A dynamic dual-partition cache mapping method and system based on multi-dimensional feature classification

Abstract

The present invention discloses a dynamic dual-partition cache mapping method and system based on multidimensional feature classification. The method comprises cache partitions including a set-associative master region (SAMR), a direct-mapped region (DMR), and a fully associative region (FAR), and sets the initial capacity of each partition in the cache partition; collects multidimensional feature data of data blocks in real time, the multidimensional feature data including access frequency f, spatial locality s, prefetch benefit p, and temporal volatility t; inputs the multidimensional feature data into a weighted classifier module to calculate a comprehensive score Fscore of the data block; generates a migration trigger instruction based on the comprehensive score Fscore, and migrates the data block to the direct-mapped region (DMR) or the fully associative region (FAR) between dynamic regions. The present invention improves cache hit rate, reduces hardware overhead of fully associative mapping, optimizes cache efficiency by dynamically adjusting rules, and improves the accuracy of data migration decisions.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a dynamic dual-partition cache mapping method and system based on multi-dimensional feature classification. Background Art

[0002] Traditional implementations of cache physical mapping mechanisms have significant technical limitations. While direct mapping offers the advantages of simple structure and low hardware cost, its fixed-location storage strategy causes different main memory blocks to compete for the same cache location, leading to conflict misses. While fully associative mapping reduces the probability of conflicts by allowing data blocks to be placed arbitrarily, it requires parallel comparison of all cache line tags, significantly increasing hardware implementation complexity and cost. While set-associative mapping combines the advantages of the previous two approaches to some extent, it still struggles to achieve optimal hardware efficiency in practical applications. Existing dynamic matching scheme designs use a partitioning architecture to divide the cache into a set-associative main area, a direct-mapped area, and a fully-associative mapped area. The fixed ratio of the set-associative main area limits the system's adaptability to dynamic workloads, while the fixed ratio of the direct-mapped area may not effectively match actual cold data storage requirements. While the fully-associative mapped area is optimized for hot data, its adjustment mechanism suffers from responsiveness lags when dealing with sudden changes in access characteristics. While unified tag memories and reconfigurable address mapping tables in hardware co-design offer flexibility, they introduce additional hardware overhead and system complexity. The multi-dimensional feature scoring mechanism that data migration strategies rely on suffers from insufficient precision in threshold setting. This can lead to inaccurate data block value assessments and biased migration decisions, impacting the overall efficiency of cache resources. These technical flaws severely restrict the performance of cache systems in complex application scenarios. Summary of the Invention

[0003] In order to solve the technical problems existing in the background technology, the present invention proposes a dynamic dual-partition cache mapping method and system based on multi-dimensional feature classification.

[0004] The present invention proposes a dynamic dual-partition cache mapping method based on multi-dimensional feature classification, comprising:

[0005] Establishing cache partitions, the cache partitions including a set-associative master region SAMR, a direct-mapped region DMR, and a fully associative region FAR, and setting an initial capacity of each partition in the cache partition;

[0006] Real-time collection of multi-dimensional feature data of data blocks, the multi-dimensional feature data including but not limited to access frequency f, spatial locality s, pre-fetch benefit p, and temporal volatility t;

[0007] Input the multi-dimensional feature data into the weighted classifier module to calculate the comprehensive score Fscore of the data block;

[0008] A migration trigger instruction is generated according to the comprehensive score Fscore to migrate the data block to the direct mapping area DMR or the fully associated area FAR between dynamic areas.

[0009] Preferably, generating a migration trigger instruction according to the comprehensive score Fscore specifically includes:

[0010] If the comprehensive score Fscore>0.7, migrate to the fully associative mapping area FAR;

[0011] If the comprehensive score Fscore < the mean of the comprehensive scores Fscore corresponding to all data blocks in the current window minus the standard deviation, it is migrated to the direct mapping area DMR.

[0012] Preferably, the step of inputting the multi-dimensional feature data into a weighted classifier module and calculating the comprehensive score Fscore of the data block specifically includes:

[0013] The multi-dimensional feature data is weighted and summed according to the preset weight coefficient to obtain the comprehensive score Fscore of the data block. The calculation formula is as follows:

[0014] Fscore=w1×f+w2×g+w3×p+w4×s+w5×t;

[0015] Among them, w1-w5 are weight coefficients, which are optimized and determined by grid search method under typical load; f is the access frequency; s is the spatial locality; p is the prefetch benefit; t is the temporal volatility.

[0016] Preferably, the initial capacity of the set associative main region SAMR is 70% to 90% of the system memory, and the initial capacity of the direct mapping region DMR and the fully associative region FAR are 5% to 15% of the system memory respectively.

[0017] Preferably, it also includes:

[0018] Monitor the conflict rate C of the group-associative main area SAMR, the miss rate M of the direct mapping area DMR, and the hit rate H of the fully associative mapping area FAR in real time, and dynamically adjust the corresponding capacity of each partition in the cache partition according to the conflict rate C, miss rate M, hit rate H and the preset dynamic adjustment strategy.

[0019] Preferably, the preset dynamic adjustment strategy specifically includes:

[0020] When the conflict rate C>25%, the capacity of the fully associative mapping area FAR will be expanded by 1%, and the capacity of the direct mapping area DMR will be reduced by 1%, with the highest priority;

[0021] When the miss rate M>40%, the capacity of the direct mapping area DMR will be expanded by 1%, and the capacity of the fully associative mapping area FAR will be reduced by 1%, with the priority being second;

[0022] When the hit rate H is less than 70%, the capacity of the fully associative mapping area FAR is reduced by 1%, and the capacity of the set-associative main area SAMR is expanded by 1%, with the lowest priority.

[0023] Preferably, when the adjustment conditions of the conflict rate C>25% and the miss rate M>40% are triggered simultaneously, the adjustment operation corresponding to the conflict rate C>25% is performed preferentially.

[0024] Preferably, the dynamic area ratio adjustment of the preset dynamic adjustment strategy is subject to an upper limit and a lower limit, and the ratio range of the fully associative mapping area FAR to the direct mapping area DMR is 5%-15% respectively.

[0025] Preferably, it also includes:

[0026] When the cache space is insufficient, a replacement strategy is used to decide whether to eliminate one or more data blocks in the dynamic area. The replacement strategy is specifically as follows:

[0027] P=α*LRU_order-β*fnorm;

[0028] Among them, α is the weight of the LRU order, which controls the importance of recent access; β is the weight of the normalized access frequency, which controls the contribution of long-term popularity; LRU_order represents the "freshness" of the data block, 0 means the most recent access, and a larger value means the longer it has not been accessed; fnorm is the normalized access frequency, fnorm=f / fmax, where f is the number of times a data block is accessed, and fmax is the maximum number of times all data blocks in the current window are accessed.

[0029] The present invention proposes a dynamic dual-partition cache mapping system based on multi-dimensional feature classification, comprising:

[0030] A cache partition module is used to establish cache partitions, wherein the cache partitions include a set associative main region SAMR, a direct mapping region DMR and a fully associative region FAR, and set the initial capacity of each partition in the cache partition;

[0031] A multi-dimensional feature acquisition module, configured to acquire multi-dimensional feature data of a data block in real time, wherein the multi-dimensional feature data includes but is not limited to access frequency f, spatial locality s, prefetch benefit p, and temporal volatility t;

[0032] The weighted classification module is used to input the multi-dimensional feature data into the weighted classifier module and calculate the comprehensive score Fscore of the data block;

[0033] The dynamic partition control module is used to generate a migration trigger instruction according to the comprehensive score Fscore, and migrate the data block to the direct mapping area DMR or the fully associated area FAR between dynamic areas.

[0034] In the present invention, a dynamic dual-partition cache mapping method and system based on multi-dimensional feature classification is proposed. By establishing a set-associative main region SAMR, a direct mapping region DMR, and a fully associative region FAR, and collecting the memory access characteristics of data blocks in real time, including access frequency, spatial locality, prefetching efficiency, and time volatility, a weighted classifier is used to calculate the comprehensive score Fscore of each data block, and the capacity of DMR and FAR is dynamically adjusted. It can effectively solve the conflict miss problem caused by the direct mapping fixed-position storage strategy, optimize the partition ratio through a dynamic adjustment mechanism, and improve the cache hit rate; overcome the defects of complex implementation and high hardware cost of fully associative mapping, realize intelligent scheduling through multi-dimensional feature collection and weighted classifier, and reduce hardware overhead; improve the deficiency of group associative mapping in not achieving optimal matching, respond to workload changes in real time through a dynamic partition control module, and improve system adaptability; optimize the fixed ratio setting of direct mapping area and fully associative mapping area, respond to sudden access feature changes in time through dynamic adjustment rules, and improve cache efficiency; reduce the additional overhead and complexity problems in hardware collaborative design, and achieve efficient management through a unified tag memory and a reconfigurable address mapping table; improve the accuracy of data migration strategy, accurately reflect the actual value changes of data blocks through multi-dimensional feature scoring and weighting strategy, and optimize migration decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a schematic diagram of the workflow of a dynamic dual-partition cache mapping method based on multi-dimensional feature classification proposed by the present invention. DETAILED DESCRIPTION

[0036] Reference Figure 1 The present invention proposes a dynamic dual-partition cache mapping method based on multi-dimensional feature classification, comprising the following steps:

[0037] S1: Create a cache partition, which includes a set-associative main region SAMR, a direct-mapped region DMR, and a fully associative region FAR, and set the initial capacity of each partition in the cache partition.

[0038] In this embodiment, the initial capacity of the set-associative main region SAMR is 70% to 90% of the system memory, and the initial capacities of the direct-mapped region DMR and the fully-associative region FAR are 5% to 15% of the system memory, respectively.

[0039] Multi-dimensional feature data of data blocks are collected in real time. The multi-dimensional feature data includes but is not limited to access frequency f, spatial locality s, prefetch benefit p, and time volatility t.

[0040] In this embodiment, the specific attributes of the multi-dimensional feature data are as follows:

[0041] (1) Access frequency f:

[0042] Definition: The number of times a data block is accessed within the unit time window Twindows.

[0043] Formula: f=number of visits / Twindows;

[0044] Hardware implementation: Each cache line has a built-in 4-bit sliding window counter that is automatically updated every cycle.

[0045] (2) Spatial locality s:

[0046] Definition: Reflects the access association between a data block and adjacent addresses.

[0047] Formula: s = number of adjacent block accesses / total number of accesses;

[0048] Hardware implementation: The address comparator detects accesses in the ±k address range, and the counter accumulates the number of hits.

[0049] (3) Prefetch benefit p:

[0050] Definition: The proportion of data blocks loaded by the prefetch mechanism that are subsequently accessed.

[0051] Formula: p = number of prefetch trigger hits / total number of prefetches

[0052] Hardware implementation: The prefetch operation marks the prefetch bit, and triggers the prefetch counter when it hits.

[0053] (4) Time volatility t:

[0054] Definition: The stability of access time intervals. The smaller the standard deviation, the more regular the access pattern.

[0055] formula:

[0056] The time difference between the i-th access and the i-1-th access of data block X (unit: cycle); σ interval is the visit interval standard deviation: ≈ Example: If the interval is [5, 3], then σ≈(5−3) / 2=1. interval is the average access interval: ,Example: The interval is [5, 3], then μ interval= 4; Δt is the time interval between consecutive accesses within the window; ε (e.g., 1e-5) is a minimum constant to prevent division by zero, which is obtained and stored in a register. Value source: max(Δt) and min(Δt) are recorded in the extreme value registers for the maximum and minimum intervals, and the difference approximates the standard deviation. For example, t = 1-1 / (4+1e-5) ≈ 0.75.

[0057] S2: Input the multi-dimensional feature data into the weighted classifier module to calculate the comprehensive score Fscore of the data block.

[0058] In this embodiment, the multi-dimensional feature data is input into the weighted classifier module to calculate the comprehensive score Fscore of the data block, which specifically includes:

[0059] S3: Perform weighted summation on the multi-dimensional feature data according to the preset weight coefficient to obtain the comprehensive score Fscore of the data block. The specific calculation formula is:

[0060] Fscore=w1×f+w2×g+w3×p+w4×s+w5×t;

[0061] Among them, w1-w5 are weight coefficients, which are optimized and determined by grid search method under typical load; f is the access frequency; s is the spatial locality; p is the prefetch benefit; t is the temporal volatility.

[0062] For example: w1=0.4,w2=0.2,w3=0.1,w4=0.2,w5=0.1;

[0063] If f=0.12, g=0.5, p=0.25, s=0.4, t=0.5, then:

[0064] Fscore=0.4×0.12+0.2×0.5+0.1×0.25+0.2×0.4+0.1×0.5=0.048+0.1+0.025+0.08+0.05=0.303;

[0065] Because Fscore=0.303<0.5, it is mapped to DMR.

[0066] S4: Generate a migration trigger instruction based on the comprehensive score Fscore, and migrate the data block to the direct mapping area DMR or the fully associative area FAR between dynamic areas.

[0067] In this embodiment, the migration trigger instruction is generated according to the comprehensive score Fscore, specifically including:

[0068] If the comprehensive score Fscore>0.7, migrate to the fully associative mapping area FAR;

[0069] If the comprehensive score Fscore < the mean of the comprehensive scores Fscore corresponding to all data blocks in the current window minus the standard deviation, it is migrated to the direct mapping area DMR.

[0070] In this embodiment, it also includes:

[0071] S5: Monitor the conflict rate C of the group-associative main area SAMR, the miss rate M of the direct mapping area DMR, and the hit rate H of the fully associative mapping area FAR in real time, and dynamically adjust the corresponding capacity of each partition in the cache partition according to the conflict rate C, miss rate M, hit rate H and the preset dynamic adjustment strategy.

[0072] Specifically, the preset dynamic adjustment strategy includes:

[0073] When the conflict rate C>25%, the capacity of the fully associative mapping area FAR will be expanded by 1%, and the capacity of the direct mapping area DMR will be reduced by 1%, with the highest priority;

[0074] When the miss rate M>40%, the capacity of the direct mapping area DMR will be expanded by 1%, and the capacity of the fully associative mapping area FAR will be reduced by 1%, with the priority being second;

[0075] When the hit rate H is less than 70%, the capacity of the fully associative mapping area FAR is reduced by 1%, and the capacity of the set-associative main area SAMR is expanded by 1%, with the lowest priority.

[0076] It should be noted that when the adjustment conditions of the conflict rate C>25% and the miss rate M>40% are triggered at the same time, the adjustment operation corresponding to the conflict rate C>25% is performed first.

[0077] In this embodiment, the dynamic area ratio adjustment of the preset dynamic adjustment strategy is limited by an upper limit and a lower limit, and the ratio ranges of the fully associative mapping area FAR and the direct mapping area DMR are 5%-15% respectively.

[0078] Specifically, the priority and execution restrictions of the trigger conditions need to be considered in this process. Priority rule: C>M>H: Prioritize processing when the conflict rate is too high to avoid performance avalanche. Execution restrictions are controlled by single-cycle single-action, and only one adjustment action is executed per clock cycle to avoid multiple triggers causing the ratio to be out of control. A bounded protection mechanism is also required to handle the situation when the upper and lower limits are reached. If the FAR has reached 15%, even if C>25%, further expansion is prohibited, and only the replacement strategy is optimized (such as increasing the FAR elimination priority). The processing when the lower limit is reached is as follows:

[0079] If the FAR reaches 5%, even if H is less than 70%, further scaling is prohibited and only the prefetch strategy is optimized.

[0080] It should be noted that the boundary constraints:

[0081] FAR ratio range: 5%~15%;

[0082] DMR ratio range: 5%~15%;

[0083] SAMR ratio: 70%~90%;

[0084] If C >25% and M >40% are triggered simultaneously, C >25% is prioritized (FAR increases by 1%, DMR decreases by 1%)

[0085] The hardware implementation associates a 2-bit flag with each cache line (00 = SAMR, 01 = DMR, 10 = FAR). During rebalancing, low-heat data (rows with the lowest F-score) is migrated. The flag is adjusted to determine which partition the data block belongs to. Partitioning doesn't change the physical properties of the cache; it essentially changes the mapping method. Therefore, partitioning logic can be implemented based on changes in the data block flag.

[0086] In this embodiment, the following data collection is achieved through the multi-dimensional feature collection circuit:

[0087] Access frequency counter: 4-bit saturating counter to prevent overflow.

[0088] Neighboring block detector: compares block access signals of addresses ±1 in parallel.

[0089] Extreme value register: records the maximum / minimum access interval within the window.

[0090] Dynamic partition adjustment circuitry: A reconfigurable address mapping table with a 2-bit flag encoding the region each cache line belongs to. A dedicated state machine circuit generates new scale values ​​based on trigger conditions, thereby adjusting the scale logic.

[0091] Specifically, the reconstructable address mapping table uses a 2-bit flag encoding:

[0092] 00: SAMR (set associative);

[0093] 01: DMR (direct mapping);

[0094] 10: FAR (Fully Associative);

[0095] Migration process: Select the target row from the reduced area, modify the flag, and update the label.

[0096] In this embodiment, it also includes:

[0097] S6: When the cache space is insufficient, a replacement strategy is used to decide whether to eliminate one or more data blocks in the dynamic area. The replacement strategy is as follows:

[0098] P=α*LRU_order-β*fnorm;

[0099] Among them, α is the weight of the LRU order, which controls the importance of recent access; β is the weight of the normalized access frequency, which controls the contribution of long-term popularity; LRU_order represents the "freshness" of the data block, 0 means the most recent access, and a larger value means the longer it has not been accessed; fnorm is the normalized access frequency, fnorm=f / fmax, where f is the number of times a data block is accessed, and fmax is the maximum number of times all data blocks in the current window are accessed.

[0100] In this embodiment, the SAMR group associative mapping index is:

[0101] ;

[0102] Example: If the block size is 64B, 2 6 =64, requires 6-bit address, i.e. address bits 5-0. Number of SAMR groups Nsets=1024=2 10 , a 10-bit address is required. Therefore, address bits 15-6 are used for the group index.

[0103] In this embodiment, the DMR directly maps the index:

[0104] ;

[0105] Example: If the DMR capacity is 64KB (1024 rows), then Ndirect = 1024.

[0106] In this embodiment, the FAR fully associative mapping index compares the 32-bit Tag fields of all rows in parallel.

[0107] Reference Figure 1 The present invention proposes a dynamic dual-partition cache mapping system based on multi-dimensional feature classification, comprising:

[0108] A cache partition module is used to establish cache partitions, which include set-associative master regions SAMR, direct-mapped regions DMR, and fully associative regions FAR, and set the initial capacity of each partition in the cache partition;

[0109] A multi-dimensional feature acquisition module, which is used to collect multi-dimensional feature data of data blocks in real time. The multi-dimensional feature data includes but is not limited to access frequency f, spatial locality s, prefetch benefit p, and time volatility t;

[0110] The weighted classification module is used to input the multi-dimensional feature data into the weighted classifier module and calculate the comprehensive score Fscore of the data block;

[0111] The dynamic partition control module is used to generate a migration trigger instruction according to the comprehensive score Fscore, and migrate the data block to the direct mapping area DMR or the fully associated area FAR between dynamic areas.

[0112] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A dynamic partition cache mapping method based on multi-dimensional feature classification, characterized in that: include: Establishing a cache partition, wherein the cache partition includes a set-associative main region SAMR, a direct mapping region DMR, and a fully associative mapping region FAR, and setting an initial capacity of each partition in the cache partition; Real-time collection of multi-dimensional feature data of data blocks, wherein the multi-dimensional feature data includes access frequency f, spatial locality s, prefetch benefit p, and temporal volatility t; Input the multi-dimensional feature data into the weighted classifier module to calculate the comprehensive score Fscore of the data block; Generate a migration trigger instruction based on the comprehensive score Fscore to migrate the data block to the direct mapping area DMR or the fully associative mapping area FAR; Also includes: Monitor the conflict rate C of the group-associative master region SAMR, the miss rate M of the direct-mapped region DMR, and the hit rate H of the fully-associative mapped region FAR in real time, and dynamically adjust the corresponding capacity of each partition in the cache partition based on the conflict rate C, miss rate M, hit rate H, and a preset dynamic adjustment strategy; The preset dynamic adjustment strategy specifically includes: When the conflict rate C>25%, the capacity of the fully associative mapping area FAR will be expanded by 1%, and the capacity of the direct mapping area DMR will be reduced by 1%, with the highest priority; When the miss rate M>40%, the capacity of the direct mapping area DMR will be expanded by 1%, and the capacity of the fully associative mapping area FAR will be reduced by 1%, with the priority being second; When the hit rate H is less than 70%, the capacity of the fully associative mapping area FAR is reduced by 1%, and the capacity of the set-associative main area SAMR is expanded by 1%, with the lowest priority.

2. The dynamic partition cache mapping method based on multi-dimensional feature classification according to claim 1 is characterized in that: The multi-dimensional feature data is input into the weighted classifier module to calculate the comprehensive score Fscore of the data block, which specifically includes: The multi-dimensional feature data is weighted and summed according to the preset weight coefficient to obtain the comprehensive score Fscore of the data block. The calculation formula is as follows: Fscore=w1×f+w2×g+w3×p+w4×s+w5×t; Among them, w1, w2, w3, w4 and w5 are weight coefficients, which are optimized and determined by the grid search method under typical load; f is the access frequency; s is the spatial locality; p is the prefetch benefit; and t is the temporal volatility.

3. The dynamic partition cache mapping method based on multi-dimensional feature classification according to claim 1, characterized in that: The initial capacity of the set associative main region SAMR is 70%-90% of the system memory, and the initial capacities of the direct mapping region DMR and the fully associative mapping region FAR are 5%-15% of the system memory respectively.

4. The dynamic partition cache mapping method based on multi-dimensional feature classification according to claim 1, characterized in that: The dynamic area ratio adjustment of the preset dynamic adjustment strategy is subject to upper and lower limits, and the ratio ranges of the fully associative mapping area FAR and the direct mapping area DMR are 5%-15% respectively.

5. The dynamic partition cache mapping method based on multi-dimensional feature classification according to claim 1, characterized in that: Also includes: When the cache space is insufficient, a replacement strategy is used to decide whether to eliminate one or more data blocks in the dynamic area. The replacement strategy is specifically as follows: P=α*LRU_order-β*fnorm; Among them, α is the weight of the LRU order, which controls the importance of recent access; β is the weight of the normalized access frequency, which controls the contribution of long-term popularity; LRU_order represents the freshness of the data block, 0 means the most recent access, and a larger value means the longer it has not been accessed; fnorm is the normalized access frequency, fnorm=f / fmax, where f is the number of times a data block is accessed, and fmax is the maximum number of times all data blocks in the current window are accessed.

6. A dynamic partition cache mapping system based on multi-dimensional feature classification, characterized in that: The system is applied to the dynamic partition cache mapping method based on multidimensional feature classification according to any one of claims 1 to 5, comprising: A cache partition module is used to establish cache partitions, wherein the cache partitions include a set associative main region SAMR, a direct mapping region DMR and a fully associative mapping region FAR, and set the initial capacity of each partition in the cache partition; A multi-dimensional feature acquisition module is used to acquire multi-dimensional feature data of data blocks in real time, wherein the multi-dimensional feature data includes access frequency f, spatial locality s, prefetch benefit p and temporal volatility t; The weighted classification module is used to input the multi-dimensional feature data into the weighted classifier module and calculate the comprehensive score Fscore of the data block; The dynamic partition control module is used to generate a migration trigger instruction according to the comprehensive score Fscore to migrate the data block to the direct mapping area DMR or the fully associative mapping area FAR.

Images