Server motherboard fault location method based on memory channel error correction event migration
By constructing component reference events and residing reference events, the problem of difficult fault location during server memory module migration was solved, enabling accurate location of faults in memory modules and motherboard channel paths, and improving the accuracy and efficiency of fault location.
Patent Information
- Application Number
- CN202611571181.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-12
- Publication Date
- 2026-11-17
AI Technical Summary
During the migration of server memory modules, existing technologies struggle to effectively distinguish whether a fault is caused by the memory module itself or the motherboard channel path, making it difficult to pinpoint the location of a server motherboard fault.
By obtaining the first and second anchoring states of the target server, component reference events and retention reference events are constructed respectively. The migrated error correction events are then mapped to the two types of reference events to generate corresponding results and determine the fault anchoring object.
This technology enables accurate differentiation of whether a fault is caused by the memory module itself or the motherboard channel path after memory module migration, improving the accuracy and efficiency of server motherboard fault location.
Smart Images

Figure CN123114835A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer hardware fault location technology, and more specifically to a server motherboard fault location method based on memory channel error correction event migration. Background Technology
[0002] A server memory subsystem typically consists of a processor memory controller, motherboard memory channels, connectors, memory slots, and memory modules. Error correction events generated during server operation can include error addresses, memory channels, rank, error syndromes, and related register information. This information reflects the operating status of the memory modules and their respective physical channels.
[0003] When multiple memory modules exhibit similar error characteristics, or when some error correction fields are missing, the migrated error evidence may not be able to effectively distinguish whether the fault migrated with the memory module or remained on the original motherboard channel path, thus affecting the further determination of the server motherboard fault location. Summary of the Invention
[0004] The purpose of this invention is to provide a server motherboard fault location method based on memory channel error correction event migration.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a server motherboard fault location method based on memory channel error correction event migration, the method comprising: Obtain the first anchoring state, the second anchoring state, and the motherboard topology field of the target server. The first anchoring state is the state in which the fault is anchored to the memory module itself and migrates with the location of the memory module. The second anchoring state is the state in which the fault is anchored to the motherboard memory channel path and does not migrate with the location of the memory module. Based on the first anchoring state, the motherboard topology field is remapped from the source motherboard topology location to the target motherboard topology location, while maintaining the relative error field within the memory module, to obtain the component reference event; Based on the second anchoring state, the motherboard topology field remains unchanged in the source motherboard topology position, and the memory module identity is rebound to the memory module identity that occupies the source motherboard topology position after migration, thus obtaining the retention reference event; The migrated error correction events are then mapped to the accompanying reference events and the retention reference events, generating corresponding results, and the fault anchoring objects are determined based on the corresponding results. When the fault anchor object is the motherboard channel path, output the location of the server motherboard fault.
[0006] In a preferred embodiment, the migrated error correction events are mapped to accompanying reference events and retention reference events, respectively, including: Obtain the sub-matching quantities corresponding to the motherboard topology and memory module identities respectively; The feature weights are determined based on the repeatability of each feature in error correction events before and after migration, and the distinguishability of each feature between accompanying reference events and retention reference events. The corresponding results are generated based on the feature weights and the feature change relationship between the first anchoring state and the second anchoring state.
[0007] In a preferred embodiment, when the error correction event after migration has a target error feature that cannot be normalized, the error features that have been normalized within the same observation window are determined as the context feature set. Historical event windows containing the context feature set are selected from historical error correction events. Each known error feature that appears in conjunction with the context feature set and its corresponding co-occurrence frequency are obtained. Each known error feature is weighted according to the co-occurrence frequency to obtain the surrogate feature of the target error feature. The surrogate feature is then used to participate in the acquisition of the sub-matching quantity.
[0008] In a preferred embodiment, the historical event window is filtered according to the platform type or compatible platform type corresponding to the target server, the co-occurrence frequency of each known error feature is normalized, and the feature vectors of each known error feature are weighted and combined according to the normalized co-occurrence frequency to obtain the proxy feature. When the number of historical event windows that satisfy the context feature set is less than the preset effective number, the matching weight corresponding to the target error feature is reduced.
[0009] In a preferred embodiment, the migrated error correction events are mapped to accompanying reference events and retention reference events, respectively, including: The migration-corrected events and their corresponding reference events are used as two types of nodes, and the event correspondence is established based on the comprehensive matching amount between each node. When there are multiple error correction events corresponding to multiple memory modules in the same observation window, the reference event corresponding to each migrated error correction event is determined based on the comprehensive matching quantity.
[0010] In a preferred embodiment, before obtaining the component reference event and the leave reference event, the error correction event is field-standardized, the resolvable memory address is converted into memory level coordinates, and the field associated with the motherboard physical path is determined as the motherboard topology field.
[0011] In a preferred embodiment, the normalized row address in the relative error field within the memory module is obtained by grouping the row addresses according to a preset address granularity.
[0012] In a preferred embodiment, a first observation window is constructed before the memory module is migrated, and a second observation window is constructed after the memory module is migrated and the target server completes memory training. Error correction events are obtained from the first and second observation windows respectively. Duplicate error correction events from different fault record sources are deduplicated, and the event intensity of the error correction events is normalized according to at least one of memory access volume, memory bandwidth, and observation duration.
[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: Compared to methods that directly compare location or error fields between error correction events before and after migration, this invention first obtains a first anchoring state, a second anchoring state, and a motherboard topology field. Second, in the first anchoring state, the motherboard topology field is remapped from the source motherboard topology location to the target motherboard topology location, while maintaining the relative error fields within the memory module, forming a component-specific reference event. Then, in the second anchoring state, the source motherboard topology location is maintained, and the memory module identity is rebound to the identity of the memory module occupying that location after migration, forming a retention reference event. Finally, the error correction events after migration are mapped to the two types of reference events respectively. Because the two types of reference events retain invariants corresponding to the memory module itself and the motherboard physical path, and apply different processing to fields that change with migration, the same error correction event can form comparable event representations in both fault attachment states, thus providing clear data evidence for distinguishing between memory module faults and motherboard channel path faults. Attached Figure Description
[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0015] Figure 1 This is a schematic diagram of the server motherboard fault location method in this invention.
[0016] Figure 2 This is a schematic diagram of the transformation relationship of the dual-anchored state field in this invention.
[0017] Figure 3 This is a schematic diagram illustrating the memory module migration and fault anchoring relationship in this invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on 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.
[0019] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.
[0020] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of the invention described herein can be implemented in an order other than that illustrated or described herein.
[0021] In this embodiment of the invention, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.
[0022] Unless otherwise defined, all technical and scientific terms used in the embodiments of this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in the embodiments of this invention is for descriptive purposes only and is not intended to limit the invention.
[0023] Before further explaining the specific implementation of the present invention, the main data, states, and identifiers involved in the present invention will be explained. The first anchoring state describes the hypothetical state where the root cause of the fault is attached to the memory module body and migrates with the position of the memory module; the second anchoring state describes the hypothetical state where the root cause of the fault is attached to the motherboard memory channel trace, connector, or slot path, and remains on the original board-level path after the memory module is migrated. The above two states are used to construct different reference events, and are not fault conclusions obtained directly from detection by a single sensor.
[0024] In some implementations, the first and second anchoring states are pre-configured by the diagnostic program and loaded when the diagnostic task starts, used to define the field transformation rules of the reference event; the source motherboard topology location and the target motherboard topology location are determined by the memory module migration mapping. The memory module migration mapping can be established after reading the SPD identity and slot topology information before and after the memory module is disassembled and reassembled, or it can be imported from maintenance records or BMC asset change records. Therefore, the first and second anchoring states, and their corresponding topology transformations, all have clear data sources and acquisition paths.
[0025] In some implementations, memory module identity is used to identify the same DIMM across observation intervals before and after migration. It can preferentially use the serial number recorded in the SPD or a unique identifier provided by the platform. When the platform does not have a single unique field, at least two of the following can be combined: manufacturer identifier, part number, serial number, capacity, and Rank organization. The combined result can then be encoded or hashed to form a stable memory module identity. Motherboard topology location is used to characterize the correspondence between the memory module and the physical path of the motherboard memory channels. It can consist of one or more of the following: processor socket, memory controller port, channel, slot, and board-level Route-ID. The Route-ID can be obtained from motherboard design data, the BMC asset table, or a model topology table corresponding to the target server model.
[0026] It should be noted that the composite identity is only used for identity maintenance across observation windows when it can uniquely correspond to a memory module in the current diagnostic task. If the composite fields of multiple memory modules are completely identical and cannot be distinguished, the maintenance tool can assign a temporary unique identifier to the corresponding memory module before disassembly and reassembly, and bind the temporary unique identifier to the slot record before and after disassembly and reassembly. If the identity ambiguity still cannot be eliminated, the relevant event will be marked as insufficient identity information, and the memory module identity will not be used to form evidence of consistency with the component, so as to avoid incorrect rebinding caused by non-unique composite identifiers.
[0027] For ease of subsequent explanation, the key terms used in this invention and their optional implementations are described below.
[0028] The memory module identification identifier U is used to identify the same DIMM across migration observation intervals. It can be formed by a combination of SPD serial number, platform unique identifier, or manufacturer identifier, part number, serial number, capacity and Rank organization identifier.
[0029] The motherboard-side topology location P is used to indicate the relationship between the DIMM and the physical path of the motherboard memory channel, and may include Socket, memory controller port, Channel, Slot and Route-ID.
[0030] Error correction event e is associated with at least the memory module identity, motherboard-side topology location, and error characteristics, and may also include event time, CE / UE type, MCA Bank, syndrome, Rank, Bank Group, Bank, Row / Column cluster, Device / DQ, and read / write volume.
[0031] The memory module migration mapping (M) records the correspondence between the memory module's migration location from the source motherboard topology location to the target motherboard topology location. This can be generated through maintenance record import, BMC asset change comparison, or automatic recording by maintenance tools.
[0032] The relative error fingerprint R within a module extracts relative module representations from error features that are as independent of the absolute value of the host physical address as possible, such as Rank, Bank, Device / DQ, normalized Row clusters, and syndrome families.
[0033] The accompanying reference event is an event transformed from the pre-migration error correction event in the first anchoring state. It maintains the memory module identity and stable intra-module relative error fields, and remaps the motherboard topology location to the target location.
[0034] The reference event is an event transformed from the error correction event before migration in the second anchoring state. It maintains the source motherboard channel path and rebinds the memory module identity to the memory module identity that occupies the path after migration.
[0035] Transfer consistency is a comprehensive matching quantity formed by pairing the real error correction event after transfer with the corresponding reference event. It is used to characterize the degree of conformity between the real event and a certain anchor state.
[0036] Based on the above data definition, a memory module location change can be broken down into two independent physical relationships: "whether the memory module identity is maintained" and "whether the motherboard physical path is maintained." In subsequent processing, instead of directly using the host physical address before and after the migration as the sole basis for comparison, the motherboard topology field and the relative error field within the module are used to establish a comparable relationship between events. This reduces the impact of memory retraining, changes in address interleaving methods, and differences in address decoding across different platforms.
[0037] The term "error correction event" in this invention refers to a collective term for memory error events recorded by the ECC or RAS mechanism, which can include correctable errors (CE) and uncorrectable errors (UE). CE represents a record formed after an error has been corrected by the ECC mechanism; UE represents an error record detected by the ECC mechanism but unable to be corrected. Including UE in the unified error correction event data structure is only for fault evidence analysis and does not indicate that the UE has been corrected.
[0038] In some implementations, this invention can be executed locally on the target server or collaboratively by the target server and the management server. When executed locally on the target server, the processor reads error correction events from at least one data source, including the MCA or MCI register, EDAC, mcelog, rasdaemon, BIOS RAS record, BMC SEL, and vendor extended memory diagnostic logs, and performs field standardization, reference event construction, event mapping, and fault location output locally. When using a collaborative approach, the target server is responsible for collecting and uploading error correction events, SPD information, and topology information, while the management server is responsible for uniformly maintaining memory module migration relationships and historical fault windows, and completing subsequent data processing.
[0039] The software logic for implementing the above processing can be functionally divided into a data acquisition unit, an event standardization unit, an anchoring transformation unit, an event mapping unit, a fault anchoring unit, and a historical feature management unit. The data acquisition unit acquires error correction events before and after migration, as well as memory module asset information; the event standardization unit maps fields from different log sources to a unified data structure; the anchoring transformation unit constructs accompanying reference events and retention reference events for the same pre-migration error correction event; the event mapping unit compares the correlation between the real event after migration and the two types of reference events; the fault anchoring unit determines the fault attachment object and motherboard fault location based on the mapping results; and the historical feature management unit saves confirmed fault windows for calculating feature repeatability, distinguishability, and proxy representations of unnormalized error features. These functions can be completed by the same program or by multiple program modules working together.
[0040] like Figure 1 As shown, the present invention first obtains the first anchoring state, the second anchoring state, and the motherboard topology field, constructs the component reference event and the reservation reference event respectively, then maps the migrated error correction event to the two types of reference events, determines the fault anchoring object based on the mapping result, and outputs the server motherboard fault location when the fault anchoring object is the motherboard channel path.
[0041] Example 1 This embodiment provides an implementation method for constructing reference events and locating motherboard faults based on dual anchoring states. The method includes: obtaining a first anchoring state, a second anchoring state, and a motherboard topology field of the target server, wherein the first anchoring state is a state in which the fault is anchored to the memory module itself and migrates with the location of the memory module, and the second anchoring state is a state in which the fault is anchored to the motherboard memory channel path and does not migrate with the location of the memory module; remapping the motherboard topology field from the source motherboard topology location to the target motherboard topology location based on the first anchoring state, while maintaining the relative error field within the memory module, to obtain a component reference event; keeping the motherboard topology field unchanged in the source motherboard topology location based on the second anchoring state, and rebinding the memory module identity to the identity of the memory module occupying the source motherboard topology location after migration, to obtain a retention reference event; mapping the migrated error correction event to the component reference event and the retention reference event respectively, generating corresponding results, and determining the fault anchoring object based on the corresponding results; and outputting the server motherboard fault location when the fault anchoring object is the motherboard channel path.
[0042] In this embodiment, both the first anchoring state and the second anchoring state serve as state constraints used when transforming the same pre-migration error correction event. In the first anchoring state, the root cause of the fault is considered to be located inside the DIMM; therefore, error characteristics with the same origin as this fault should move with the same DIMM to the new motherboard location. In the second anchoring state, the root cause of the fault is considered to be located in the motherboard memory channel trace, connector, or slot contact path; therefore, the fault should continue to manifest on the original motherboard channel path, while the identity of the DIMM occupying that path may change after migration. By applying two state constraints to the same pre-migration error correction event, two competing theoretical references can be formed.
[0043] In some implementations, a first observation window is built before the memory module migration, and a second observation window is built after the memory module migration and the target server has completed memory training and entered a stable operating state. The first observation window can be selected from the most recent stable operating range before the migration, or it can be formed using a preset memory stress load in maintenance mode. The second observation window can use the same or comparable diagnostic load as the first observation window. To reduce the discrepancy in the number of events caused by workload differences, the CE count can be converted into the event intensity per unit of memory access, per unit of memory bandwidth, or per unit of time; when the platform cannot read the memory access count, the two observation windows can be kept with similar observation durations and diagnostic loads.
[0044] Error correction events can be obtained from multiple fault log sources. For cases where the same original error may appear in both the BIOS log and the BMC log, deduplication can be performed based on at least two of the following: timestamp, physical address, error syndrome, and event source. For error correction events with resolvable physical addresses, the address can be converted to memory-level coordinates according to the target server's address decoding rules. For events where complete addresses cannot be resolved, fields such as Channel, Slot, Rank, Bank, Device / DQ, and syndrome that can be identified can be retained, and missing fields can be marked as unobserved, allowing the event to still participate in subsequent matching.
[0045] In some implementations, the normalized Row clusters in the relative error field within the memory module can be grouped according to a preset address granularity. For example, consecutive 2 k Each row is grouped into the same cluster, where k is an integer greater than or equal to 0, determined by the server's memory organization and the desired address resolution. When the platform cannot obtain the Row field, at least two of Rank, Bank, Device / DQ, and syndrome families can be used to form a degradation-relative error fingerprint. This process ensures that the relative error field within the module is as independent as possible of the absolute value of the host's physical address, which may change after migration.
[0046] like Figure 2 As shown, in the first anchoring state, the same pre-migration error correction event maintains the original memory module identity and the relative error field within the module that is stable during migration, and remaps the motherboard topology field from the source location to the target location; in the second anchoring state, it maintains the source motherboard topology location, and rebinds the memory module identity to the memory module identity that occupies the source location after migration, thereby forming a component reference event and a reserved reference event respectively.
[0047] To clarify the processing relationship of each field under the two anchoring states, the specific hierarchical and transformation rules of each field are as follows.
[0048] The identity fields include the DIMM serial number, SPD unique identifier, and manufacturer / part number combination identifier. In the first anchoring state, the original DIMM identity is maintained; in the second anchoring state, the memory module identity is rebound to the DIMM identity that occupies the source location after migration.
[0049] The motherboard topology fields include Socket, controller port, Channel, Slot, and Route-ID. In the first anchoring state, the motherboard topology fields are transformed from the source location to the target location according to the migration relationship; in the second anchoring state, the source motherboard channel path and Route-ID remain unchanged.
[0050] The relative fields within the module include Rank, Bank Group, Bank, Device / DQ, and normalized Row clusters. In the first anchoring state, the intra-module fault fingerprint with migration stability is maintained; in the second anchoring state, the corresponding fields are not required to follow the original DIMM, and only the comparable parts related to the board-level signal path are retained.
[0051] Error attribute fields include CE / UE category, syndrome family, parity mode, and burst / repetitive mode. In the first anchoring state, consistency constraints are maintained on inherent error attributes within the DIMM; in the second anchoring state, consistency constraints are maintained on channel signal quality-related error attributes.
[0052] The time / load field includes CE rate per unit read / write volume, event interval, and event intensity in the observation window. Both are compared after being normalized by load or observation duration in the first and second anchoring states.
[0053] The above field processing relationship does not require each platform to provide all fields simultaneously. Fields that can be reliably obtained from the target server can be used as input for corresponding matching; for missing or low-repeatability fields, their participation can be reduced or proxy features from subsequent embodiments can be used. This allows error correction logs from different CPU generations and different BIOS or BMC versions to enter the same dual-reference processing framework.
[0054] Specifically, any error correction event before migration can be represented as an event record consisting of the memory module identity, motherboard-side topology location, and error characteristics. The memory module migration relationship records the change of the memory module from the source motherboard topology location to the target motherboard topology location. When two DIMMs are swapped, the migration relationship simultaneously records the migration of the first DIMM from the first position to the second position and the migration of the second DIMM from the second position to the first position, thus enabling the query to determine which memory module occupies any source location after migration.
[0055] In the first anchoring state, the original memory module identity from the pre-migration event is maintained. The motherboard topology field is rewritten from the source motherboard topology location to the target motherboard topology location after the actual migration of the memory module. At least one of the following is retained: Rank, Bank, Device / DQ, Normalized Row Cluster, and Syndrome Family, which are determined to be migration-stable. This forms a component-specific reference event. Physically, if the fault is located within the memory module itself, error correction evidence originating from the same source as the original fault after migration should primarily appear at the new location of the original memory module, preserving relative error characteristics related to the internal organization of that memory module.
[0056] In the second anchoring state, the source motherboard topology location corresponding to the event before migration remains unchanged. Based on the memory module migration relationship, the memory module that actually occupies the source motherboard topology location after migration is queried, and the memory module identity in the event is rebound to the queried identity. Fields that only characterize the original DIMM's internal organization are not required to be retained, but comparison constraints are retained for at least one of Route-ID, Channel, Slot, channel signal quality related syndromes, and repeated CE patterns, thus forming a reference event for retention. Its physical meaning is that if the fault is located in the motherboard channel path, the error still tends to appear in the original board-level physical path, but the DIMM identity recording the error can change with the relocation.
[0057] After forming two types of reference events, the migrated real error correction events in the second observation window are mapped to the accompanying reference events and the retained reference events, respectively. When the number of events is small, the reference event with the highest comprehensive matching degree can be selected for each event; when the number of events is large, a global event mapping relationship can be established according to the method in Example 2. If the number of real events that can be explained by the retained reference events is higher than that of the accompanying reference events, the fault anchoring object tends to be the motherboard channel path, and the Route-ID, Channel, or Slot that the retained reference events with high contributions point to can be used as the motherboard fault location; if the accompanying reference events are dominant, the fault tends to be more inclined to the memory module itself; when neither type of reference can form a stable mapping, an insufficient evidence state can be output and observation data can continue to be collected.
[0058] like Figure 3 As shown, when DIMM-A is migrated from C2D0 to C2D1 and DIMM-B is migrated from C2D1 to C2D0, the component reference relationship moves to C2D1 along with DIMM-A, while the reserved reference relationship keeps the motherboard channel path corresponding to C2D0 unchanged, and updates the memory module identity at that position to DIMM-B.
[0059] For example, slots C2D0 and C2D1 are set on Channel 2 of CPU0. Before migration, DIMM-A is located in C2D0, and DIMM-B is located in C2D1. In the first observation window, DIMM-A exhibits multiple CE events in C2D0, with similar Rank, Bank Group, Device / DQ, and syndrome families. After swapping DIMM-A and DIMM-B, the accompanying reference event changes the motherboard topology location corresponding to the original C2D0 to C2D1, while maintaining the identity of DIMM-A and a stable intra-module relative error field; the retention reference event still maintains the motherboard topology location corresponding to C2D0 and rebinds the memory module identity to the currently located DIMM-B in C2D0. If the actual CE in the second observation window mainly appears in DIMM-B on C2D0 and maintains a high degree of correspondence with the retention reference event, the fault can be anchored to the motherboard channel path corresponding to C2D0; if the actual CE mainly moves to DIMM-A on C2D1, it is more consistent with the first anchoring state.
[0060] This embodiment establishes two different field transformation relationships for the same pre-migration error correction event, so that the memory module body and the motherboard channel path each correspond to a set of computable reference events. This process preserves the physical invariants caused by different fault attachment objects, so that the post-migration event does not need to rely solely on the single phenomenon of the slot reporting an error again for judgment, and can continue to use the memory module identity, board-level topology, and relative error fields within the module to establish the correspondence between the pre- and post-migration events when the address mapping changes.
[0061] Example 2 This embodiment, based on the dual-reference event formed in Embodiment 1, further explains the correspondence process between the migrated error correction event and the two types of reference events. When the migrated error correction event is mapped to the accompanying reference event and the retention reference event respectively, the sub-matching quantities corresponding to the motherboard topology and memory module identity can be obtained respectively. The feature weights are determined based on the repeatability of each feature in the error correction event before and after migration and the distinguishability of each feature between the accompanying reference event and the retention reference event. Then, the corresponding results are generated based on the feature weights and the feature change relationship corresponding to the first anchoring state and the second anchoring state.
[0062] In some implementations, the characteristics corresponding to the events may include motherboard topology, memory module identity, error syndrome, relative error location within the module, and event time density. The motherboard topology sub-matching quantity can be obtained based on the similarity or hierarchical proximity of Socket, controller port, Channel, Slot, and Route-ID; the memory module identity sub-matching quantity can be obtained based on whether the SPD serial number or combined identity is consistent; the sub-matching quantities of error syndrome and relative error location within the module can be obtained based on the similarity of syndrome family, Rank, Bank, Device / DQ, and normalized Row cluster; the event time density can be compared after load or observation duration normalization.
[0063] As one possible implementation, all sub-matching values can be standardized to the same normalization scale. For discrete fields such as memory module identity, Channel, Slot, and Route-ID, a higher sub-matching value is assigned when the fields are consistent, and a lower sub-matching value is assigned when the fields are inconsistent. For motherboard topology fields with hierarchical relationships, the sub-matching value can be determined based on the finest topology level that the two events both match. For continuous fields such as event time density, the sub-matching value can be determined based on the degree of difference after normalization. Missing fields are not included in the calculation of the sub-matching value for the current event to avoid misjudging "unobserved" as "inconsistent".
[0064] Feature repeatability characterizes whether a feature can be stably obtained within the historical observation window of the same DIMM or the same topology anchor. Feature discriminability characterizes whether the expected value of the feature between the following reference event and the remaining reference event is sufficient to distinguish the differences between the two anchoring states. For example, the memory module identity should be maintained in the first anchoring state, but may be rebound in the second anchoring state, thus having high discriminability when a transposition occurs; conversely, some changes in absolute physical address caused by retraining after migration may have low repeatability, which can reduce its impact on the final correspondence result.
[0065] In some implementations, for each feature dimension, its stability and observability are first statistically analyzed based on historically confirmed fault windows to determine repeatability. Then, the expected relationship of the feature in two types of reference events is compared to form discriminability. Subsequently, normalization is performed according to the principle that the feature weight increases as repeatability and discriminability increase simultaneously. The above weights can be recalculated for each new diagnostic window, allowing the same set of corresponding logic to adapt to different server platforms and different log completeness levels.
[0066] One feasible way to obtain repeatability is to count the number of windows in the same memory module or the same topological anchor point where the target feature can be stably observed and its value is within the allowable fluctuation range, and then proportionally compare this number to the total number of historically valid windows. Discrimination can be determined based on the difference in the expected relationship between the feature in incident reference events and standby reference events. When determining feature weights, the product of repeatability and discrimination can be used as the unnormalized weight, and then the unnormalized weights of the current valid features can be normalized; when all unnormalized weights are zero, it can degenerate into equal-weighting the current valid features. The allowable fluctuation range can be determined by historically confirmed fault windows or the diagnostic configuration corresponding to the target server.
[0067] After determining the feature weights, we can first calculate the basic consistency degree under a certain anchoring state. One feasible calculation method is as follows: ; Among them, B h This represents the basic consistency degree corresponding to the h-th anchoring state, where h can represent either the first or second anchoring state; w j sim represents the weight of the j-th feature dimension; h,j This represents the sub-match quantity obtained for the j-th feature dimension in the h-th anchor state of the real event after migration; the summation range covers all valid feature dimensions corresponding to the currently participating event. For missing fields, the corresponding sub-match quantity may not be included in the summation, and the weights of the remaining valid features are renormalized.
[0068] Where J represents the total number of valid feature dimensions corresponding to the current participating event, and j represents the feature dimension index. If a feature is not included in the matching of the current event due to missing logs, then that feature is not counted in j, and the feature weight is renormalized among the remaining valid features.
[0069] To further reflect the physical relationship of "which fields should be maintained and which fields should be changed" under the two anchoring states, a feature correction can be calculated for each anchoring state. A positive correction is generated when the direction of field change in the actual event matches the expected change relationship of the current anchoring state, and a negative correction is generated when the direction of change is opposite. A feasible form of migration consistency derived from the basic consistency and feature correction is as follows: One feasible method for obtaining feature correction quantities is as follows: Examine the direction of change of the real event relative to the event before migration on each valid feature; if the change conforms to the expected relationship of "should be maintained" or "should be changed" under the current anchoring state, a positive correction contribution is formed; if the direction of change is opposite to the expected relationship, a negative correction contribution is formed; if the corresponding feature is missing or the direction of change cannot be determined, no correction contribution is formed. By summarizing the correction contributions of each valid feature according to their respective feature weights, the feature correction quantity corresponding to the current anchoring state can be obtained, thus making the feature correction quantity acquisition process repeatable.
[0070] ; Among them, S h Q represents the migration consistency degree corresponding to the h-th anchoring state; h This represents the feature correction amount obtained based on the feature retention or change relationship of the anchoring state; This indicates the weight of the correction term. The determination can be based on historically confirmed fault samples, or it can be provided by the diagnostic configuration corresponding to the server model and limited to a preset range. The above calculation method is used to illustrate one possible implementation method and does not require different platforms to use the same numerical scale. It is only necessary to ensure that the two anchoring states within the same diagnostic window use consistent calculation rules.
[0071] When multiple memory modules and multiple error correction events exist within the same observation window, the migrated error correction events can be treated as one type of node, and the reference events as another type of node. The overall matching quantity can be used as the edge weight between nodes, thus forming an event correspondence graph. Performing maximum weight matching separately for accompanying reference events and reserved reference events can obtain the global interpretation of the current real event set for each of the two anchoring states, avoiding the same real event being repeatedly assigned to multiple reference events. For cases with a small number of events or significant differences between events, a per-event maximum matching method can also be used.
[0072] For example, when Channel 1 and Channel 2 of CPU0 both experience CE (Error Detection), and multiple DIMMs are cross-migrated according to a replacement relationship during a maintenance, individually selecting the local best reference for each event might result in two real events pointing to the same pre-migration event. By using global event mapping, the system selects the event combination with the higher overall edge weight and no conflict within each type of reference event. If multiple real events point to the same Route-ID in the second anchoring state, while only scattered mappings can be formed in the first anchoring state, the consistency of the hold-over migration is relatively high, which can be used to determine the motherboard channel path as the primary fault anchoring object.
[0073] In some implementations, the number of events that can form a valid correspondence can also be considered when determining the fault anchoring object. When the retention migration consistency relative to the component migration consistency reaches a preset dominance condition, and the number of explainable valid events reaches the minimum number of valid events, the fault is anchored to the motherboard channel path; when the component migration consistency meets the corresponding dominance condition, the fault is anchored to the memory module itself; when neither is met, an insufficient evidence status is output. The dominance condition and the minimum number of valid events can be determined based on historically confirmed fault samples of the target server, vendor diagnostic strategies, or statistical results from the testing phase.
[0074] For example, the difference between the retention migration consistency and the component migration consistency can be no less than a preset advantage threshold, and the number of valid corresponding events can be no less than the minimum number of valid events, serving as an advantage criterion for anchoring the fault to the motherboard channel path. The preset advantage threshold is used to limit the minimum difference that needs to be achieved between the two types of migration consistency, and the minimum number of valid events is used to limit the scale of valid events that can support the current conclusion; both can be determined statistically from historical confirmed fault samples of the same or compatible server models and saved in the diagnostic configuration. When historical samples are insufficient, an initial configuration can be provided based on the known fault verification results during the testing phase or the manufacturer's diagnostic strategy.
[0075] This embodiment incorporates both the repeatability of features and the distinguishability between the two types of references into the event correspondence process. This allows stable fields that reflect differences in the objects to which faults are attached to to contribute more, while reducing reliance on fields that are prone to change after memory retraining or that are missing in the current log. Simultaneously, anchoring state-specific correction ensures that the final consistency reflects not only the static similarity but also whether the direction of field changes conforms to the corresponding physical relationship, thus maintaining a consistent interpretation logic for the correspondence results under multiple event conditions.
[0076] Example 3 This embodiment illustrates the handling method when the migrated error correction event has target error features that cannot be normalized. When vendor-specific error codes, unknown syndrome texts, or missing device-level fields occur, the normalized error features within the same observation window are determined as the context feature set. Historical event windows containing this context feature set are selected from historical error correction events. The known error features that co-occur with the context feature set and their corresponding co-occurrence frequencies are obtained. The known error features are weighted according to the co-occurrence frequencies to obtain the proxy features of the target error features. The proxy features are then used in the sub-matching quantity calculation in Embodiment 2.
[0077] Historical error correction events can be organized according to the platform type or compatible platform type of the target server. The platform type can be determined by at least one of the following: CPU generation, motherboard model, BIOS or BMC log format version, and memory organization method. When filtering historical event windows, priority should be given to windows that can simultaneously contain key features of the current context, so that the agent features reflect historical co-occurrence relationships related to the current diagnostic environment, rather than using all historical errors for unconditional statistics.
[0078] In some implementations, suppose there are n known erroneous features in the filtered historical event window that can be used to construct surrogate representations, and the co-occurrence frequency of the i-th known erroneous feature is f. i After normalizing the co-occurrence frequencies, the co-occurrence weight of the i-th known erroneous feature can be obtained. An implementable calculation form is as follows: ; in, f represents the normalized co-occurrence weight of the i-th known erroneous feature; i This indicates the co-occurrence frequency of the known error feature within the historical event window that satisfies the current context conditions; the summation term covers all known error features used to construct the surrogate representation. For duplicate reports of events from the same source within the same historical window, deduplication can be performed first by timestamp, physical address, and syndrome before calculating the co-occurrence frequency.
[0079] After obtaining the feature vectors corresponding to each known error feature from the error feature dictionary or historical fault clustering results, a surrogate feature vector for the target error feature can be constructed based on the normalized co-occurrence weights. An achievable computational form is as follows: ; in, The surrogate feature vector representing the target error features; v i This represents the feature vector corresponding to the i-th known error feature; This represents the corresponding normalized co-occurrence weight. The surrogate feature vector is used for calculating the similarity of erroneous attributes in event correspondence. Its role is to enable target erroneous features that cannot be directly parsed to participate in matching with combinations of known features in the historical context, rather than directly forcing the target erroneous features to be interpreted as any single known erroneous feature.
[0080] When the number of historical event windows satisfying the context feature set is lower than the preset effective number, the matching weight corresponding to the target error feature can be reduced, or only other fields that have been normalized in the current event can be used for dual-reference correspondence. The preset effective number can be determined based on the size of the historical event database and the distribution of platform error types. By setting this degradation path, excessively high matching contributions from a small number of accidental co-occurrence relationships can be avoided when historical samples are insufficient.
[0081] When there are no known erroneous features in the screening results that can be used to construct a surrogate representation, or when the sum of the co-occurrence frequencies of all known erroneous features is zero, the co-occurrence weight normalization and surrogate feature vector construction are not performed. Instead, the above-mentioned degenerate path is directly adopted, and only other valid fields that have been normalized in the current event are used to participate in the correspondence of the two reference events, so as to avoid the situation where there is no valid denominator in the co-occurrence weight calculation.
[0082] For example, after the target server upgrades its BIOS, a new error code X9 appears in the second observation window. The existing parser cannot directly map X9 to an existing syndrome family. The system can take Channel2, Repeat CE, Rank0, and DQ group D3, which are already identified in the same window, as context, and search for records containing the above features in the historical event window of the same model or compatible platform. If it is known that syndromes G3 and G5 have a high co-occurrence frequency in these historical windows, the feature vectors corresponding to G3 and G5 can be combined according to their normalized co-occurrence weights to form a surrogate feature vector for X9, and this surrogate feature vector is compared with the error attributes in the accompanying reference event and the retention reference event.
[0083] In other implementations, the target error feature may be a missing device-level field rather than a newly added text error code. For example, a platform's BMC log may only record the Channel and syndrome, but not the Device / DQ. In this case, known Channel, Rank, Bank, and syndrome within the same window can be used as context. Device / DQ features that frequently appear with this context can be statistically analyzed from the historical fault window to form a surrogate representation. The surrogate representation only participates in the sub-matching quantity corresponding to the missing field; other observed motherboard topology and memory module identity fields are still directly compared according to the methods in Embodiments 1 and 2.
[0084] This embodiment utilizes the known context consistent with the current diagnostic window to filter historical events, and then constructs surrogate representations of target error features based on co-occurrence frequency. This allows some unknown fields generated by different CPU generations, BIOS or BMC versions, and manufacturer extended log formats to still enter the dual-reference event correspondence process. Simultaneously, by reducing the contribution of surrogate fields when historical samples are insufficient, the primary role of direct evidence such as known motherboard topology and memory module identity in fault anchoring can be maintained.
[0085] Example 4 This embodiment further explains the observation window, fault location output, and fault troubleshooting on the processor memory controller side. A first observation window is constructed before the memory module migration, and a second observation window is constructed after the memory module migration and the target server's memory training is completed. Error correction events are obtained from both observation windows. Duplicate error correction events from different fault record sources are deduplicated, and the event intensity is normalized based on at least one of memory access volume, memory bandwidth, or observation duration. This processing ensures that the event density before and after the migration is at a comparable data scale.
[0086] After the first observation window ends, memory module migration relationships can be formed through maintenance record import, BMC asset change comparison, or automatic recording by maintenance tools. When the maintenance tool automatically records, it can read the SPD identity and original topology location before the memory module is disassembled or reassembled. After the target server re-completes memory training, it reads the SPD identity of each slot again and generates a migration correspondence from the source location to the target location based on the identity changes before and after. For scenarios involving the swapping of two or more memory modules, the migration relationship retains both the independent source and target locations of each memory module, enabling the second anchoring state to determine the identity of the memory module actually occupying the original source location after migration.
[0087] After the second observation window begins, first complete the event field standardization and dual-reference event construction according to Example 1, and then complete the event mapping according to Example 2 or Example 3. When the result of the retention reference event mapping is dominant, obtain the motherboard topology field that is commonly associated with each error correction event that forms a valid retention mapping. If multiple valid events point to the same Route-ID, the motherboard memory channel path corresponding to that Route-ID can be used as the fault location; if the platform can only resolve to the Channel or Slot level, the finest topology level that can be stably determined can be output, without forcibly generating a finer position that the platform cannot support.
[0088] To avoid directly categorizing CPU internal memory controller malfunctions as motherboard channel path failures, controller-side exclusion conditions can be set. When the MCA or CPU RAS register explicitly indicates an internal IMC failure, or multiple different motherboard Route-IDs synchronously exhibit associated anomalies on the same memory controller port, or the platform self-test result points to an internal controller error, the corresponding fault is marked as a memory controller-side fault, and the output is stopped as a motherboard channel path failure. This exclusion process can be executed as an event filtering condition before dual-reference mapping, or as a result verification condition after the doubling reference event dominates.
[0089] For example, in a dual-socket server, the same memory controller port of CPU0 connects to two different channels. In the second observation window, different Route-IDs of the two channels simultaneously experience CEs (Error Detection) at similar times, and the CPURAS register reports an internal error of the memory controller port. Even if the hold reference events of a certain channel have a high degree of correspondence, the result can be marked as a memory controller-side anomaly based on controller-side exclusion criteria, rather than directly outputting it as a motherboard trace or slot fault. Conversely, if the CPU RAS does not report an internal controller anomaly, and the hold reference events consistently focus on a single Route-ID, then that Route-ID, Channel, or Slot can be output as the location of a server motherboard fault.
[0090] In the implementation of the electronic device, the electronic device may include a processor and a memory, the memory storing a computer program for executing the above methods. When the processor runs the computer program, it reads error correction events and memory module asset data from the fault record source of the target server, constructs two types of reference events based on pre-formed or real-time migration relationships, and completes event mapping, fault anchoring, and output of motherboard fault location. The processor may be a general-purpose processor, a server management processor, or other processing circuits with data processing capabilities; the memory may be volatile memory, non-volatile memory, or a combination of both.
[0091] The aforementioned computer program can also be stored in a computer-readable storage medium and executed after being loaded by the processor. Program data may include the target server's topology configuration, SPD identity mapping, historical error correction event windows, error signature dictionary, and diagnostic configurations for determining dominant conditions and the minimum number of valid events. This data can be stored in the same storage space or stored separately and loaded into the processor-accessible workspace when the diagnostic task starts.
[0092] This embodiment clarifies the observation window, migration relationship generation, event intensity normalization, and controller-side exclusion conditions, ensuring that dual-reference events have definite data acquisition paths and result output boundaries. For different server platforms, while maintaining the field transformation logic between the first and second anchor states, the specific data source can be adjusted according to the topology hierarchy and RAS field provided by the platform, thereby matching the fault location output with the platform's actual observable capabilities.
[0093] In some implementations, the error signature dictionary can maintain a mapping between syndrome families, CE / UE types, MCA banks, and vendor error codes according to platform version. For multiple records of the same original error generated from different log sources, the deduplication result should retain the most complete set of fields that can be parsed. For example, if the BMC log contains timestamps and slot information while the MCA record contains physical addresses and syndromes, their complementary fields can be merged after confirming that the two records belong to the same original error, rather than simply deleting either record.
[0094] In some implementations, Route-IDs can be provided directly from motherboard design data or generated from the model topology table based on a combination of Socket, controller port, Channel, and Slot. Regardless of the method used, the same Route-ID generation rules should be used for the same target server before and after migration. For platforms where board-level Route-IDs are unavailable, a reliably obtainable combination of Channel and Slot can be used as the motherboard topology anchor point, thus maintaining the meaning of "the source motherboard physical path remains unchanged" in the second anchoring state.
[0095] In some implementations, minimum valid event conditions can be set for both the pre- and post-migration observation windows. When the number of target events in the first observation window is insufficient to form a repeatable intra-module relative error fingerprint, the observation time can be extended or the diagnostic load increased; when the number of real events in the second observation window is insufficient to form a valid correspondence with both types of reference events simultaneously, an insufficient evidence status can be output. This process ensures that the fault anchoring result is consistent with the amount of currently available error-correcting evidence, and does not require the forced generation of motherboard fault locations when evidence is insufficient.
[0096] In summary, this invention uses the motherboard topology change resulting from the actual migration of memory module locations as a known condition. It projects the same pre-migration error correction event into two states: "fault migrates with memory module" and "fault remains on the motherboard channel path." Reference events are constructed by selectively maintaining or remapping motherboard topology fields, memory module identity, and relative error fields within the module. Subsequent event correspondence, dynamic feature weighting, unknown error feature proxying, and controller-side exclusion all revolve around this dual-reference relationship, enabling interpretable fault anchoring results under different log completeness levels and memory topology conditions.
[0097] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of the present invention are included within the scope of protection of the present invention.
Claims
1. A server motherboard fault location method based on memory channel error correction event migration, characterized in that, The method includes: Obtain the first anchoring state, the second anchoring state, and the motherboard topology field of the target server. The first anchoring state is a state in which the fault is anchored to the memory module body and migrates with the location of the memory module. The second anchoring state is a state in which the fault is anchored to the motherboard memory channel path and does not migrate with the location of the memory module. Based on the first anchoring state, the motherboard topology field is remapped from the source motherboard topology location to the target motherboard topology location, while maintaining the relative error field within the memory module, to obtain the component reference event; Based on the second anchoring state, the motherboard topology field remains unchanged in the source motherboard topology position, and the memory module identity is rebound to the memory module identity that occupies the source motherboard topology position after migration, thus obtaining the retention reference event; The migrated error correction events are then mapped to the accompanying reference events and the retention reference events, generating corresponding results, and the fault anchoring objects are determined based on the mapping results. When the fault anchoring object is the motherboard channel path, the server motherboard fault location is output.
2. The server motherboard fault location method based on memory channel error correction event migration according to claim 1, characterized in that, The process of mapping the migrated error correction events to accompanying reference events and retention reference events includes: Obtain the sub-matching quantities corresponding to the motherboard topology and memory module identities respectively; The feature weights are determined based on the repeatability of each feature in the error correction events before and after migration, and the distinguishability of each feature between the accompanying reference events and the retention reference events. The corresponding result is generated based on the feature weights and the feature change relationship corresponding to the first anchoring state and the second anchoring state.
3. The server motherboard fault location method based on memory channel error correction event migration according to claim 2, characterized in that, When the error correction event after migration has a target error feature that cannot be normalized, the error features that have been normalized within the same observation window are determined as the context feature set. Historical event windows containing the context feature set are selected from historical error correction events. Each known error feature that co-occurs with the context feature set and its corresponding co-occurrence frequency are obtained. Each known error feature is weighted according to the co-occurrence frequency to obtain the surrogate feature of the target error feature. The surrogate feature is then used to participate in the acquisition of the sub-matching quantity.
4. The server motherboard fault location method based on memory channel error correction event migration according to claim 3, characterized in that, The historical event window is filtered according to the platform type or compatible platform type corresponding to the target server. The co-occurrence frequency of each known error feature is normalized, and the feature vectors of each known error feature are weighted and combined according to the normalized co-occurrence frequency to obtain the proxy feature. When the number of historical event windows that satisfy the context feature set is less than the preset effective number, the matching weight corresponding to the target error feature is reduced.
5. The server motherboard fault location method based on memory channel error correction event migration according to claim 4, characterized in that, The process of mapping the migrated error correction events to accompanying reference events and retention reference events includes: The migration-corrected events and their corresponding reference events are used as two types of nodes, and the event correspondence is established based on the comprehensive matching amount between each node. When there are multiple error correction events corresponding to multiple memory modules in the same observation window, the reference event corresponding to each migrated error correction event is determined according to the comprehensive matching quantity.
6. The server motherboard fault location method based on memory channel error correction event migration according to claim 1, characterized in that, Before obtaining the accompanying reference event and the reserved reference event, the error correction event is standardized by converting the resolvable memory address into memory level coordinates and determining the field associated with the motherboard physical path as the motherboard topology field.
7. The server motherboard fault location method based on memory channel error correction event migration according to claim 6, characterized in that, The normalized row address in the relative error field within the memory module is obtained by grouping the row addresses according to a preset address granularity.
8. The server motherboard fault location method based on memory channel error correction event migration according to claim 1, characterized in that, A first observation window is built before the memory module is migrated, and a second observation window is built after the memory module is migrated and the target server completes memory training. Error correction events are obtained from the first and second observation windows respectively. Duplicate error correction events from different fault record sources are deduplicated, and the event intensity of the error correction events is normalized according to at least one of memory access volume, memory bandwidth, and observation duration.
9. An electronic device, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the server motherboard fault location method based on memory channel error correction event migration as described in any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the server motherboard fault location method based on memory channel error correction event migration as described in any one of claims 1 to 8.
