risk, and deletes the first index of pre-retransmission data packets that are converted from high risk to low risk from the preprocessing table according to the dynamic changes in link quality, which can adaptively match the dynamic changes in data link quality.

[0061] d. The active retransmission method, switching chip, and communication system provided by this invention are not only applicable to PCIe 5.0 / 6.0, but can also be seamlessly extended to other high-speed link interfaces such as CXL, SerDes, and Chiplet interconnects. This provides a highly scalable intelligent link reliability enhancement approach for future multi-chip modules (MCMs), high-density interconnects, and server-level high-speed switching chips, and has significant industrial application value and promotion prospects. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 A flowchart of the PCIe retransmission mechanism in the prior art;

[0064] Figure 2 A flowchart of a retransmission mechanism including pre-retransmission and fast retransmission processes is provided as an exemplary embodiment of the present invention.

[0065] Figure 3 A schematic diagram of a switching chip module provided for an exemplary embodiment of the present invention;

[0066] Figure 4 A flowchart of an active retransmission triggered by a link anomaly is provided as an exemplary embodiment of the present invention;

[0067] Figure 5 A flowchart of a PCIe proactive retransmission method based on GRU neural network prediction, provided as an exemplary embodiment of the present invention;

[0068] Figure 6 A flowchart for constructing a fast retransmission queue is provided as an exemplary embodiment of the present invention. Detailed Implementation

[0069] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0070] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0071] In recent years, with the widespread use of GPUs (Graphics Processing Units) in inference, it has become possible to predict high-speed link parameters in real time using lightweight deep neural networks. By modeling the physical layer (PHY) telemetry data of the data link, AI can identify potential risks in the data link before errors occur, enabling PCIe chips to evolve from "passive retransmission" to "active retransmission." This invention is proposed against this technological background, aiming to use GRU (Gated Recurrent Unit) neural networks to predict link error trends and prepare retransmission resources in advance, thereby significantly reducing retransmission latency and improving link reliability.

[0072] In one embodiment of the present invention, a PCIe proactive retransmission method based on GRU neural network prediction is provided, see [link to relevant documentation]. Figures 2 to 5 This includes the following steps:

[0073] Quality metrics of each data link are collected, and window aggregation is performed on the quality metrics at a preset inference period to obtain a first feature vector. The quality metrics include multiple of the following: bit error rate, adaptive tap coefficient, carrier frequency offset, and clock offset.

[0074] A GRU recurrent neural network is used to perform inference and prediction based on the first feature vector corresponding to one or more prior time windows to obtain the pre-retransmission data packets in the data packet queue to be transmitted. The retransmission risk probability corresponding to the pre-retransmission data packets is higher than the retransmission risk probability corresponding to the non-pre-retransmission data packets.

[0075] The storage location of the pre-retransmission data packet in the retransmission buffer is locked, and a preprocessing table is constructed. The preprocessing table includes a first index corresponding to each pre-retransmission data packet, and the first index includes a data pointer for indicating the storage location of the pre-retransmission data packet in the retransmission buffer.

[0076] Monitor the acknowledgment signal corresponding to each of the pre-retransmission data packets. If the acknowledgment signal corresponding to the pre-retransmission data packet is NACK, then determine that the pre-retransmission data packet is the target retransmission data packet, and construct a fast retransmission queue. The fast retransmission queue includes a second index corresponding to each target retransmission data packet, and the second index includes the data pointer.

[0077] The retransmission controller retransmits the target retransmission data packets indicated by each second index in the fast retransmission queue, including: the retransmission controller determines the storage location of the target retransmission data packet in the retransmission buffer according to the data pointer as the target storage location, and retransmits the target retransmission data packet stored at the target storage location.

[0078] This invention enhances the reliability of data links in PCI Express (PCIe) switching chips and reduces retransmission latency. When a high-risk data link condition is predicted, it proactively enters a "prepared retransmission mode," pre-locking high-risk data packets, shortening the retransmission timer, and generating a first index based on the TLP packet header, data pointer, and packet status, adding it to the preprocessing table. Upon a genuine anomaly or a NACK, the high-risk data packet is determined to be a packet requiring retransmission. A second index is generated based on the first index and placed in the Fast Path queue, allowing immediate retransmission of the corresponding TLP packet with zero waiting time. This effectively eliminates the NACK round-trip latency and retransmission timer startup delay found in traditional solutions, significantly reducing retransmission response time and improving link reliability and throughput stability. This invention elevates the PCIe retransmission mechanism from "passive remediation" to "proactive prediction," making it particularly suitable for applications highly sensitive to tail latency, such as data centers, AI acceleration clusters, and high-speed storage. Its applications include data center server motherboards, PCIe switching chips, AI accelerator cards, NVMe SSD arrays, cloud computing cluster interconnects, high-speed network equipment, and any high-reliability, high-bandwidth system using the PCIe protocol. This invention can also be extended to other high-speed serial links, such as CXL, Ethernet SerDes, InfiniBand, and Chiplet Interconnect.

[0079] In one embodiment of the present invention, a pre-configured telemetry acquisition module is used to telemetry the quality indicators of each data link at the PCIe PHY layer, including: bit error rate (instantaneous BER estimate), random jitter, deterministic jitter, CTLE / DFE adaptive tap coefficients, carrier frequency offset and clock offset, and operating environment parameters such as temperature and voltage. Specifically, the sampling frequency is based on the GRU inference period Tinf (e.g., 5–20µs).

[0080] Within each inference cycle, window aggregation (mean, variance, quantiles, maximum / minimum) is performed on the above parameters to obtain a fixed-length feature vector. Further, a normalization operation is performed on the fixed-length feature vector obtained from the window aggregation to obtain the first feature vector, thereby enhancing model robustness. The processed first feature vector is written to a telemetry FIFO buffer, independent of other buffers, via an AXI bus. The data in the telemetry FIFO buffer serves as input to the GRU prediction engine.

[0081] In this embodiment, the telemetry acquisition module is further configured to acquire feature indicators of each data packet in the data packet queue to be transmitted, and perform window aggregation and normalization processing on the feature indicators at a preset inference period to obtain a second feature vector. The feature indicators include multiple parameters such as packet length, packet type, address sequence pattern, and routing information. The time window for acquiring the feature indicators and the quality indicators is the same, and the inference period for both is also the same.

[0082] See Figure 2 and Figure 5 A pre-trained GRU prediction engine is set inside the switching chip. The GRU prediction engine includes a fully connected layer and a softmax output layer. The input of the GRU prediction engine is the first feature vector and the second feature vector. The output of the GRU prediction engine is the retransmission risk probability corresponding to each data packet.

[0083] A GRU recurrent neural network is used to perform inference and prediction based on the first feature vector and the second feature vector corresponding to one or more prior time windows to obtain the pre-retransmission data packets in the data packet queue to be transmitted. The retransmission risk probability corresponding to the pre-retransmission data packets is higher than the retransmission risk probability corresponding to the non-pre-retransmission data packets.

[0084] If the retransmission risk probability corresponding to the data packet is greater than the preset retransmission risk probability threshold for M consecutive times and M is not less than the preset second number threshold, then the data packet is determined to be a pre-retransmission data packet.

[0085] This invention employs a lightweight GRU neural network (≤100k parameters), setting the pre-trained GRU prediction engine to run on the on-chip NPU or lightweight hard core of the switching chip. Through fully connected layers and softmax, it outputs: the probability of a link error occurring in the future window, such as the next window, and the retransmission risk distribution of the corresponding TLP data packets to be transmitted in the future window. The total time for inference and data handling is ≤2-5µs, which is far lower than the minimum threshold for PCIe retransmission latency.

[0086] When the GRU prediction engine is able to support inference prediction normally, inference is triggered once per cycle. If the GRU prediction engine fails, such as when the GRU times out, or when the inference unit (NPU) fails, it automatically degrades to the traditional NACK process.

[0087] A GRU recurrent neural network is used to predict the probability of a link error occurring in the data link within a future target time window based on a first feature vector corresponding to one or more prior time windows. If the probability of a link error occurring in the data link is greater than a preset second failure rate threshold, and the second failure rate threshold is greater than the first failure rate threshold, then the retransmission controller determines the data packet that needs to be transmitted from the data link in the preprocessing table as the target retransmission data packet. The retransmission controller determines the data pointer corresponding to the target retransmission data packet from the preprocessing table as the target data pointer, and retransmits the data packet corresponding to the target data pointer in the retransmission buffer.

[0088] After determining the data packet to be retransmitted, the retransmission controller enters the pre-retransmission mode. Compared to existing PCIe data transmission, which introduces an optional intermediate state, the pre-retransmission mode, into the original retransmission state machine before confirming the success of data transmission, this does not require changes to the PCIe protocol.

[0089] The pre-retransmission data packet is a high-risk retransmission data packet among the TLP data packets to be transmitted. The storage location of the pre-retransmission data packet in the retransmission buffer is locked, and the data pointer of the pre-retransmission data packet is fixed to prevent it from being overwritten by the normal transmission process. It should be noted that in the implementation of a traditional PCIe retransmission buffer, the sender typically manages the storage area of ​​unacknowledged TLP data packets through two types of key pointers: a head pointer, which points to the earliest TLP data packet that has not yet received an ACK, and a tail pointer, which points to the next position where a new TLP can be written. In this invention, high-risk TLP data packets have not yet been ACKed and need to remain in the retransmission buffer for fast retransmission; they cannot be overwritten by the normal writing of the tail pointer, so their corresponding data pointers need to be fixed. This is mainly achieved by recording the retransmission index bit (first index) of each pre-retransmission data packet and locking its flag bit's data pointer. Simultaneously, the tail pointer automatically skips the locked area. This mechanism ensures that the normal transmission process does not overwrite any entries marked as high-risk in the retransmission buffer when writing new TLP packets, thus guaranteeing that these pre-retransmission packets can be used for immediate scheduling in the Transmission Pre-Queue (TPQ). Entering pre-retransmission mode reduces subsequent retransmission waiting times by at least half, thereby minimizing redundancy due to timeouts.

[0090] In the event of NACK, PHY anomaly, or DFE tap saturation, if the retransmission controller is already in pre-retransmission mode and TPQ is ready, it directly enters fast retransmission mode: zero-wait retransmission, data packets in TPQ are directly injected into the data link, ensuring the SeqNo order remains unchanged. Upon receiving an ACK, the corresponding retransmission buffer entry is cleared, the corresponding first index is removed from the preprocessing table (PrepTable), and the corresponding second index is deleted from the fast retransmission queue. If another NACK is received, the same packet is immediately retransmitted again (≤2 times). If no acknowledgment or timeout occurs, it is considered a link defect, exiting the normal mechanism and triggering the traditional retransmission process. Specifically, the retransmission controller activates fast retransmission mode and executes fast retransmission operations under the following conditions.

[0091] Condition 1: If there exists an acknowledgment signal corresponding to one of the pre-retransmission data packets that is NACK.

[0092] In response to detecting a NACK signal corresponding to a pre-retransmission data packet, the retransmission controller determines that the pre-retransmission data packet is the target retransmission data packet. The retransmission controller determines the data pointer corresponding to the target retransmission data packet from the fast retransmission queue as the target data pointer, and retransmits the data packet corresponding to the target data pointer in the retransmission buffer.

[0093] Condition 2: There is a data link in an abnormal state.

[0094] If the data link meets one of the following conditions, it is determined that the data link is abnormal: the change in any physical layer parameter in the data link is greater than the corresponding preset change threshold; the jitter magnitude of the data link is greater than the preset jitter magnitude threshold; or the DFE tap is saturated.

[0095] The retransmission controller is further configured to determine, when an anomaly is detected in the data link, the data packet in the data packet queue that needs to be transmitted from the data link is the target retransmission data packet.

[0096] The retransmission controller is further configured to determine whether an equivalent link exists on the data link corresponding to the target retransmission data packet. If an equivalent link exists, the retransmission controller controls the target retransmission data packet to be retransmitted through the equivalent link. A data link and its equivalent link correspond to the same sender and destination.

[0097] More preferably, two parallel data links are constructed inside the sending end. One is configured as a regular sending link for the normal transaction layer (TLP), and the other is a fast retransmission link dedicated to fast retransmission. This does not increase the number of physical ports, but rather adds an independent fast retransmission data path inside the link layer. This is used to immediately send the data packets that need to be retransmitted through the fast retransmission link when a high risk is predicted for the link or when a NACK is received.

[0098] If the acknowledgment signal corresponding to the pre-retransmission data packet is ACK, the retransmission controller will delete the packet header and data pointer corresponding to ACK in the fast retransmission queue, and also delete the pre-retransmission data packet corresponding to the retransmission buffer.

[0099] The retransmission controller periodically refreshes the queue of data packets to be transmitted, updating the time or number of cycles elapsed since each data packet entered the unacknowledged state, the probability or risk level of future errors of the data packets, and updating the preprocessing table and the fast retransmission queue. Through this periodic refresh mechanism, the retransmission controller obtains these parameters, enabling the system to continuously track high-risk TLPs in real time and ensuring that resource scheduling in the fast retransmission path remains accurate and effective.

[0100] The proposed active retransmission method for PCIe switching chips based on GRU neural network prediction offers significant advantages over existing passive retransmission processes that rely on NACK, in terms of response speed, resource utilization, and retransmission efficiency in link error handling. Firstly, traditional PCIe link layers must wait for a NACK signal from the receiver before initiating a retransmission process, with trigger latency limited by both link round-trip delay and the lower limit of ReplayTimer (typically several microseconds). This invention, by real-time acquisition of data link quality indicators at the PHY layer, including BER, jitter, DFE taps, CFO, and eye diagram opening, constructs a first feature vector reflecting link quality. This vector is then used for prediction using an on-chip lightweight GRU model, enabling the switching chip to identify error trends before link errors occur. Through this predictive judgment, the system can enter a preparatory state before potential errors appear, selecting high-risk packets and completing resource locking and rapid retransmission preparation without altering protocol semantics, fundamentally eliminating the latency bottleneck caused by NACK waiting in traditional mechanisms.

[0101] Secondly, this invention constructs a two-level structure of a preprocessing table and a Fast Retransmit Queue (TPQ) for recording high-risk data packets. This enables pre-indexing and queued loading of high-risk packets, forming a fast retransmit path parallel to the retransmit buffer. The first index in the preprocessing table, corresponding to each high-risk data packet, records the packet header, sequence number, data pointer, and packet status information, including acknowledged and unacknowledged status, without copying the packet content. The second index in the TPQ, corresponding to each target retransmitted data packet, caches the packet header and data pointer. This ensures that in the event of a sudden change in link quality or a NACK, the retransmit controller can directly schedule data from the retransmit buffer and send it quickly with zero lookup and zero loading latency, achieving true instant retransmission. This mechanism effectively reduces critical path latency caused by delayed access and scheduling during retransmission, significantly improving the available bandwidth and throughput of the link in high-error-rate scenarios.

[0102] Furthermore, this invention maintains full PCIe protocol compatibility: the Replay Timer, SeqNo, and ACK / NACK mechanisms remain unchanged; only prediction, preparation, and fast path logic are added at the control level, thus ensuring secure interoperability with existing PCIe hosts and devices. In the event of GRU inference timeout, model failure, or prediction uncertainty, the system automatically degrades to a traditional retransmission process, exhibiting good security and robustness. In terms of hardware implementation, the GRU model has fewer than 100k parameters and can be implemented within a small on-chip NPU or logic unit, resulting in minimal impact on chip area, power consumption, and BOM cost. The storage capacity of TPQ and PrepTable is also very limited, making the overall solution highly feasible in engineering.

[0103] Finally, the active retransmission mechanism proposed in this patent is not only applicable to PCIe 5.0 / 6.0, but can also be seamlessly extended to other high-speed link interfaces such as CXL, SerDes, and Chiplet interconnects. It provides a highly scalable intelligent link reliability enhancement approach for future multi-chip modules (MCM), high-density interconnects, and server-grade high-speed switching chips, and has significant industrial application value and promotion prospects.

[0104] In another embodiment of the present invention, compared to the above embodiment, this embodiment does not consider the impact of data packets on retransmission risk. A pre-trained GRU prediction engine is set inside the switching chip. The GRU prediction engine includes a fully connected layer and a softmax output layer. The input of the GRU prediction engine is the first feature vector, and the output of the GRU prediction engine is the predicted value of the probability of link errors occurring in each data link. Specifically, a GRU recurrent neural network can be used to predict the probability of link errors occurring in a future target time window based on the first feature vector corresponding to one or more prior time windows.

[0105] See Figure 6 If the probability of a link error is greater than a preset first failure rate threshold for N consecutive times, and N is not less than a preset first number of failures threshold, then the data packets transmitted through the data link within the future target time window are determined to be pre-retransmission data packets. Before the acknowledgment signal of the data packet is returned, the probability of a link error occurring in the data link is continuously updated. If the predicted probability of a link error is not greater than the first failure rate threshold for several consecutive times, then the preprocessed data packets that were originally high-risk are determined to be low-risk data packets, and the corresponding first index in the preprocessing table is deleted.

[0106] For each preprocessed data packet in the preprocessing table, if the corresponding data link quality further deteriorates, and the probability of a predicted link error exceeds a second failure rate threshold, which in turn exceeds a preset first failure rate threshold, then the preprocessed data packet is determined to be target retransmission data. Based on its corresponding first index, a second index is determined, and the packet is placed in the fast retransmission queue. It should be noted that the second index can be the same as the first index. More preferably, the second index is a further simplified version of the first index; the second index simply indicates the storage location of the target retransmission data in the retransmission buffer.

[0107] Furthermore, for each preprocessed data packet in the preprocessing table, if its corresponding acknowledgment signal is NACK, then the preprocessed data packet is determined to be target retransmission data, and a second index is determined based on its corresponding first index and placed in the fast retransmission queue. If its corresponding acknowledgment signal is ACK, it indicates that it has been received by the receiving end / target end, and the first index information corresponding to the preprocessing table can be deleted.

[0108] In one embodiment of the present invention, a switching chip suitable for active retransmission is provided, see [link to relevant documentation]. Figure 2 The switch includes a telemetry acquisition module, a data processing module, a prediction engine, a retransmission controller, and a retransmission buffer.

[0109] The telemetry acquisition module is configured to collect quality indicators of each data link.

[0110] The data processing module is configured to perform window aggregation on the quality indicators at a preset inference period to obtain a first feature vector. The quality indicators include multiple of the following: bit error rate, adaptive tap coefficient, carrier frequency offset, and clock offset.

[0111] The GRU prediction engine is configured to perform inference and prediction on the first feature vector corresponding to one or more prior time windows based on the GRU recurrent neural network to obtain the pre-retransmission data packets in the data packet queue to be transmitted, wherein the retransmission risk probability corresponding to the pre-retransmission data packets is higher than the retransmission risk probability corresponding to the non-pre-retransmission data packets.

[0112] The retransmission controller is configured to perform a pre-retransmission operation and a fast retransmission operation. The pre-retransmission operation includes: locking the storage location of the pre-retransmission data packet in the retransmission buffer, and constructing a preprocessing table, the preprocessing table including a first index corresponding to each pre-retransmission data packet, the first index including a data pointer for indicating the storage location of the pre-retransmission data packet, and obtaining the acknowledgment signal corresponding to each pre-retransmission data packet.

[0113] The fast retransmission operation includes: in response to determining that the acknowledgment signal corresponding to the pre-retransmission data packet is NACK, the retransmission controller determines that the pre-retransmission data packet is a target retransmission data packet and constructs a fast retransmission queue, the fast retransmission queue including a second index corresponding to each target retransmission data packet, the second index including the data pointer; the retransmission controller determines the storage location of the target retransmission data packet according to the second index and retransmits it.

[0114] In a specific application, such as Figure 3 As shown, the upstream host and multiple downstream terminal devices (such as...) Figure 1 The service data of terminal devices 1 and 2 shown first enters the switching chip through the multi-port PCIe PHY / SerDes layer. All TLP data packets to be sent are written to the retransmission buffer in the lower right corner as real data storage units. The PHY / SerDes module simultaneously outputs the link physical status to the telemetry acquisition module to realize PHY telemetry acquisition (such as BER, jitter, eye diagram, DFE tap, etc.). After feature construction and normalization, these telemetry data form a fixed-length first feature vector, which is sent to the GRU prediction engine in the middle of the chip. The GRU prediction engine outputs the probability of future link errors or failures and the retransmission risk distribution of each data packet to be confirmed. Then, it makes a judgment based on the preset risk fusion and threshold logic. If the predicted risk exceeds the threshold, a first index or a second index is generated and placed in the preprocessing table or fast retransmission queue. At the same time, a ready state flag (ENTER_PRE) signal or an immediate fast retransmission flag (FAST_TRIGGER) signal is generated and sent to the retransmission controller.

[0115] Upon receiving a high-risk trigger, the retransmission controller selects high-risk pre-retransmission data packets from the currently unacknowledged sequence in the retransmission buffer, writes their metadata into a first index, and sets a data pointer lock for the corresponding retransmission buffer entry, achieving pre-loading. When a NACK or link physical anomaly actually occurs, based on the first index of the high-risk packets, the packet headers and data pointers of the data packets that must be retransmitted in these high-risk packets form a second index and are copied to the TPQ queue. The retransmission controller immediately enters fast retransmission mode, directly retrieving the corresponding data from the retransmission buffer according to the indication of the second index in TPQ and sending it quickly via PHY, thereby bypassing the delay of traditional retransmission buffer retrieval and scheduling. After a successful ACK, the retransmission buffer, the high-risk packet index, and the corresponding entry in TPQ are synchronously cleared; if the risk decreases, the pre-retransmission mode is exited and normal transmission resumes. Throughout the process, the retransmission buffer always stores the original data of the acknowledged TLP data packets, the preprocessing table is used as a high-risk index, and TPQ serves as a pre-loading queue for fast retransmission. Together, these three constitute a predictable, pre-prepared, and instantly retransmittable proactive retransmission data path.

[0116] In one embodiment of the present invention, a communication system suitable for active retransmission is provided, including a switching chip and multiple devices to be communicated, each of the devices being electrically connected to the switching chip, and the switching chip performing communication between the multiple devices based on the PCIe active retransmission method based on GRU neural network prediction as described in any of the above embodiments.

[0117] It should be noted that the switching chip and communication system embodiments for active retransmission provided by the present invention are based on the same inventive concept as the above-described PCIe active retransmission method embodiments based on GRU neural network prediction. All contents of the PCIe active retransmission method embodiments based on GRU neural network prediction are incorporated into the switching chip and communication system embodiments for active retransmission by reference.

[0118] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0119] The above description is only a specific embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A PCIe active retransmission method based on GRU neural network prediction, characterized in that, Includes the following steps: Quality metrics of each data link are collected, and window aggregation is performed on the quality metrics at a preset inference period to obtain a first feature vector. The quality metrics include multiple of the following: bit error rate, adaptive tap coefficient, carrier frequency offset, and clock offset. A GRU recurrent neural network is used to predict the probability of a link error occurring in the data link within a future target time window based on the first feature vector corresponding to one or more prior time windows. If the probability of a link error is greater than a preset first failure rate threshold for N consecutive times, and N is not less than a preset first number threshold, then the data packets transmitted through the data link within the future target time window are determined to be pre-retransmission data packets. The retransmission risk probability corresponding to the pre-retransmission data packets is higher than the retransmission risk probability corresponding to non-pre-retransmission data packets. The storage location of the pre-retransmission data packet in the retransmission buffer is locked, and a preprocessing table is constructed. The preprocessing table includes a first index corresponding to each pre-retransmission data packet, and the first index includes a data pointer for indicating the storage location of the pre-retransmission data packet. Monitor the acknowledgment signal corresponding to each of the pre-retransmission data packets. If the acknowledgment signal corresponding to the pre-retransmission data packet is NACK, then determine that the pre-retransmission data packet is the target retransmission data packet, and construct a fast retransmission queue. The fast retransmission queue includes a second index corresponding to each target retransmission data packet, and the second index includes the data pointer. The retransmission controller retransmits the target retransmission data packets indicated by each second index in the fast retransmission queue, including: the retransmission controller determines the storage location of the target retransmission data packet in the retransmission buffer according to the data pointer as the target storage location, and retransmits the target retransmission data packet stored at the target storage location.

2. The PCIe proactive retransmission method based on GRU neural network prediction according to claim 1, characterized in that, A pre-trained GRU prediction engine is set inside the switching chip. The GRU prediction engine includes a fully connected layer and a softmax output layer. The input of the GRU prediction engine is the first feature vector, and the output of the GRU prediction engine is the probability of each data link experiencing a link error.

3. The PCIe proactive retransmission method based on GRU neural network prediction according to claim 2, characterized in that, A GRU recurrent neural network is used to predict the probability of a link error occurring in the data link within a future target time window based on the first feature vector corresponding to one or more prior time windows. If the probability of a link error occurring in the data link is greater than a preset second failure rate threshold, and the second failure rate threshold is greater than the first failure rate threshold, then the data packets that need to be transmitted from the data link in the preprocessing table are determined to be target retransmission data packets. For other pre-retransmission data packets in the preprocessing table, if the acknowledgment signal corresponding to the pre-retransmission data packet is NACK, then the pre-retransmission data packet is determined to be the target retransmission data packet.

4. The PCIe proactive retransmission method based on GRU neural network prediction according to claim 1, characterized in that, It also includes the following steps: The feature indicators of each data packet in the queue of data packets to be transmitted are collected and window aggregation is performed on the feature indicators at a preset inference period to obtain a second feature vector. The feature indicators include multiple of the following: packet length, packet type, address sequence pattern, and routing information. A GRU recurrent neural network is used to perform inference and prediction based on the first feature vector and the second feature vector corresponding to one or more prior time windows to obtain the pre-retransmission data packets in the data packet queue to be transmitted. The retransmission risk probability corresponding to the pre-retransmission data packets is higher than the retransmission risk probability corresponding to the non-pre-retransmission data packets.

5. The PCIe proactive retransmission method based on GRU neural network prediction according to claim 4, characterized in that, The GRU recurrent neural network is used to predict pre-retransmission packets in the packet queue to be transmitted in the following way: A pre-trained GRU prediction engine is set inside the switching chip. The GRU prediction engine includes a fully connected layer and a softmax output layer. The input of the GRU prediction engine is the first feature vector and the second feature vector. The output of the GRU prediction engine is the retransmission risk probability corresponding to each data packet. If the retransmission risk probability corresponding to the data packet is greater than the preset retransmission risk probability threshold for M consecutive times and M is not less than the preset second number threshold, then the data packet is determined to be a pre-retransmission data packet.

6. The PCIe active retransmission method based on GRU neural network prediction according to claim 1, characterized in that, If the acknowledgment signal corresponding to the pre-retransmitted data packet is ACK, then the retransmission controller deletes the first index corresponding to ACK in the preprocessing table; And / or, If the acknowledgment signal corresponding to the target retransmitted data packet is ACK, then the retransmission controller deletes the second index corresponding to ACK from the fast retransmission queue; And / or, If the acknowledgment signal corresponding to the target retransmitted data packet is still NACK, the retransmission controller will retransmit the target retransmitted data packet again. If the acknowledgment signal corresponding to the target retransmitted data packet is NACK two or more times consecutively, the NACK acknowledgment signal will be returned to the sending end.

7. The PCIe proactive retransmission method based on GRU neural network prediction according to claim 1, characterized in that, The retransmission controller is further configured to determine, when it is determined that there is an anomaly in the data link, the data packet in the data packet queue that needs to be transmitted from the data link is the target retransmission data packet; If the data link meets one of the following conditions, then the data link is determined to be abnormal: The change in any physical layer parameter in the data link is greater than the corresponding preset change threshold. The jitter magnitude of the data link is greater than a preset jitter magnitude threshold; The DFE tap corresponding to the data link is saturated.

8. The PCIe proactive retransmission method based on GRU neural network prediction according to claim 1, characterized in that, The retransmission controller is further configured to determine whether an equivalent link exists on the data link corresponding to the target retransmission data packet; if so, the retransmission controller controls the target retransmission data packet to be retransmitted through the equivalent link. And / or, For each sending end, there are equivalent first and second links. The first link is configured for data transmission in non-retransmission mode, and the second link is configured for data transmission in retransmission mode.

9. The PCIe proactive retransmission method based on GRU neural network prediction according to claim 1, characterized in that, Performing window aggregation includes calculating one or more of the following: mean, variance, quantiles, and extreme values; and / or, It also includes performing window aggregation on the quality indicators at a preset inference cycle and normalizing the quality indicators after window aggregation to obtain a first feature vector; and / or, The first feature vector is configured to be stored in a preset telemetry first-in-first-out buffer, independent of other buffers; and / or, The first index includes a packet header, sequence number, data pointer, and packet status information corresponding to each of the pre-retransmission data packets; and / or, The second index includes a header and a data pointer corresponding to each of the said target retransmission data packets.

10. A switch chip suitable for active retransmission, characterized by, It includes a telemetry acquisition module, a data processing module, a GRU prediction engine, a retransmission controller, and a retransmission buffer; The telemetry acquisition module is configured to collect quality indicators of each data link. The data processing module is configured to perform window aggregation on the quality indicators at a preset inference period to obtain a first feature vector. The quality indicators include multiple of the following: bit error rate, adaptive tap coefficient, carrier frequency offset, and clock offset. A GRU recurrent neural network is used to predict the probability of a link error occurring in the data link within a future target time window based on the first feature vector corresponding to one or more prior time windows. If the probability of a link error is greater than a preset first failure rate threshold for N consecutive times, and N is not less than a preset first number threshold, then the data packets transmitted through the data link within the future target time window are determined to be pre-retransmission data packets. The retransmission risk probability corresponding to the pre-retransmission data packets is higher than the retransmission risk probability corresponding to non-pre-retransmission data packets. The retransmission controller is configured to perform a pre-retransmission operation and a fast retransmission operation. The pre-retransmission operation includes: locking the storage location of the pre-retransmission data packet in the retransmission buffer, and constructing a preprocessing table, the preprocessing table including a first index corresponding to each pre-retransmission data packet, the first index including a data pointer for indicating the storage location of the pre-retransmission data packet, and obtaining the acknowledgment signal corresponding to each pre-retransmission data packet. The fast retransmission operation includes: in response to determining that the acknowledgment signal corresponding to the pre-retransmission data packet is NACK, the retransmission controller determines that the pre-retransmission data packet is a target retransmission data packet and constructs a fast retransmission queue, the fast retransmission queue including a second index corresponding to each target retransmission data packet, the second index including the data pointer; the retransmission controller determines the storage location of the target retransmission data packet according to the second index and retransmits it.

11. A communication system suitable for active retransmission, characterized in that, The device includes a switching chip and multiple devices to be communicated with, each of which is electrically connected to the switching chip. The switching chip performs communication between the multiple devices based on the PCIe active retransmission method based on GRU neural network prediction as described in claim 1.

Images