An ai-based carbon data intelligent calculation operating system
By allocating a non-paged contiguous physical space in memory and hardware-level addressing reconstruction, the problems of processor power loss and cache failure in ultra-large-scale data processing are solved, and efficient and deterministic energy carbon data processing is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SINRIDIGITALCITYTECCO LTD
- Filing Date
- 2026-05-06
- Publication Date
- 2026-08-04
AI Technical Summary
In ultra-large-scale data processing scenarios, the high-frequency and highly volatile data characteristics in existing technologies conflict with the general kernel scheduling mechanism, resulting in processor computing power loss, cache failure and bus congestion, making it difficult to maintain deterministic computing latency.
The energy carbon tensor management module allocates a non-paged contiguous physical space in memory. Combined with the bypass scheduling module and the mapping processing module, it achieves efficient data flow processing by real-time monitoring of data exchange frequency and shielding hardware interrupts through hardware-level addressing reconstruction and memory access topology mapping.
It effectively eliminates software-level observation overhead, ensures that processor computing resources are used for tensor computation, guarantees the absolute continuity and determinism of computation, reduces the bus bandwidth occupation of cache coherence probe signals, and improves the memory access bandwidth utilization of processor cores.
Smart Images
Figure CN122132186B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electrical digital data processing technology, and in particular relates to an AI-based intelligent computing operating system for energy and carbon data. Background Technology
[0002] In the current energy and carbon emission monitoring technology system, the high-frequency time-series data flow generated by tens of millions of terminals mainly relies on the general operating system kernel. The kernel adopts a standard network protocol stack and a virtual file system to process the heterogeneous energy and carbon data streams as electrical digital data, and after being parsed by the kernel-mode protocol, it enters the user-mode computing framework through context switching.
[0003] In ultra-large-scale data processing scenarios, the high frequency and highly volatile characteristics of data conflict with general kernel scheduling mechanisms. Standard interrupt handling mechanisms and fair scheduling algorithms generate frequent context switching when dealing with sudden traffic surges, leading to processor computational power loss due to interrupt responses and memory copying. Software-level comparison logic executed by the system to identify data flow fluctuations evicts the processor cache, causing cache invalidation and bus congestion, resulting in significant observational overhead. Furthermore, in dynamic simulations of ultra-large-scale topology networks, frequent disconnections of edge nodes lead to missing tensor dimensions. Existing technologies typically employ application-layer tensor reshaping to address this, causing pipeline flushing. In multi-processor architectures, data interaction between different physical regions triggers cross-node cache consistency probing storms, resulting in bus bandwidth being occupied by protocol probing signals. To address these issues, simply increasing the processor clock speed or adding hardware resources cannot eliminate the inherent overhead of kernel-to-user mode switching, nor can it resolve hardware logic conflicts induced by cross-processor node memory access. Existing technologies that address node disconnections through application-layer operator reorganization disrupt the continuity of the instruction flow, making it difficult to maintain deterministic computational latency.
[0004] Therefore, how to achieve efficient processing of heterogeneous data streams while avoiding state observation overhead is the technical problem that this invention aims to solve. Summary of the Invention
[0005] This invention provides an AI-based intelligent computing operating system for energy and carbon data, comprising:
[0006] The energy carbon tensor management module is used to allocate non-paged contiguous physical space in memory to store time-series energy consumption data streams that characterize the status of heterogeneous IoT terminals.
[0007] The bypass scheduling module is used to obtain the non-consistent memory access topology map of the physical server, identify the local memory nodes corresponding to different processor cores, and monitor the data exchange frequency between the processor cores and the energy carbon tensor management module in real time.
[0008] The mapping processing module is used to obtain the energy and carbon computing topology correlation matrix and perform graph partitioning, identify non-zero elements in the energy and carbon computing topology correlation matrix that represent cross-regional data interaction, extract topology boundary nodes connecting different computing regions, and generate kernel-mode topology boundary cut sets. The operating system is used to perform physical mirroring multi-projection mapping based on the non-consistent memory access topology graph and kernel-mode topology boundary cut sets, mirroring the original physical pages corresponding to the topology boundary nodes to multiple local memory nodes, and establishing page table alias redirection paths to transform cross-bus remote memory access into local memory page addressing access for the processor core.
[0009] Preferably, the energy carbon tensor management module is used to perform continuous physical space pre-allocation during the system startup phase, and the operating system performs memory attribute locking to prevent the continuous physical space from participating in virtual memory page swapping; the mapping processing module is used to address the input pointer of the computation graph operator to the continuous physical space according to the characteristic dimension of the time-series energy consumption data stream, and establish an addressing path from the network card direct memory access interface to the processor core to execute the operator in-place operation.
[0010] Preferably, the bypass scheduling module is used to monitor the data exchange frequency between the processor core and the energy carbon tensor management module by disabling hardware interrupts.
[0011] Preferably, the mapping processing module is used to calculate the weights of different non-zero elements in the topological association matrix when performing graph segmentation, and to identify the associated nodes whose weights exceed a preset weight threshold as topological boundary nodes; the kernel-mode topological boundary cut set is used to record the logical address, physical page frame number, and original physical page image index of the topological boundary node.
[0012] Preferably, the operating system includes a kernel-mode synchronization module for synchronizing the data content of the original physical page and the mirrored physical page; when any processor core performs a write operation on the original physical page, the kernel-mode synchronization module uses a multicast signal based on the hardware bus to synchronously update the corresponding mirrored physical page in all local memory nodes.
[0013] Preferably, the operating system includes a fluctuation sensing module for real-time detection of the data throughput of the time-series energy consumption data stream and for adjusting the scheduling priority of the bypass scheduling module according to the rate of change of the data throughput.
[0014] Preferably, the page table alias redirection path is used to point the local page table entries of different processor cores to different physical page mirrors of the same logical address, so that when different processor cores concurrently access data at the topology boundary node, the access actions are converged to their respective local memory buses, thereby eliminating hardware cache consistency probe signals.
[0015] Preferably, the mapping processing module is used to predict the data flow topology in the energy carbon tensor management module based on the preloaded deep learning computation graph, and preheat the physical memory pages involved in the data flow topology to the local memory node corresponding to the target processor core.
[0016] Preferably, when executing the global tensor prediction task, the operating system uses the mapping processing module to divide the carbon tracing operator into tasks based on the kernel-state topology boundary cut set, ensuring that each divided sub-operator obtains the data of the topology boundary node through the local memory node.
[0017] Compared to existing technologies, the AI-based intelligent computing operating system for energy and carbon data of this invention has the following advantages:
[0018] 1. In the intelligent computing operation of energy and carbon data in AI, a fluctuation-aware bypass scheduler resides in the kernel state and extracts the cache line failure values recorded in the hardware performance monitoring unit in real time. This enables passive physical perception of the state changes of the energy and carbon tensor memory pool. The traditional monitoring logic based on software polling and numerical comparison is transformed into a direct physical mirror reading of the state of the underlying hardware registers. Since the cache coherence protocol at the underlying hardware triggers the cache line failure of the corresponding physical address and causes the hardware counter value to increment when the network interface controller writes data different from the original state to the memory pool through direct memory access, the system does not need to use the central processing unit clock cycle to frequently traverse and read the numerical changes in the memory address. This eliminates the forced eviction of the L1 and L2 caches by the software-level polling observation, avoids cache pollution and bus congestion caused by the state detection link, and ensures that the processor's computing power resources can be exclusively used to perform tensor calculations.
[0019] 2. The operator mapping engine utilizes the page table aliasing mechanism of the memory management unit. When the topology state machine detects an interruption in the data update of a specific physical address segment, it issues a page table modification instruction to the memory management unit to redirect and anchor the failed virtual address to the physical page frame of the adjacent active node. This hardware-level addressing and reconstruction process enables the continuously running dedicated tensor computation thread to transparently pull substitute data and perform matrix operations through the original high-dimensional tensor pointer without being aware of the backend physical link disconnection. No operator overloading, weight unloading, or instruction pipeline flushing is required at the software application layer. This method solves the memory page reallocation and garbage collection overhead caused by frequent tensor dimension reconstruction in traditional solutions. Under extreme conditions of large-scale disconnection of edge nodes, it ensures the absolute continuity and determinism of instruction flow processing of complex simulation models.
[0020] 3. By extracting cross-regional boundary cut sets from the power grid topology correlation matrix and reconfiguring the hardware multicast function of the peripheral component interconnection standard root complex, the system synchronously distributes the received heterogeneous data streams from boundary nodes to the local physical memory of all nodes with non-consistent memory access that have cross-domain computational dependencies. Read-only shadow physical pages are generated in situ, reducing the high-frequency remote memory access required for cross-regional tensor interaction and that needs to pass through the hyperpath interconnection bus to ultra-fast addressing of local physical memory pages. This physically blocks high-frequency cache consistency probe signals across nodes. By eliminating cross-bus data pull actions and the associated bus protocol handshake overhead, the memory access bandwidth of different physical processor cores is converged within their respective local domains when processing strongly coupled global topology prediction, eliminating computing power starvation and latency nonlinear degradation caused by congestion of the underlying communication links. Attached Figure Description
[0021] Figure 1 This is a flowchart illustrating the overall architecture and memory access redirection process of the AI-powered carbon-based intelligent computing operating system of this invention.
[0022] Figure 2 This invention provides a logic diagram for identifying topological boundary nodes and generating kernel-state cut sets in carbon computing. Detailed Implementation
[0023] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.
[0024] It should be noted that all directional and positional terms used in this invention, such as: up, down, left, right, front, back, vertical, horizontal, inner, outer, top, bottom, transverse, longitudinal, center, etc., are only used to explain the relative positional relationship and connection between components in a specific state (as shown in the accompanying drawings). They are only for the convenience of describing this invention and do not require that this invention be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the descriptions of "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.
[0025] In the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.
[0026] In the description of this specification, references to the terms "an embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0027] An AI-based intelligent computing operating system for energy and carbon data includes:
[0028] The energy carbon tensor management module is used to allocate non-paged contiguous physical space in memory to store time-series energy consumption data streams that characterize the status of heterogeneous IoT terminals.
[0029] The bypass scheduling module is used to obtain the non-consistent memory access topology map of the physical server, identify the local memory nodes corresponding to different processor cores, and monitor the data exchange frequency between the processor cores and the energy carbon tensor management module in real time.
[0030] The mapping processing module is used to obtain the energy and carbon computing topology correlation matrix and perform graph partitioning, identify non-zero elements in the energy and carbon computing topology correlation matrix that represent cross-regional data interaction, extract topology boundary nodes connecting different computing regions, and generate kernel-mode topology boundary cut sets. The operating system is used to perform physical mirroring multi-projection mapping based on the non-consistent memory access topology graph and kernel-mode topology boundary cut sets, mirroring the original physical pages corresponding to the topology boundary nodes to multiple local memory nodes, and establishing page table alias redirection paths to transform cross-bus remote memory access into local memory page addressing access for the processor core.
[0031] Preferably, the energy carbon tensor management module is used to perform continuous physical space pre-allocation during the system startup phase, and the operating system performs memory attribute locking to prevent the continuous physical space from participating in virtual memory page swapping; the mapping processing module is used to address the input pointer of the computation graph operator to the continuous physical space according to the characteristic dimension of the time-series energy consumption data stream, and establish an addressing path from the network card direct memory access interface to the processor core to execute the operator in-place operation.
[0032] Preferably, the bypass scheduling module is used to monitor the data exchange frequency between the processor core and the energy carbon tensor management module by disabling hardware interrupts.
[0033] Preferably, the mapping processing module is used to calculate the weights of different non-zero elements in the topological association matrix when performing graph segmentation, and to identify the associated nodes whose weights exceed a preset weight threshold as topological boundary nodes; the kernel-mode topological boundary cut set is used to record the logical address, physical page frame number, and original physical page image index of the topological boundary node.
[0034] Preferably, the operating system includes a kernel-mode synchronization module for synchronizing the data content of the original physical page and the mirrored physical page; when any processor core performs a write operation on the original physical page, the kernel-mode synchronization module uses a multicast signal based on the hardware bus to synchronously update the corresponding mirrored physical page in all local memory nodes.
[0035] Preferably, the operating system includes a fluctuation sensing module for real-time detection of the data throughput of the time-series energy consumption data stream and for adjusting the scheduling priority of the bypass scheduling module according to the rate of change of the data throughput.
[0036] Preferably, the page table alias redirection path is used to point the local page table entries of different processor cores to different physical page mirrors of the same logical address, so that when different processor cores concurrently access data at the topology boundary node, the access actions are converged to their respective local memory buses, thereby eliminating hardware cache consistency probe signals.
[0037] Preferably, the mapping processing module is used to predict the data flow topology in the energy carbon tensor management module based on the preloaded deep learning computation graph, and preheat the physical memory pages involved in the data flow topology to the local memory node corresponding to the target processor core.
[0038] Preferably, when executing the global tensor prediction task, the operating system uses the mapping processing module to divide the carbon tracing operator into tasks based on the kernel-state topology boundary cut set, ensuring that each divided sub-operator obtains the data of the topology boundary node through the local memory node.
[0039] Example 1: In the infrastructure of a large-scale computing center with 1.25 million concurrent energy and carbon sensor terminals, the system faces the need for real-time processing of high-frequency time-series energy consumption data and carbon emission factors. Such operating conditions require the underlying operating system to provide microsecond-level deterministic latency and extremely high throughput. Traditional energy and carbon data processing architectures based on general-purpose kernels treat heterogeneous data streams as standard input and output byte streams. Data enters user space via standard network protocol stacks and virtual file systems. In scenarios with sudden high-frequency traffic such as power grid load changes, frequent context switching occurs, resulting in processor clock cycles being consumed by interrupt response and data copying. This causes the user-space artificial intelligence tensor calculation thread to face end-to-end computation latency due to computing power starvation and nonlinear deterioration. During the system startup phase, an energy and carbon tensor memory pool is pre-allocated in physical memory. This memory pool has a continuous, locked, and page-prohibited physical space. After receiving the heterogeneous data stream, the network interface controller uses the direct memory access mechanism to unpack the data stream according to the physical dimensions of voltage amplitude, active power, and carbon emission factors, and directly maps it to the energy and carbon tensor memory pool, avoiding the memory parsing and copying operations of the standard network protocol stack.
[0040] The fluctuation-aware bypass scheduler residing in kernel mode employs a physical mirror state monitoring mechanism based on the hardware performance monitoring unit. It periodically reads the CPU's hardware performance monitoring unit registers to obtain the cache line miss event count triggered by a specific physical address segment within a unit clock cycle. When the network interface controller writes new data to the memory pool via direct memory access, triggering a hardware-level cache coherency protocol and incrementing the hardware count, the fluctuation-aware bypass scheduler intercepts and masks standard kernel-triggered regular input / output hardware interrupts, waking up the dedicated AI tensor computation thread bound to the processor core within 8μs. During the system cold start phase, the fluctuation-aware bypass scheduler collects hardware characteristic benchmarks, reading the L3 cache miss register storing retired instruction values from the processor's hardware performance monitoring unit to obtain the cache line miss count sequence under idle conditions. The sampling window is set to 1024 clock cycles, and the arithmetic mean of the sequence is calculated. with standard deviation Determine the benchmark value Determined as The obtained cache line invalidation value and The ratio is used as a normalized input parameter for the carbon state information entropy H(E). This process filters out hardware noise caused by processor branch prediction errors and determines that the logic signal-to-noise ratio is maintained above 26dB.
[0041] The operator image engine running in the intelligent computing middleware layer points the computation graph operator input pointer of the artificial intelligence model to the physical address of the energy carbon tensor memory pool. This enables the awakened dedicated thread for artificial intelligence tensor computation to directly pull the latest energy carbon data from physical address 0x7F8A0000 and perform matrix operations. When a local power grid experiences a sudden load change, the system reduces the context switching overhead between kernel mode and user mode from 85,000 times per second to less than 340 times per second. The P99 tail latency of artificial intelligence model inference is stabilized from 145ms to 12.4ms. This achieves pass-through from direct writing to the network card to in-place operator operation, ensuring the deterministic computing throughput of ultra-large-scale energy carbon topology simulation tasks under extreme loads.
[0042] Example 2: This example verifies the addressing determinism and interrupt suppression capabilities of the intelligent computing operating system when processing high-frequency energy carbon data streaming. A server equipped with a dual-socket 32-core processor, 512GB of non-uniform memory access architecture, and a dual-port 100Gbps network interface controller was used as the test platform. A city-level virtual power plant topology data stream generated through computational fluid dynamics simulation was used as the data source. Its physical model was constructed based on the power grid flow balance equation. To simulate signal interference in a real industrial environment, Gaussian white noise with a signal-to-noise ratio of 15dB and 50Hz power frequency interference harmonics were superimposed on the input signal. The sampling period in the experiment was... The setting depends on the maximum update frequency of the carbon terminal. Minimum time slice for operating system kernel response To balance these factors, the sampling period is set according to the following proportional relationship: , where k is the sampling factor, and its value is not less than 2. Under the experimental conditions, k is set to 2.5, and the determined sampling frequency is 2.5MHz.
[0043] In the standard operating condition test with a data exchange frequency of 100kHz, the control group adopted the standard kernel scheduling mechanism, and the measured end-to-end processing latency fluctuated between 12.4ms and 18.6ms. Its latency curve showed random dispersion characteristics with the injection of background noise, indicating that the general protocol stack generated buffer overflow when processing small packet data mixed with noise. In contrast, the sample of this invention used the non-paged continuous physical space pre-allocated by the energy carbon tensor management module to stabilize the latency at 3.2ms and the jitter rate was less than 5%. When the data exchange frequency was increased to a high load state of 500kHz, the control group had an increased number of context switches due to frequent hardware interrupts, and the processor clock cycle was consumed by the interrupt handler, and the measured latency rose to 85.4ms. In contrast, the sample of this invention used a fluctuation-aware bypass scheduler to read the cache line failure count value in the hardware performance monitoring unit register in real time. The system used the lockless memory pool of the intelligent computing middleware layer to buffer noise disturbances and maintain a processing latency of 4.5ms.
[0044] When the data exchange frequency reached the performance inflection point of 1MHz, the control group experienced an interrupt storm, with latency surging to 312.4ms. 75% of processor utilization was consumed by kernel-mode soft interrupt execution. The fluctuation-aware bypass scheduler intercepted and masked hardware interrupts triggered by the standard kernel, binding the dedicated thread for AI tensor computation to the processor core. In actual tests, the latency converged and stabilized at 12.6ms. When the data exchange frequency reached the preset performance inflection point, the operating system used the advanced programmable interrupt controller mask register to mask external hardware interrupts to the processor core. Non-real-time system kernel tasks were migrated to the auxiliary management core for execution, and the bypass scheduling module locked the AI tensor computation thread. The dedicated thread is located on the target physical core. During interrupt masking, the dedicated thread occupies the instruction pipeline and L1 cache resources. A watchdog timer monitors the thread execution status. If the cycle of a single operator operation exceeds 100μs, the interrupt masking is released and a state rollback procedure is executed. After entering the 2MHz boundary state test, the control group system stalled because the bandwidth of the superpath interconnect bus was occupied by the cache coherency probe signal, and the packet loss rate exceeded 92%, showing a performance collapse. However, the prototype of this invention uses the operator mapping engine to directly address the operator input pointer to the physical address of the energy carbon tensor memory pool, avoiding the latency loss caused by cross-bus access. The measured latency is 18.5ms.
[0045] Example 3: In access In a city-level virtual power plant intelligent computing infrastructure with distributed energy nodes, when the energy topology expands due to the addition of energy storage equipment, the system faces the need to reconstruct the underlying memory pool while maintaining online tensor computation. General-purpose operating systems, lacking awareness mechanisms for heterogeneous hardware platforms, typically set their scheduling thresholds to static values, leading to erroneous bypass decisions when the hardware performance monitoring unit's baseline shifts. Furthermore, physical mirror mapping faces data inconsistency risks during topology mutations. The system performs baseline sampling based on the hardware performance monitoring unit. The fluctuation-aware bypass scheduler reads the cache line miss count from the processor registers in an idle state, with a sampling length of 1024 clock cycles, and calculates the arithmetic mean of the count values. with standard deviation The steady-state threshold of the system is determined as When the average value is measured 1250 times and standard deviation When the number of iterations is 45, the system determines the trigger baseline value to be 1385. The mapping processing module reads the energy carbon calculation topology correlation matrix and counts the number of non-zero elements in each row to obtain the node connectivity. The mapping processing module marks the node as a topology boundary node, extracts its physical page frame number, and writes it into the physical page image index table of the kernel-mode topology boundary cut set.
[0046] To maintain consistency in physical mirror multi-projection mapping, the kernel-mode synchronization module performs write-protected locking on the page table entries associated with the original physical pages. When a processor core writes data to the original physical page, the system intercepts the write instruction and triggers a multicast update protocol. Atomic swap instructions are used to synchronize the updated data stream to the mirrored physical pages distributed across different local memory nodes. After all mirrored physical pages return write completion signals, a global storage barrier instruction is issued to refresh the processor core's storage queue. This ensures that all processor cores read consistent energy and carbon state data in subsequent local addressing accesses, reducing the cross-node access conflict rate to below 0.08% under topology expansion-induced fluctuations. The processor core obtains a consistent energy and carbon state tensor by addressing local mirrored pages. It supports the deterministic execution of large-scale virtual power plant simulation tasks. The mapping processing module identifies the kernel-mode topology boundary cut set, calls the kernel-mode memory allocation interface to reserve a memory page frame in the local memory node of the target processor core with an amount equal to the original physical page, performs instruction-level atomic swap operations to mirror the data content of the original physical page to the local memory page frame, modifies the corresponding page table entry of the target processor core, redirects the virtual address to the physical address of the local memory page frame, sets the write protection bit of the page table entry, the processor core initiates a write request to trigger a page fault exception, the kernel-mode synchronization module intercepts the exception, calls the hardware multicast function of the peripheral component interconnect standard root complex to synchronously update the mirror copy, issues a global storage barrier instruction to refresh the storage queue, and ensures that the energy carbon state tensor read by each core remains consistent.
[0047] Example 4: In the energy and carbon data processing environment of a heterogeneous computing power cluster, the fluctuation-aware bypass scheduler reads the instruction count register and clock cycle register in the hardware performance monitoring unit during the system boot phase, extracts the processor's baseline flip-flop count under zero load, and calculates the hardware state reference vector. The fluctuation-aware bypass scheduler will combine the carbon state information entropy H(E) with the hardware state reference vector. The system correlates and compensates for cache response deviations in different processor architectures and calibrates judgment thresholds to achieve real-time filtering of noise from computing node monitoring.
[0048] The energy-carbon tensor management module receives the energy-carbon computational topological correlation matrix, statistically analyzes the distribution characteristics of non-zero elements in the matrix, and determines the average degree of the energy-carbon connectivity map. Total number of nodes and according to the formula Determine the size of the physical address space that needs to be locked, where, The physical size of the carbon tensor memory pool. The total number of nodes. The value is the average degree, and 64 represents the standard width of a cache line in bytes. The system allocates contiguous address segments in physical memory and marks them as non-pageable. The mapping module then injects the physical page frame number into the kernel page table, establishing a static binding between the topology logical connection and the physical memory access path. This suppresses the probability of address resolution failures under dynamic expansion conditions. Magnitude.
[0049] Example 5: When a computing node with a multi-processor architecture performs cold start and energy and carbon monitoring task initialization, the fluctuation-aware bypass scheduler initiates a benchmark calibration procedure based on hardware characteristics. It reads the instruction count register and clock cycle register in the hardware performance monitoring unit, accumulates the count values of 1024 cache line miss events in a 10μs time window under no-load conditions, and calculates the arithmetic mean of the count value sequence. with standard deviation and according to the relationship Determine the judgment benchmark value, where, To determine the benchmark value, This represents the average cache line invalidation count. As the standard deviation of the count value, the system uses the judgment benchmark value as the zero-point offset compensation term of the energy carbon state information entropy H(E), which is used to filter out random hardware noise caused by processor branch prediction errors and improve the signal-to-noise ratio of the judgment criterion to more than 26dB.
[0050] After the energy and carbon tensor management module obtains the topological correlation matrix of the virtual power plant, the mapping processing module establishes the addressing path through the atomic injection procedure of the physical address space. The mapping processing module then determines the addressing path based on the average degree of the energy and carbon connectivity graph. The system allocates physically contiguous page table space and uses atomic compare-and-swap instructions to write the physical page frame numbers corresponding to the identified topology boundary nodes into the reserved bits of the kernel-mode page table, establishing a static mapping from the topology logical connection to the physical memory access path. To address the performance bottleneck caused by the hardware cache coherence probe signal, the system forcibly disables the broadcast listening protocol for the specific physical memory address range 0x7F800000 to 0x7F8FFFFF by modifying the listening filter bit in the listening control register of the processor configuration space, preventing each processor core from sending listening broadcast signals through the bus. Simultaneously, the kernel-mode synchronization module calls the hardware of the peripheral component interconnect standard root complex. The multicast configuration register maps the multicast group to this address range, ensuring that when the main core performs a write operation, the hardware logic directly pushes the update payload to the shadow page frame of the remote local memory node through point-to-point data packets within 10ns. To ensure the visibility of data during the physical image multi-projection mapping process, the kernel-mode synchronization module issues a global storage barrier instruction after updating the image physical page, refreshing the processor core's storage buffer queue, so that the processor core reads a consistent energy carbon state tensor in local addressing access. When topology nodes are dynamically added or removed, the system completes the address space correction operation within 15μs, eliminating the suspension of computing resources caused by cross-node memory access conflicts.
[0051] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.
Claims
1. An AI-based intelligent computing operating system for energy and carbon data, characterized in that, include: The energy carbon tensor management module is used to allocate non-paged contiguous physical space in memory to store time-series energy consumption data streams that characterize the status of heterogeneous IoT terminals. The bypass scheduling module is used to obtain the non-consistent memory access topology map of the physical server, identify the local memory nodes corresponding to different processor cores, and monitor the data exchange frequency between the processor cores and the energy carbon tensor management module in real time. The mapping processing module is used to obtain the energy and carbon computing topology correlation matrix and perform graph partitioning, identify non-zero elements in the energy and carbon computing topology correlation matrix that represent cross-regional data interaction, extract topology boundary nodes connecting different computing regions, and generate kernel-mode topology boundary cut sets. The operating system is used to perform physical mirroring multi-projection mapping based on the non-consistent memory access topology graph and kernel-mode topology boundary cut sets, mirroring the original physical pages corresponding to the topology boundary nodes to multiple local memory nodes, and establishing page table alias redirection paths to transform cross-bus remote memory access into local memory page addressing access for the processor core.
2. The AI-based intelligent computing operating system for energy and carbon data according to claim 1, characterized in that, The energy carbon tensor management module is used to perform contiguous physical space pre-allocation during the system startup phase, and the operating system performs memory attribute locking to prevent contiguous physical space from participating in virtual memory page swapping; The mapping processing module is used to address the input pointers of the computation graph operators in a continuous physical space according to the characteristic dimensions of the time-series energy consumption data stream, and to establish an addressing path that directly writes from the network card direct memory access interface to the processor core to execute the operator in-place operation.
3. The AI-based intelligent computing operating system for energy and carbon data according to claim 1, characterized in that, The bypass scheduling module is used to monitor the data exchange frequency between the processor core and the energy carbon tensor management module by disabling hardware interrupts.
4. The AI-based intelligent computing operating system for energy and carbon data according to claim 1, characterized in that, The mapping processing module is used to calculate the weights of different non-zero elements in the topological association matrix when performing graph partitioning, and to identify the associated nodes whose weights exceed the preset weight threshold as topological boundary nodes; the kernel-mode topological boundary cut set is used to record the logical address, physical page frame number, and original physical page image index of the topological boundary node.
5. The AI-based intelligent computing operating system for energy and carbon data according to claim 1, characterized in that, The operating system includes a kernel-mode synchronization module, which is used to synchronize the data content of the original physical page and the mirror physical page. When any processor core performs a write operation on the original physical page, the kernel-mode synchronization module uses a multicast signal based on the hardware bus to synchronously update the corresponding mirror physical page in all local memory nodes.
6. The AI-based intelligent computing operating system for energy and carbon data according to claim 1, characterized in that, The operating system includes a fluctuation sensing module, which is used to detect the data throughput of the time-series energy consumption data stream in real time and adjust the scheduling priority of the bypass scheduling module according to the rate of change of the data throughput.
7. The AI-based intelligent computing operating system for energy and carbon data according to claim 1, characterized in that, Page table alias redirection paths are used to point local page table entries of different processor cores to different physical page mirrors of the same logical address. This ensures that when different processor cores concurrently access data at topology boundary nodes, the access actions are converged to their respective local memory buses, eliminating hardware cache coherency probes.
8. The AI-based intelligent computing operating system for energy and carbon data according to claim 1, characterized in that, The mapping processing module is used to predict the data flow topology in the energy carbon tensor management module based on the preloaded deep learning computation graph, and preheat the physical memory pages involved in the data flow topology to the local memory node corresponding to the target processor core.
9. The AI-based intelligent computing operating system for energy and carbon data according to claim 1, characterized in that, When executing global tensor prediction tasks, the operating system uses the mapping processing module to divide the carbon tracing operator into tasks based on the kernel-state topology boundary cut set, ensuring that each sub-operator obtains the data of the topology boundary node through the local memory node.