An abnormality detection method for a high-speed communication protocol of a computing power chip

Abstract

This invention relates to the field of communication protocol monitoring technology, specifically to an anomaly detection method for high-speed communication protocols of computing chips. The method includes: acquiring link status records, data unit records, bit error monitoring records, and channel training sequence records of the high-speed digital communication port of the computing chip, and grouping them into synchronization windows according to digital transmission timestamps; obtaining violation events based on the link status records and data unit records, obtaining channel transmission defect events based on the bit error monitoring records, and obtaining channel skew warning events based on the channel training sequence records; then constructing each anomaly event into anomaly event item, establishing a fault propagation relationship diagram based on the occurrence sequence, physical channel, data unit, and transmission layer relationships, and generating a communication link diagnostic report.

Description

Technical Field

[0001] This invention relates to the field of communication protocol monitoring technology, specifically to a method for detecting anomalies in high-speed communication protocols of computing chips. Background Technology

[0002] With the continuous increase in the demand for inter-chip interconnect bandwidth in high-speed digital communication, the Chiplet heterogeneous integration architecture has been widely used. In this architecture, dies with different functions transmit digital information through high-speed serial interconnect protocols such as UCIe, PCIe, and CXL. The single-channel transmission rate has reached tens of GT / s, and the reliability of the communication link directly affects the efficiency of digital information transmission.

[0003] These protocols typically include protocol layers corresponding to physical transport, link control, and transaction processing. The physical transport layer is responsible for signal encoding and bitstream transmission and reception; the link control layer is responsible for link training, flow control, and transmission synchronization; and the transaction layer is responsible for data unit routing and field processing. Degradation of physical channel quality may trigger link state anomalies or retraining, further causing data unit transmission anomalies. Existing monitoring schemes are usually deployed independently according to protocol layers, with a lack of correlation between anomaly records between layers, making it difficult to reconstruct the triggering location and propagation path of cross-layer anomalies.

[0004] Regarding physical channel bit error monitoring, existing schemes mainly calculate the bit error rate within a single moment or time interval, lacking sufficient analysis of the distribution characteristics of bit errors along the time and channel axes, making it difficult to accurately distinguish between continuous bit errors in a single channel, synchronous bit errors in multiple channels, and periodic bit errors. Regarding parallel channel timing offset monitoring, existing schemes typically only compare the current offset with the static tolerance, lacking sufficient analysis of the offset's changing trend over time, making it difficult to detect gradual anomalies in channel timing in a timely manner.

[0005] To address this, an anomaly detection method for high-speed communication protocols of computing chips is proposed. Summary of the Invention

[0006] This invention provides an anomaly detection method for high-speed communication protocols of computing chips. By synchronously collecting and hierarchically processing multi-source protocol data, a fault propagation relationship graph between abnormal events is constructed, and root cause candidates are screened according to the graph topology, thereby realizing the detection and assisted localization of anomalies in the high-speed communication protocols of computing chips.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting anomalies in a high-speed communication protocol of a computing chip includes: The link status records, data unit records, error monitoring records of each physical channel, and channel training sequence records of the high-speed digital communication port of the computing chip are acquired and merged into a synchronization window according to the digital transmission timestamp. Based on the link status record and data unit record, the link status violation event and the data unit violation event are obtained; The error monitoring records are aggregated using the physical channel as the row and the time sub-windows within the synchronization window as the column to obtain the channel error distribution table; the channel transmission defect event is obtained by matching the channel error distribution table with the preset transmission defect judgment rules. Based on the channel training sequence record, calculate the transmission delay difference of each physical channel relative to the reference physical channel, and form a channel skew sequence in chronological order; calculate the static skew amount and dynamic skew change rate based on the channel skew sequence, compare them with the preset link transmission tolerance, and obtain the channel skew warning event. Link status violation events, data unit violation events, channel transmission defect events, and channel skew warning events within the observation time window are constructed as abnormal event items. These are then correlated according to their occurrence sequence, physical channel correlation, data unit correlation, and transmission level correlation to obtain a fault propagation relationship diagram. A communication link diagnostic report is then generated based on this diagram.

[0008] Preferably, the synchronization window division includes: link status records corresponding to the link control layer, data unit records corresponding to the transaction layer, and error monitoring records and channel training sequence records corresponding to the physical transmission layer; dividing the four types of records into segments with a fixed time length; merging records whose timestamps fall within the same partition into the same synchronization window; and using null values ​​to fill in missing timestamps for records.

[0009] Preferably, obtaining link state violation events and data unit violation events includes: comparing the state transition sequence in the link state record with a preset legal state transition table hop by hop; for state transitions not in the legal state transition table, recording the predecessor state, successor state, and timestamp to obtain link state violation events; verifying the field values ​​of each data unit in the data unit record with the field constraints defined by the protocol specification field by field; for data units that violate field constraints and / or transaction sequence order, recording their data unit identifier, violation fields, and timestamp to obtain data unit violation events.

[0010] Preferably, obtaining the channel error distribution table includes: dividing the synchronization window into several time sub-windows with a fixed time step; constructing a two-dimensional table with the physical channel as the row index and the time sub-window number as the column index; assigning the error monitoring records of each physical channel to the corresponding time sub-window according to the timestamp, counting the number of error entries in each time sub-window and filling it into the corresponding table entry; setting the table entry with no error monitoring records to zero, thus obtaining the channel error distribution table.

[0011] Preferably, obtaining the channel transmission defect event includes: calculating the length of a continuous non-zero sub-window along the time axis of the channel error distribution table to obtain the burst continuity feature; calculating the Pearson correlation coefficient between the sequence of error entries formed by adjacent physical channels within each time sub-window along the channel axis of the channel error distribution table to obtain the channel synchronization correlation feature; calculating the normalized autocorrelation function of the number of error entries globally on the channel error distribution table, extracting the lag order corresponding to the first effective autocorrelation peak to obtain the periodic oscillation feature; and using the burst continuity feature, the channel synchronization correlation feature, and the periodic oscillation feature as error distribution judgment quantities, matching them with preset transmission defect judgment rules to obtain the channel transmission defect event carrying the defect mode type.

[0012] Preferably, obtaining the channel skew warning event includes: the preset link transmission tolerance includes a static tolerance and a dynamic tolerance; arranging the transmission delay differences of each physical channel within a continuous synchronization window in chronological order to obtain a channel skew sequence; taking the delay difference corresponding to the current synchronization window of the channel skew sequence as the static skew amount; performing linear fitting on the channel skew sequence using a sliding window of a preset length, and taking the slope of the fitted line as the dynamic skew change rate; comparing the static skew amount with the static tolerance, and comparing the dynamic skew change rate with the dynamic tolerance, and generating a channel skew warning event if either exceeds the limit.

[0013] Preferably, obtaining the fault propagation graph includes: constructing abnormal event items based on timestamps, protocol layers, event attributes corresponding to the event type, and physical channel identifiers or data unit identifiers carried by the source records for link state violation events, data unit violation events, channel transmission defect events, and channel skew warning events; using physical channel identifiers or data unit identifiers as the association basis, pointing abnormal event items carrying the same valid identifier and with an earlier timestamp to the next abnormal event item with the nearest timestamp; establishing directed associations from physical transport layer abnormal event items to link control layer abnormal event items based on the same physical channel identifier, and establishing directed associations from link control layer abnormal event items to transaction layer abnormal event items based on the same data unit identifier, wherein the start timestamp is not later than the end timestamp; abnormal event items without corresponding valid identifiers do not participate in this type of association; when the same abnormal event item satisfies multiple association conditions, all corresponding associations are retained to obtain the fault propagation graph.

[0014] Preferably, generating a communication link diagnostic report based on the fault propagation graph includes: identifying abnormal event items with zero in-degree and non-zero out-degree in the fault propagation graph as root cause candidates; when there is only one root cause candidate, it is identified as the highest priority root cause candidate; when there are multiple root cause candidates, they are arranged in descending order of their out-degree, and the candidate with the largest out-degree is identified as the highest priority root cause candidate; when there are multiple candidates with the same largest out-degree, all of the candidates are identified as the highest priority root cause candidates; and writing the protocol layer to which the highest priority root cause candidate belongs, the associated path reachable from the candidate, and the physical channel identifier or data unit identifier it carries into the communication link diagnostic report.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention merges link status records, data unit records, bit error monitoring records, and channel training sequence records into a synchronization window according to digital transmission timestamps, enabling communication records generated under different transmission layers to be processed under a unified time reference. This facilitates subsequent correlation analysis of abnormal events at the link control layer, transaction layer, and physical transmission layer.

[0016] 2. This invention constructs a channel error distribution table and obtains channel transmission defect events by combining it with preset transmission defect judgment rules. At the same time, it calculates the channel skew sequence, static skew amount and dynamic skew change rate based on the channel training sequence record. It can reflect the transmission status of the physical channel from two aspects: error distribution and channel delay change, which facilitates the hierarchical judgment of channel transmission anomalies.

[0017] 3. This invention constructs link status violation events, data unit violation events, channel transmission defect events, and channel skew warning events as abnormal event items, and establishes a fault propagation relationship diagram based on the occurrence sequence, physical channel correlation, data unit correlation, and transmission level relationship. It can present the correlation path between abnormal events in the communication link diagnostic report, providing a reference for anomaly investigation of high-speed communication protocols of computing power chips. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of an anomaly detection method for a high-speed communication protocol of a computing chip according to the present invention. Figure 2 This is a schematic diagram of the detection process of the present invention; Figure 3 This is a schematic diagram illustrating the generation of mapping and diagnostic reports according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: The link status records, data unit records, error monitoring records of each physical channel, and channel training sequence records of the high-speed digital communication port of the computing chip are acquired and merged into a synchronization window according to the digital transmission timestamp. Based on the link status record and data unit record, the link status violation event and the data unit violation event are obtained; The error monitoring records are aggregated using the physical channel as the row and the time sub-windows within the synchronization window as the column to obtain the channel error distribution table; the channel transmission defect event is obtained by matching the channel error distribution table with the preset transmission defect judgment rules. Based on the channel training sequence record, calculate the transmission delay difference of each physical channel relative to the reference physical channel, and form a channel skew sequence in chronological order; calculate the static skew amount and dynamic skew change rate based on the channel skew sequence, compare them with the preset link transmission tolerance, and obtain the channel skew warning event. Link status violation events, data unit violation events, channel transmission defect events, and channel skew warning events within the observation time window are constructed as abnormal event items. These are then correlated according to their occurrence sequence, physical channel correlation, data unit correlation, and transmission level correlation to obtain a fault propagation relationship diagram. A communication link diagnostic report is then generated based on this diagram.

[0021] The observation time window referred to in this application is the scope of abnormal event association processing composed of multiple consecutive synchronization windows in chronological order. It is used to limit the abnormal event items included in the current fault propagation relationship diagram without changing the segmentation rules of the synchronization window and time sub-window.

[0022] Furthermore, the synchronization window division includes: link status records corresponding to the link control layer, data unit records corresponding to the transaction layer, and error monitoring records and channel training sequence records corresponding to the physical transmission layer; the four types of records are divided with a fixed time length; records whose timestamps fall within the same partition are merged into the same synchronization window; and records with missing timestamps are filled with null values.

[0023] Specifically, during the initialization phase of the monitoring platform for the high-speed digital communication port of the computing chip, the system establishes a static mapping relationship between four types of records and protocol layers based on the pre-selected target high-speed communication protocol and its version. The target high-speed communication protocol referred to in this application is a high-speed serial interconnection protocol and its version pre-selected during deployment; the legal state transition table, field constraints, and channel training sequence parsing rules all originate from the protocol specification of the corresponding version of the target high-speed communication protocol. Link state records originate from the register snapshots of the link training and state machine, corresponding to the link control layer; data unit records originate from the parsing results of the data units of the transaction layer of the target high-speed communication protocol, corresponding to the transaction layer; bit error monitoring records originate from the valid bit error statistics records output by the bit error monitoring modules of each physical channel according to a fixed statistical caliber, corresponding to the physical transmission layer; and channel training sequence records originate from the link initialization or retraining process specified by the target high-speed communication protocol, corresponding to the physical transmission layer. When the above four types of records are generated at their respective collection points, they are all calibrated with digital transmission timestamps according to the global reference clock uniformly configured by the monitoring platform, making records from different protocol layers comparable in terms of timestamp dimensions.

[0024] After completing the hierarchical mapping, the monitoring platform divides the global digital transmission timeline into equal intervals with a fixed time length, forming a continuous and non-overlapping sequence of time intervals, resulting in several sequentially numbered intervals. The four types of records share the same global digital transmission timeline, the same time segmentation starting point, and the same segmentation boundary; no separate intervals are set for each type of record. The fixed time length is determined based on the reporting cycle of the four types of records and is used to form a shared window index. The system extracts the digital transmission timestamps carried by each of the four types of records, merging records whose timestamps fall within the same partition into the same synchronization window. When no valid record of a certain type is obtained within the synchronization window, a null value is set for that record type; this null value does not participate in the generation of the corresponding abnormal event.

[0025] To address situations where hardware probe sampling funnel effects or bus congestion lead to the loss of some record timestamps, for records received by the monitoring platform during the current synchronization window but lacking a timestamp field, the system sets a null value at the corresponding record type position within that synchronization window and marks the record as invalid. Null values ​​and invalid records do not participate in the generation of abnormal event items. For late records arriving after the synchronization window has closed but carrying valid timestamps that fall within the closed synchronization window, the system does not change the already output real-time abnormal event items and fault propagation relationship diagram. Instead, it stores the late record in the late record cache and records it as window boundary evidence in subsequent offline communication link diagnostic reports. Late records do not participate again in the real-time calculation of the closed synchronization window.

[0026] This method achieves window-level time slice alignment between microsecond-level sampled data from the physical transport layer and millisecond-level reported data from the link control layer under the premise of a unified time base by dividing the synchronization window based on a unified global reference clock. This provides a reliable time slice basis for subsequent cross-layer correlation analysis.

[0027] Furthermore, obtaining link state violation events and data unit violation events includes: comparing the state transition sequence in the link state record with a pre-set legal state transition table hop by hop; for state transitions not in the legal state transition table, recording the predecessor state, successor state, and timestamp to obtain link state violation events; verifying the field values ​​of each data unit in the data unit record with the field constraints defined by the protocol specification field by field; for data units that violate field constraints and / or transaction sequence order, recording their data unit identifier, violation fields, and timestamp to obtain data unit violation events.

[0028] Specifically, in detecting link state violation events, the monitoring platform pre-loads a legal state transition table into memory. This table is compiled based on the specification text regarding link training and state machines in the pre-selected target high-speed communication protocol and its version, using state nodes as row and column indices to record all permitted transition relationships. During the processing of each synchronization window, the detection engine extracts the actual state transition sequence from the link state record and connects it with the state recorded at the end of the previous synchronization window as the preceding state for the first transition comparison in this synchronization window. Each state transition in the connected sequence is then compared hop-by-hop with the legal state transition table, ensuring that state transitions crossing synchronization window boundaries are also included in the comparison.

[0029] Once a state transition not in the valid state transition table is detected, such as a link transitioning directly from a low-power sleep state to an active state without undergoing standard recovery state negotiation, the system immediately triggers an alarm and records the predecessor state, successor state, and timestamp of the transition. These three attributes are encapsulated as a standardized link state violation event, which serves as the input node attributes for subsequently constructing the fault propagation relationship graph.

[0030] In the detection of data unit violation events, the detection engine verifies each data unit in the data unit record field by field, comparing each field value with the field constraints defined by the protocol specification. The field constraints cover the legal range of the requester identifier, memory address boundaries, upper limit of payload length, and routing label format, etc. For example, under the memory access semantics of the target protocol, the host physical address must strictly fall within the legal memory range declared by the device; exceeding the boundary constitutes a violation.

[0031] After completing field-level verification, the detection engine further performs transaction sequence order verification on the data unit sequence: Based on the initial and upper limits of the flow control credit defined in the protocol specification, it verifies whether the count value after credit consumption and release exceeds the legal range allowed by the specification; if it does, a credit imbalance is determined. According to the data unit sequence number increment and legal wraparound rules defined in the protocol specification, it verifies whether the change in the data unit sequence number conforms to the increment and wraparound rules; if not, it is determined as an abnormal sequence number jump or duplication. For any data unit that violates field constraints, has a credit imbalance, or has an abnormal sequence number, the system extracts and records its data unit identifier, the specific name of the violating field, and the timestamp of the occurrence, encapsulating it as a data unit violation event.

[0032] This method accurately covers the compliance detection blind spots of the link control layer and transaction layer by comparing each hop of the pre-set legal state transition table and verifying each field of protocol field constraints. The generated violation events carry complete location attributes, providing traceable input for the subsequent construction of nodes in the graph model.

[0033] Further, obtaining the channel error distribution table includes: dividing the synchronization window into several time sub-windows with a fixed time step; constructing a two-dimensional table with the physical channel as the row index and the time sub-window number as the column index; assigning the error monitoring records of each physical channel to the corresponding time sub-window according to the timestamp, counting the number of error entries in each time sub-window and filling it into the corresponding table entry; setting the table entry with no error monitoring records to zero, thus obtaining the channel error distribution table.

[0034] Specifically, since bit error rate statistics at a single time point are insufficient to effectively distinguish different physical causes in high-speed parallel link scenarios, this method organizes bit error monitoring records into a two-dimensional structure to characterize the distribution of bit errors along both the time and channel axes. In processing each synchronization window, the system first further divides the current synchronization window into several time sub-windows with a fixed time step, and each sub-window is numbered sequentially. The fixed time step is equal to the fixed statistical period of the bit error monitoring module, and the duration of the synchronization window is an integer multiple of the fixed time step. The fixed statistical period is derived from the hardware configuration of the bit error monitoring module and is measured in seconds. The total number of time sub-windows is obtained by dividing the duration of the synchronization window by the fixed time step, and is a dimensionless positive integer, not set separately by the operator.

[0035] Each time sub-window is divided using a method that includes a start boundary but excludes an end boundary. If the timestamp of a bit error monitoring record matches the start boundary of a time sub-window, the record is assigned to the time sub-window starting from that boundary; if the timestamp matches the end boundary of a time sub-window, the record is assigned to the time sub-window starting from the next boundary. Each bit error monitoring record is assigned to only one time sub-window, and duplicates are not counted. Time sub-windows are only used for aggregating bit error monitoring records and do not split or truncate data unit records.

[0036] Subsequently, the system constructs a two-dimensional table structure in memory with the number of rows equal to the total number of physical channels and the number of columns equal to the total number of time sub-windows, using all physical channel identifiers involved in the current synchronization window as row indexes and the sequence number of each time sub-window as column indexes.

[0037] During the table filling phase, the system iterates through the error monitoring records of each physical channel, determines the time sub-window number to which each record belongs based on the timestamp carried by the record, and fills the number of error entries corresponding to that record into the corresponding entry of the two-dimensional table. When the error monitoring module is in an effective monitoring state within the corresponding time sub-window and no error records are generated, the corresponding entry is set to zero; when the error monitoring module is in an ineffective monitoring state, the record timestamp is missing, or the record is marked as invalid, the corresponding entry is set to a null value, and the null value is not included in subsequent feature calculations. After completing the table filling, a channel error distribution table is obtained, which serves as the input matrix for subsequent feature extraction.

[0038] This method transforms spurious transient error logs into a structured two-dimensional distribution table, enabling subsequent feature extraction to perform directional statistical analysis along the time axis and channel axis respectively, laying a data foundation for refined identification of error patterns.

[0039] Further, obtaining the channel transmission defect event includes: calculating the length of a continuous non-zero sub-window along the time axis of the channel error distribution table to obtain the burst continuity feature; calculating the Pearson correlation coefficient between the sequence of error entries formed by adjacent physical channels within each time sub-window along the channel axis of the channel error distribution table to obtain the channel synchronization correlation feature; globally calculating the normalized autocorrelation function of the number of error entries in the channel error distribution table, extracting the lag order corresponding to the first effective autocorrelation peak to obtain the periodic oscillation feature; and using the burst continuity feature, the channel synchronization correlation feature, and the periodic oscillation feature as error distribution judgment quantities, matching them with preset transmission defect judgment rules to obtain the channel transmission defect event carrying the defect mode type.

[0040] Specifically, in the extraction of burst continuity features, the system scans line by line along the time axis of the channel error distribution table, and counts the length of the continuous non-zero sub-window of the number of error entries in the row of each physical channel. That is, it counts the maximum number of continuous sub-windows in which errors continue to occur in the time dimension. For example, if a physical channel has error entries recorded in 6 consecutive time sub-windows, the burst continuity feature value is 6. A larger feature value indicates that the corresponding physical channel has continuous bit error phenomena in time. It can be used as one of the criteria for judging defect modes such as transient interference, channel attenuation, connection anomaly or clock recovery anomaly, but is not used alone to confirm specific physical causes.

[0041] In the extraction of channel synchronization correlation features, the system extracts the sequence of error entries within an effective time sub-window for each pair of adjacent physical channels along the channel axis of the channel error distribution table. The effective time sub-window is a time sub-window in which at least one adjacent physical channel has a non-zero error rate. Let the sequence of error entries for two adjacent physical channels within each effective time sub-window be as follows: to and to The Pearson correlation coefficient is calculated using the following formula: ; in, The number of sub-windows with valid time; The valid time sub-window sequence number; and These are two adjacent physical channels at the 1st... The number of error entries within each valid time sub-window is derived from the channel error distribution table and is a dimensionless count value. and These are the arithmetic mean of the corresponding sequence of bit error counts; is the Pearson correlation coefficient, a dimensionless quantity, with a value ranging from -1 to 1.

[0042] When the number of effective time sub-windows is less than three, or the variance of any error rate sequence is zero, the Pearson correlation coefficient for that channel pair is not calculated, and the channel synchronization correlation feature is marked as invalid. A number of effective time sub-windows of three only satisfies the minimum calculation condition for the Pearson correlation coefficient and is not considered proof of statistical significance. The calibration data and online detection data use the same effective time sub-window selection rules.

[0043] The 99th percentile of the Pearson correlation coefficient obtained under the crosstalk-free consistency test conditions specified in the target high-speed communication protocol for the same type of link is determined as the correlation judgment threshold. When the current Pearson correlation coefficient is greater than this judgment threshold and falls within the value range of the channel synchronization correlation feature corresponding to the crosstalk defect mode, the channel pair is marked as a crosstalk candidate channel pair with a statistical correlation of synchronization error. The 99th percentile is derived from the normal operating condition test data and is not set by the operator. The channel synchronization correlation feature is used to characterize the degree of linear synchronization change between the sequence of error entries of adjacent physical channels, and participates in defect mode matching together with the burst continuity feature and the periodic oscillation feature, but is not used alone as the basis for confirming the existence of physical crosstalk or causal relationship between adjacent physical channels.

[0044] In the extraction of periodic oscillation characteristics, the system sums the number of bit error entries for all physical channels within each time sub-window to obtain a global bit error total sequence arranged according to the time sub-window number. Let the global bit error total sequence contain... For each data point, the lag order is... The corresponding normalized autocorrelation function is: ; in, The time sub-window sequence number; For the first The total number of bit error entries for all physical channels within a time sub-window is derived from the channel bit error distribution table and is a dimensionless count. For the first The total number of error entries for all physical channels within a time sub-window is a dimensionless count. for The arithmetic mean of the total number of global error entries corresponding to each time sub-window is a dimensionless quantity. The total number of time sub-windows actually obtained by dividing the synchronization window is a dimensionless positive integer. is the lag order, which is a dimensionless positive integer; Lag order The corresponding normalized autocorrelation function value is dimensionless. When the denominator of the formula is zero, it indicates that the total global bit error rate sequence does not change, and periodic oscillation characteristics are not calculated. The system only... Greater than or equal to 1 and less than or equal to The search for peak values ​​is conducted within a specified range, without considering the beginning and end of the search range as local maxima. Peak values ​​higher than the adjacent autocorrelation function values ​​and higher than the candidate peak selection threshold are considered. The first position is determined as the lag order corresponding to the first effective autocorrelation peak. Among them, This represents the total number of time sub-windows actually obtained by dividing the synchronization window. This is a fixed initial screening threshold determined based on the approximate 95% confidence range of the white noise autocorrelation coefficient. It is only used for the preliminary screening of candidate autocorrelation peaks and is not used as the sole basis for confirming periodic transmission defects.

[0045] When the periodic oscillation characteristic triggers the defect determination, the physical channel identifier with the highest number of bit error entries in the current synchronization window is written into the channel transmission defect event; when multiple physical channels have the same number of bit error entries, the corresponding physical channel identifiers are written together.

[0046] After extracting the three types of features, the system uses burst continuity features, channel synchronization correlation features, and periodic oscillation features to form the error distribution judgment metric. The preset transmission defect judgment rules include normal operating condition judgment intervals, single-channel continuous error defect pattern judgment intervals, multi-channel synchronization error defect pattern judgment intervals, and periodic error defect pattern judgment intervals. Each judgment interval is determined by calibration data obtained from communication links of the same model under the same target protocol version, the same power consumption state, the same fixed time step, and the same synchronization window length. The calibration data for the preset transmission defect judgment rules comes from normal operating condition data collected from communication links of the same model under the consistency test conditions specified by the target high-speed communication protocol, as well as fault data collected under known fault injection conditions. The mean and standard deviation of the three features under normal operating conditions and various known fault conditions are calculated respectively, and the mean plus or minus three times the standard deviation is used as the fixed engineering judgment interval. The judgment interval is only used for defect pattern matching under the same model link, the same power consumption state, the same fixed time step, and the same synchronization window length, and is not used as statistical proof that the three features follow a normal distribution. The actual number of valid synchronization windows is recorded during calibration, without setting a minimum sample size.

[0047] When all three features fall within the value range corresponding to the same defect pattern, the defect pattern is determined to be matched. A logical AND relationship is used between the three features. When multiple defect patterns are matched simultaneously, the standardized distance is calculated using the following formula: ; in, For feature serial number; The burst continuity characteristic originates from the maximum number of consecutive non-zero time sub-windows in terms of the number of bit errors; The channel synchronization correlation characteristic is derived from the Pearson correlation coefficient of the sequence of bit error rates of adjacent physical channels. The characteristic of periodic oscillation originates from the autocorrelation peak lag order of the total global bit error rate sequence, and all three characteristics are dimensionless quantities. This is the defect mode number; and The first Class Known Defect Mode, First The mean and standard deviation of the features are derived from fault injection tests or protocol conformance test data of the same type of link; The standardized distance is dimensionless. No channel transmission defect event is generated when all three features fall within the value range corresponding to normal operating conditions. When at least one feature exceeds the value range corresponding to normal operating conditions, it is then matched against the value ranges corresponding to various known defect patterns. When multiple defect patterns are matched simultaneously, the defect pattern with the smallest standardized distance is determined as the defect pattern type. If no known defect pattern is matched, it is marked as an unclassified transmission defect. When the standard deviation of a feature is zero, if the current feature value is different from the corresponding mean, the defect pattern is excluded; if the current feature value is the same as the corresponding mean, the feature is not included in the standardized distance calculation. When none of the three features have a calculable standard deviation, or when the standardized distances of multiple defect patterns are tied for smallest, it is marked as an unclassified transmission defect.

[0048] This method classifies the error distribution pattern of physical channels by jointly extracting three types of statistical features and matching rules, and outputs defect mode types or unclassified transmission defects, providing input for physical transport layer anomaly correlation in the fault propagation relationship diagram.

[0049] Further, obtaining the channel skew warning event includes: the preset link transmission tolerance includes a static tolerance and a dynamic tolerance; arranging the transmission delay differences of each physical channel within a continuous synchronization window in chronological order to obtain a channel skew sequence; taking the delay difference corresponding to the current synchronization window of the channel skew sequence as the static skew amount; performing linear fitting on the channel skew sequence using a sliding window of a preset length, and taking the slope of the fitted line as the dynamic skew change rate; comparing the static skew amount with the static tolerance, and comparing the dynamic skew change rate with the dynamic tolerance, and generating a channel skew warning event if either exceeds the limit.

[0050] Specifically, the preset link transmission tolerances include static tolerances and dynamic tolerances. Static tolerances are derived from the maximum permissible relative delay difference of the physical channel specified in the corresponding version of the target high-speed communication protocol; dynamic tolerances are derived from the maximum permissible skew rate of change specified in the hardware timing design specifications of the same type of communication port. The monitoring platform reads the corresponding static and dynamic tolerances according to the current target protocol version and communication port model. Both are derived from existing protocol specifications or hardware timing design specifications and are not set by the operator.

[0051] In constructing the channel skew sequence, the system utilizes channel training sequence records and calculates the transmission delay difference of each physical channel relative to a reference physical channel using high-precision timestamps. The reference physical channel is preferentially selected from the reference clock channel or gated channel specified in the corresponding version of the target high-speed communication protocol. If the target high-speed communication protocol does not specify a dedicated reference channel, a physical channel with a fixed number is selected as the reference physical channel during the monitoring platform initialization phase and remains unchanged in subsequent calculations. The transmission delay difference of each physical channel is a relative delay difference relative to the reference physical channel and does not directly represent the absolute transmission delay of the measured physical channel. The channel training sequence record is generated during the link initialization phase or the link retraining process defined in the protocol specification. For synchronization windows where link initialization or retraining has not occurred, the transmission delay difference calculated from the most recent valid channel training sequence record is only used to maintain the current static skew state and compare static tolerances, and this value is marked as a reused value. The reused value is not used as a new measurement point in the calculation of the dynamic skew change rate. The system arranges the effective transmission delay difference obtained from each physical channel according to its original measurement timestamp, and marks the used values ​​and effective measurement values ​​respectively to obtain the channel skew sequence.

[0052] In the calculation of static skew, when a valid channel training sequence record exists within the current synchronization window, the system uses the delay difference corresponding to that record as the current measured static skew value. When no valid channel training sequence record exists within the current synchronization window, the delay difference corresponding to the most recent valid record is used as the most recent valid static skew value. The system compares the absolute value of the current measured static skew value or the most recent valid static skew value with the static tolerance; if it exceeds the static tolerance, it is marked as exceeding the limit. When using the value, the usage status is recorded in the channel skew warning event.

[0053] In calculating the dynamic skew rate of change, the system selects the three most recent effective channel training sequence records as measurement points within a sliding window and performs least-squares linear fitting based on the original timestamps of each effective record. Three effective measurement points represent the minimum fixed number of points required to calculate the linear trend and retain the fitting residuals. Placeholder values ​​derived from historical measurement results are not included in the linear fitting; if there are fewer than three effective measurement points, only static skew determination is performed, and the dynamic skew rate of change is not output. Let the measurement times corresponding to the three effective records be... , and The corresponding transmission delay differences are respectively , and The slope of the fitted line is: ; in, Record the sequence number for the effective channel training sequence; For the first The time, calculated from the original timestamp of each valid record, is derived from the global reference clock of the monitoring platform and is in seconds. For the first The transmission delay difference between the current physical channel and the reference physical channel in each valid record is derived from the channel training sequence record and is expressed in seconds. and These are the arithmetic mean of three time values ​​and three transmission delay differences, respectively; The dynamic skew rate of change is measured in seconds per second (s / s), or picoseconds per second. This dynamic skew rate of change characterizes the skew trend among the three most recent effective channel training measurement points, but does not represent the continuous real-time change in skew between two adjacent effective training measurements.

[0054] Ultimately, the system performs dual limit checks on both static skew and dynamic skew rate of change. The absolute value of the static skew is compared to the static tolerance; if the dynamic skew rate of change has been calculated, its absolute value is compared to the dynamic tolerance. Exceeding either limit generates a channel skew warning event. The generated event carries the identifier of the measured physical channel, the identifier of the reference physical channel, the current timestamp, the static skew, and its current measured or in-use status; if the dynamic skew rate of change has been calculated, it is also written into the event. This warning indicates an abnormal skew of the measured physical channel relative to the reference physical channel and is not used alone as confirmation of an absolute fault in the measured physical channel.

[0055] This method uses a dual tolerance judgment of static skew amount and dynamic skew change rate to capture timing degradation trends in advance before the skew amount exceeds the protocol chain breakage threshold, transforming traditional post-fault troubleshooting into proactive early warning diagnosis.

[0056] Further, obtaining the fault propagation graph includes: constructing abnormal event items based on timestamps, protocol layers, event attributes corresponding to the event type, and physical channel identifiers or data unit identifiers carried by the source records for link state violation events, data unit violation events, channel transmission defect events, and channel skew warning events; using physical channel identifiers or data unit identifiers as the association basis, pointing abnormal event items carrying the same valid identifier and with an earlier timestamp to the next abnormal event item with the nearest timestamp; establishing directed associations from physical transport layer abnormal event items to link control layer abnormal event items based on the same physical channel identifier, and establishing directed associations from link control layer abnormal event items to transaction layer abnormal event items based on the same data unit identifier, wherein the start timestamp is not later than the end timestamp; abnormal event items without corresponding valid identifiers do not participate in this type of association; when the same abnormal event item satisfies multiple association conditions, all corresponding associations are retained to obtain the fault propagation graph.

[0057] Specifically, during the construction phase of abnormal event items, the system instantiates the four types of detected events into graph network nodes. Each node carries its protocol layer, timestamp, and location attribute corresponding to the event type. Link state violation event nodes carry a predecessor state identifier, a successor state identifier, a timestamp, a link control layer flag, and the physical channel identifier carried by its source record. The data unit identifier is written to the node only if the source record of the link state violation event carries a data unit identifier that can be directly matched with a data unit record; otherwise, this attribute is set to empty and does not participate in data unit identifier association. Data unit violation event nodes carry a data unit identifier, violation field name, timestamp, and transaction layer flag, where the data unit identifier and violation field are derived from the data unit field verification result. For port-level link state records that only carry a port identifier and not a physical channel identifier, their physical channel identifier attribute is set to empty and does not participate in association based on the physical channel identifier. The determined physical channel identifier is written to the abnormal event item only if the source record directly carries a physical channel identifier, or if the physical channel identifier can be determined according to the fixed mapping relationship specified by the target high-speed communication protocol.

[0058] The channel transmission defect event node carries the physical channel row index (i.e., physical channel identifier), corresponding time sub-window sequence number, matched defect mode type, and physical transmission level label of the bit error distribution table. The physical channel row index comes from the row index definition used when constructing the two-dimensional table of the channel bit error distribution table, and the defect mode type comes from the output results after matching the burst continuity feature, channel synchronization correlation feature, and periodic oscillation feature with the preset transmission defect judgment rules. The channel skew warning event node carries the physical channel identifier, timestamp, calculated static skew amount and dynamic skew change rate, and physical transmission level label from the channel training sequence record. The static skew amount and dynamic skew change rate come from the calculation results obtained during the channel skew sequence calculation process.

[0059] During the directed association establishment phase, the system establishes directed edges between nodes based on four types of association rules. Using the physical channel identifier as the association basis, among abnormal event items under the same physical channel identifier, only the node with the earlier timestamp is pointed to the next node with the same physical channel identifier whose timestamp is closest to it (rather than pointing to all nodes with later timestamps), forming an association path where abnormal events on the same physical channel are arranged in chronological order; for example, a physical channel with timestamps... Channel transmission defect events occurred, in conjunction with The closest in time A link state violation event occurred ( Greater than If both carry the same physical channel identifier, then by Node to A directed edge is established for each node.

[0060] Based on the data unit identifier, among abnormal event items carrying the same data unit identifier, only the node with the earlier timestamp is pointed to the next node with the same data unit identifier closest to its timestamp, tracking multiple abnormal records triggered across time points when the same data unit flows through the system; based on the protocol layer and physical channel identifier, under the same physical channel identifier, a directed association is established with the physical transport layer abnormal event item as the starting point and the link control layer abnormal event item as the ending point, with the starting timestamp no later than the ending timestamp, forming an association path between the physical transport layer abnormal event and the link control layer abnormal event that occurs later; based on the protocol layer and data unit identifier, among event items carrying the same data unit identifier, a directed association is established with the link control layer abnormal event item as the starting point and the transaction layer abnormal event item as the ending point, with the starting timestamp no later than the ending timestamp, forming an association path between the link control layer abnormal event and the transaction layer abnormal event that occurs later.

[0061] When the same pair of anomalous event items simultaneously satisfy multiple association conditions, the system retains all corresponding directed association edges, enabling the fault propagation graph to fully preserve multi-dimensional information dependencies and ultimately obtain a fault propagation graph reflecting the cross-layer transmission path of anomalies. The construction scope of the fault propagation graph is not limited to anomalous event items within a single synchronization window, but is cumulatively constructed across multiple consecutive synchronization windows within a preset observation time window. The length of the observation time window is not less than the duration required to cover a complete link retraining cycle, ensuring that the observation period includes the complete fault evolution process. For historical anomalous event items and their associated edges whose timestamps exceed the range of this observation time window, the system archives them or removes them from the current fault propagation graph to prevent the graph structure from growing indefinitely as the system continues to operate.

[0062] This method, through the synergistic effect of four types of directed association rules, fully encodes the association relationships of time, physical channel, and protocol layer into a graph structure, enabling the fault propagation graph to present the temporal and identifier associations between abnormal events across protocol layers, and providing graph topology basis for root cause candidate screening.

[0063] Furthermore, the step of generating a communication link diagnostic report based on the fault propagation graph includes: in the fault propagation graph, identifying abnormal event items with zero in-degree and non-zero out-degree as root cause candidates; when there is only one root cause candidate, identifying it as the highest priority root cause candidate; when there are multiple root cause candidates, arranging them in descending order of their out-degree, identifying the candidate with the largest out-degree as the highest priority root cause candidate; when there are multiple candidates with the same largest out-degree, identifying all of them as the highest priority root cause candidates; and writing the protocol layer to which the highest priority root cause candidate belongs, the associated path reachable from the candidate, and the physical channel identifier or data unit identifier it carries into the communication link diagnostic report.

[0064] Specifically, in the root cause candidate screening, the system scans all anomalous event nodes in the fault propagation graph and calculates the in-degree and out-degree of each node. The in-degree is the number of directed edges pointing to that node, and the out-degree is the number of directed edges originating from that node. The system identifies nodes with an in-degree of zero and an out-degree of non-zero as root cause candidates. An in-degree of zero indicates that no upstream anomalous event pointing to that node has been detected under the current observation time window and current association rules; an out-degree of non-zero indicates that the node is associated with at least one anomalous event that occurred later in time. Root cause candidates are given priority for investigation but do not necessarily indicate that they have been confirmed as the root cause of the physical fault.

[0065] When there is only one root cause candidate, the system identifies it as the highest priority root cause candidate. When there are multiple root cause candidates, they are sorted in descending order of their out-degree values, and the candidate with the largest out-degree is identified as the highest priority root cause candidate. The out-degree represents the number of downstream abnormal event items associated with the candidate in the current fault propagation graph, and is not used alone as proof of a real physical causal relationship. When multiple candidates have the same maximum out-degree, all of them are identified as the highest priority root cause candidates and written into the communication link diagnostic report.

[0066] During the compilation of the diagnostic report, the system writes the protocol level to which the highest priority root cause candidate belongs into the report. Then, it traverses the directed association edges starting from the candidate in the fault propagation graph layer by layer, and writes the association paths reachable from the candidate and the corresponding abnormal event items into the report to present the association relationship between the highest priority root cause candidate and the downstream abnormal event items.

[0067] During the writing of location attributes, the system determines the attribute type carried by the highest priority root cause candidate. When the candidate carries a physical channel identifier, the physical channel identifier and its carried static skew amount or dynamic skew change rate are written to the report; when the candidate carries a data unit identifier, the data unit identifier and violation field information are written to the report. The above attributes are used to provide clues for troubleshooting hardware layer or protocol layer anomalies, and do not indicate that the candidate has been confirmed as the actual physical root cause.

[0068] This method prioritizes root cause candidates based on the in-degree, out-degree, and associated path of abnormal event items. The communication link diagnostic report outputs the protocol layer, associated path, and location attributes of the candidate items, providing clues for troubleshooting hardware or firmware layer anomalies.

[0069] Example 2: Furthermore, to address the issue that when the communication port of the computing chip frequently switches power consumption modes due to changes in workload, using a fixed-length synchronization window would lead to significant fluctuations in data processing efficiency, the synchronization window division further includes: obtaining the power consumption state machine identifier from the link status record; when the identifier indicates that the link is in a speed-down power-saving state or a non-full-speed wake-up state, multiplying the fixed time length by a dynamic expansion coefficient determined based on the current power consumption state and the record reporting cycle under full-speed operation state to obtain an elastic time span; and using the elastic time span to replace the fixed time length to segment the record to obtain a dynamic synchronization window.

[0070] Specifically, during the division of the synchronization window, the system pre-acquires the power consumption state machine identifier from the link state record. When the computing chip communication port is in full-speed operation, the system maintains the original fixed time length for segmenting various records. When the power consumption state machine identifier indicates that the link has entered a speed-down power-saving state or a non-full-speed wake-up state, the effective message transmission frequency drops sharply due to the clock gating mechanism. Continuing to use the original fixed time segmentation will generate a large number of invalid synchronization windows with empty values. At this time, the system reads the fixed record reporting period of the communication port monitoring module under the current power consumption state and the fixed record reporting period under the full-speed operation state. The fixed record reporting period under the current power consumption state is divided by the fixed record reporting period under the full-speed operation state, and the resulting ratio is rounded up to obtain the dynamic expansion coefficient; the fixed time length is multiplied by the dynamic expansion coefficient to obtain the elastic time span. Both fixed record reporting periods are derived from the hardware configuration of the communication port monitoring module for the corresponding power consumption state, and the unit is seconds; the dynamic expansion coefficient is a dimensionless positive integer, and the unit of the elastic time span is seconds. For different power consumption states, the corresponding record reporting period is read according to the power consumption state machine identifier, without setting a separate correction coefficient. Subsequently, the system uses the calculated elastic time span to replace the original fixed time length and re-divides the global digital transmission time axis to obtain a dynamic synchronization window. The dynamic synchronization window still uses the pre-configured fixed time step for the corresponding power consumption state to divide the time sub-window. The burst continuity characteristics, channel synchronization correlation characteristics, and periodic oscillation characteristics obtained under different power consumption states are only matched with the calibration intervals obtained under the same power consumption state, the same fixed time step, and the same dynamic synchronization window length, and are not directly compared with the calibration interval under the full-speed operating state; when there is no corresponding calibration interval, only the three characteristics are output, and the specific defect mode is not confirmed.

[0071] This method generates an elastic time span when the link enters a slowdown or low-power state by adjusting the dynamic expansion coefficient based on the power state machine identifier. This effectively avoids the problem of generating a large number of invalid empty matrices under an absolutely fixed time window. Under the premise of ensuring the uniformity of the time base, it helps to reduce the number of all-empty synchronization windows and the corresponding data processing overhead, and provides an adaptive time slicing basis for message integrity and cross-layer correlation analysis under low-rate conditions.

[0072] Further, after obtaining the periodic oscillation characteristics, the hysteresis order corresponding to the first effective autocorrelation peak is multiplied by a fixed time step to obtain the true period of the periodic fluctuation of the bit error rate; the reciprocal of the true period is taken to obtain the measured oscillation frequency. The reciprocal of the fixed time step is the sampling frequency of the total global bit error rate sequence. Only when the measured oscillation frequency is within the detectable range of less than half of the sampling frequency is it determined whether the measured oscillation frequency satisfies an integer multiple harmonic relationship with the power supply pulse width modulation frequency of the computing chip or the reference frequency of the adjacent serial bus; if it exceeds the detectable range, no harmonic relationship judgment is performed. The upper limit of half of the sampling frequency is derived from the Nyquist sampling criterion. When the frequency tolerance requirement is met, a suspected hardware crosstalk tag for the corresponding device is added to the channel transmission defect event.

[0073] Specifically, the larger of the measured oscillation frequency and the candidate interference source frequency is divided by the smaller value, and the resulting ratio is rounded to the nearest positive integer harmonic order; then the absolute difference between the larger frequency and the integer multiple of the smaller frequency is calculated, and the absolute difference is divided by the integer multiple of the smaller frequency to obtain the relative harmonic deviation.

[0074] The measured oscillation frequency is derived from the conversion result of the autocorrelation peak lag order and the fixed time step; the candidate interference source frequency is derived from the power controller datasheet or the reference clock configuration of the adjacent serial bus, both in Hertz; the positive integer harmonic orders and harmonic relative deviations are dimensionless quantities. When the harmonic relative deviation is not greater than the sum of the frequency relative tolerance specified in the candidate interference source datasheet and the frequency measurement relative error of the monitoring platform, a suspected hardware crosstalk tag for the corresponding device is added to the channel transmission defect event. The frequency relative tolerance and frequency measurement relative error are derived from existing equipment specifications and are not set by the operator.

[0075] This method compares the integer harmonic relationships between the measured oscillation frequency and the candidate interference source frequency, and can simultaneously cover situations where the measured oscillation frequency is higher or lower than the candidate interference source frequency. When both meet the frequency tolerance requirements, a suspected hardware crosstalk tag is added, providing frequency correlation basis for physical layer anomaly investigation in the communication link diagnostic report.

[0076] Furthermore, in response to the situation where avalanche-like anomalies occur in complex communication links, multiple concurrent failures can lead to the same number of downstream anomalies caused by multiple root cause candidates, resulting in ambiguity in the root cause location in the diagnostic report. The method of generating a communication link diagnostic report based on the fault propagation graph also includes: when there are multiple root cause candidates and their out-degrees are tied for the highest, calculating the topology depth weight of each tied candidate; the topology depth weight is the sum of the shortest path hops from the candidate to all reachable terminal abnormal nodes divided by the total number of corresponding terminal abnormal nodes; the candidate with the smallest topology depth weight is determined as the highest priority root cause candidate.

[0077] Specifically, when determining the highest priority root cause candidate based on the fault propagation graph, for cases where multiple candidate options have the largest out-degree in a multi-point concurrent failure scenario, the system introduces a topology depth weight for secondary evaluation. Starting with each root cause candidate with the largest out-degree, the system executes a shortest path tracing algorithm along directed edges until it reaches each terminal abnormal node with a zero out-degree. The system calculates the sum of the shortest path hops from the candidate to all reachable terminal abnormal nodes and divides this sum by the total number of corresponding terminal abnormal nodes to obtain the topology depth weight of the candidate. The smaller the topology depth weight, the fewer the average shortest path hops from the candidate to the terminal abnormal nodes in the current fault propagation graph. This topology depth weight is only used as a secondary sorting criterion for candidates with the same out-degree and is not used to directly distinguish between electrical hard blocking and logical soft interference. The system compares the topology depth weights of each candidate and determines the candidate with the smallest weight as the highest priority root cause candidate.

[0078] This method uses the shortest path hop count to calculate the topology depth weight, and performs a secondary sorting of the root cause candidates with the largest out-degree, providing graph topology basis for further narrowing down the candidate investigation range in the communication link diagnostic report.

[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting anomalies in a high-speed communication protocol of a computing chip, characterized in that, Includes the following steps: The link status records, data unit records, error monitoring records of each physical channel, and channel training sequence records of the high-speed digital communication port of the computing chip are acquired and merged into a synchronization window according to the digital transmission timestamp. Based on the link status record and data unit record, the link status violation event and the data unit violation event are obtained; The error monitoring records are aggregated using the physical channel as the row and the time sub-windows within the synchronization window as the column to obtain the channel error distribution table; the channel transmission defect event is obtained by matching the channel error distribution table with the preset transmission defect judgment rules. Based on the channel training sequence record, calculate the transmission delay difference of each physical channel relative to the reference physical channel, and form a channel skew sequence in chronological order; calculate the static skew amount and dynamic skew change rate based on the channel skew sequence, compare them with the preset link transmission tolerance, and obtain the channel skew warning event. Link status violation events, data unit violation events, channel transmission defect events, and channel skew warning events within the observation time window are constructed as abnormal event items. These are then correlated according to their occurrence sequence, physical channel correlation, data unit correlation, and transmission level correlation to obtain a fault propagation relationship diagram. A communication link diagnostic report is then generated based on this diagram.

2. The anomaly detection method for the high-speed communication protocol of computing chips according to claim 1, characterized in that, The synchronization window division includes: link status records corresponding to the link control layer, data unit records corresponding to the transaction layer, and error monitoring records and channel training sequence records corresponding to the physical transmission layer; the four types of records are divided with a fixed time length; records whose timestamps fall within the same partition are merged into the same synchronization window; and records with missing timestamps are filled with null values.

3. The anomaly detection method for the high-speed communication protocol of computing chips according to claim 1, characterized in that, The process of obtaining link state violation events and data unit violation events includes: comparing the state transition sequence in the link state record with a pre-set legal state transition table hop by hop; for state transitions not in the legal state transition table, recording the predecessor state, successor state, and timestamp to obtain link state violation events; verifying the field values ​​of each data unit in the data unit record with the field constraints defined by the protocol specification field by field; for data units that violate field constraints and / or transaction sequence order, recording their data unit identifier, violation fields, and timestamp to obtain data unit violation events.

4. The anomaly detection method for the high-speed communication protocol of computing chips according to claim 1, characterized in that, The process of obtaining the channel error distribution table includes: dividing the synchronization window into several time sub-windows with a fixed time step; constructing a two-dimensional table with the physical channel as the row index and the time sub-window number as the column index; assigning the error monitoring records of each physical channel to the corresponding time sub-window according to the timestamp, counting the number of error entries in each time sub-window and filling it into the corresponding table entry; setting the table entry with no error monitoring records to zero, thus obtaining the channel error distribution table.

5. The anomaly detection method for the high-speed communication protocol of computing chips according to claim 1, characterized in that, The process of obtaining channel transmission defect events includes: calculating the length of a continuous non-zero sub-window along the time axis of the channel error distribution table to obtain burst continuity characteristics; calculating the Pearson correlation coefficient between the sequence of error entries formed by adjacent physical channels within each time sub-window along the channel axis of the channel error distribution table to obtain channel synchronization correlation characteristics; globally calculating the normalized autocorrelation function of the number of error entries in the channel error distribution table and extracting the lag order corresponding to the first effective autocorrelation peak to obtain periodic oscillation characteristics; and using the burst continuity characteristics, channel synchronization correlation characteristics, and periodic oscillation characteristics as error distribution judgment quantities and matching them with preset transmission defect judgment rules to obtain channel transmission defect events carrying defect mode types.

6. The anomaly detection method for the high-speed communication protocol of computing chips according to claim 1, characterized in that, The process of obtaining a channel skew warning event includes: the preset link transmission tolerance includes a static tolerance and a dynamic tolerance; arranging the transmission delay differences of each physical channel within a continuous synchronization window in chronological order to obtain a channel skew sequence; taking the delay difference corresponding to the current synchronization window of the channel skew sequence as the static skew amount; performing linear fitting on a sliding window of a preset length on the channel skew sequence, and taking the slope of the fitted line as the dynamic skew change rate; comparing the static skew amount with the static tolerance, and comparing the dynamic skew change rate with the dynamic tolerance, and generating a channel skew warning event if either exceeds the limit.

7. The anomaly detection method for the high-speed communication protocol of computing chips according to claim 1, characterized in that, The process of obtaining the fault propagation graph involves: constructing abnormal event items based on timestamps, protocol layers, event attributes corresponding to the event type, and physical channel identifiers or data unit identifiers carried by the source records, for link state violation events, data unit violation events, channel transmission defect events, and channel skew warning events; using physical channel identifiers or data unit identifiers as the association basis, pointing abnormal event items carrying the same valid identifier and with an earlier timestamp to the next abnormal event item with the nearest timestamp; establishing directed associations from physical transport layer abnormal event items to link control layer abnormal event items based on the same physical channel identifier, and establishing directed associations from link control layer abnormal event items to transaction layer abnormal event items based on the same data unit identifier, wherein the start timestamp is no later than the end timestamp; abnormal event items without corresponding valid identifiers do not participate in this type of association; when the same abnormal event item satisfies multiple association conditions, all corresponding associations are retained, thus obtaining the fault propagation graph.

8. The anomaly detection method for the high-speed communication protocol of computing chips according to claim 1, characterized in that, The process of generating a communication link diagnostic report based on the fault propagation graph includes: identifying abnormal event items with zero in-degree and non-zero out-degree in the fault propagation graph as root cause candidates; when there is only one root cause candidate, it is identified as the highest priority root cause candidate; when there are multiple root cause candidates, they are arranged in descending order of their out-degree, and the candidate with the largest out-degree is identified as the highest priority root cause candidate; when there are multiple candidates with the same largest out-degree, all of them are identified as the highest priority root cause candidates; and writing the protocol layer to which the highest priority root cause candidate belongs, the associated path reachable from the candidate, and the physical channel identifier or data unit identifier it carries into the communication link diagnostic report.

Images