A hybrid load-oriented FDP SSD recycling unit granularity dynamic adjustment method

Abstract

This invention claims protection for a dynamic adjustment method for FDP SSD recycling unit granularity for mixed loads, relating to the field of computer storage technology. The method is divided into two collaborative layers: a data acquisition and extraction analysis layer and a modeling and execution layer. The data acquisition and extraction analysis layer is used to build an FDP simulation environment, extract multi-dimensional load characteristics, quantify the coupling relationship between RU granularity and multiple performance indicators, and output a load-RU granularity adaptation scoring model. The modeling and execution layer, based on the scoring model, achieves adaptive dynamic switching of RG-level RU granularity through load-aware granularity adjustment trigger rules, hierarchical RU granularity selection, and a multi-indicator weighted decision algorithm.

Description

Technical Field

[0001] This invention relates to the field of computer storage technology, and further focuses on the performance optimization of solid-state drives (SSDs). In particular, it proposes a method for dynamically adjusting the granularity of FDP SSD recycling units for mixed workloads, specifically addressing the resource scheduling and space management issues of flexible data placement solid-state drives (FDP SSDs). Background Technology

[0002] With the rapid development of emerging applications such as cloud computing, artificial intelligence, and big data analytics, the global data volume is experiencing explosive growth. According to IDC forecasts, the global data volume will reach 527.47 ZB in 2029. While this massive data volume is driving huge demand in the storage market, storage challenges are becoming increasingly prominent. On the one hand, the rate of increase in storage capacity lags significantly behind the growth rate of computing power, while data read / write speeds struggle to match the requirements of high-speed computing, resulting in slow inference response times and high costs. Storage issues are gradually becoming a key bottleneck restricting the practical application of large models. On the other hand, with the improvement of storage utilization, SSD-based storage systems are directly affected by the Web Amplification Factor (WAF), generally leading to reduced performance, shortened lifespan, and a sharp increase in operating costs. Data placement technology, as a low-cost solution for optimizing WAF, has received widespread attention from academia and industry. However, existing mainstream data placement technologies, such as Open-Channel SSDs and Partition Namespace (ZNS) SSDs, require deep modifications to the data model and software stack on the host side to reduce WAF, resulting in insufficient backward compatibility and high implementation difficulty.

[0003] The emergence of Flexible Data Placement (FDP) technology offers a groundbreaking solution to the aforementioned problems. FDP technology enables deep collaboration between hosts and storage devices without disrupting the existing NVMe ecosystem, simplifying ecosystem integration and providing excellent backward compatibility. Specifically, FDP SSDs expose device configuration information and data classification and placement interfaces to the host, allowing the host to understand the internal resource topology and effectively distribute data using its data placement interface, thereby reducing the impact of Web Application Firewall (WAF). These optimizations make it a potential next-generation enterprise-grade SSD standard. However, FDP SSDs use Reclaimed Units (RUs) as the basic management unit, and their granularity directly affects space management accuracy and parallel efficiency. Larger RU capacities result in lower erase frequencies and stronger parallel throughput, but coarse-grained space management leads to data mixing, causing more data to be moved during garbage collection (GC). In current commercial and engineering applications of FDP SSDs, storage vendors only provide data placement interfaces, while the internal structure and strategies are usually hardcoded. This opaque hardware behavior leads to significant differences in test results among different vendors' commercial products. In summary, fixed RU granularity solutions have two major drawbacks:

[0004] First, the optimization effect of fixed RU granularity depends entirely on the host-side upper-layer strategy, resulting in extremely poor robustness. Its performance difference depends entirely on the file system data classification strategy and the actual load skew. Once the data classification is inaccurate, the optimization effect of FDP will be limited, or even degrade to the level of traditional SSDs.

[0005] Secondly, fixed RU granularity cannot adapt to differentiated business scenarios and mixed loads, limiting its scenario adaptability. Different RU granularity sizes are suitable for different scenarios; fine-grained RUs can achieve more refined data isolation and reduce WAF, while coarse-grained RUs can better leverage parallel capabilities. The fixed RU granularity strategy of existing equipment cannot cope with complex situations and mixed loads.

[0006] The aforementioned problems cannot be fundamentally solved by simply using a fixed RU granularity strategy and a host-side data classification and adjustment strategy. They require a load-aware, dynamic RU granularity adjustment method adapted to the characteristics of the FDP architecture. Therefore, through preliminary research and literature review, it is easy to see that existing research on the impact of RU granularity size on performance has significant limitations, which are elaborated below:

[0007] 1. Commercial-grade FDP SSDs use hard-coded RU granularity configurations. Currently, although storage vendors offer commercial-grade FDP SSDs, their RU granularity is a factory-fixed static value, which does not support on-demand configuration on the host side or dynamic adjustment at runtime. It can only achieve basic data distribution through open data placement interfaces. Furthermore, the WAF optimization effect, GC efficiency, throughput performance, and other characteristics of FDP SSDs are strongly correlated with the data heat distribution, access skew, and read / write ratio of the business load. A fixed RU granularity can only achieve limited optimization effects in a single specific load scenario. Faced with the dynamic mixed loads widely present in enterprise-level scenarios, optimization effects are prone to failure, and may even lead to performance fluctuations and accelerated device lifespan degradation.

[0008] 2. Existing research on RU granularity only considers single-dimensional performance impact analysis. Most existing studies focus solely on the impact of RU granularity on a single core performance indicator such as WAF or latency, failing to systematically analyze the coupling relationship and boundaries between RU granularity and multiple performance indicators such as GC overhead, device parallelism, and I / O tail latency. They cannot quantify the optimal adaptation range of RU granularity under different load characteristics, making it difficult to achieve synergistic optimization of multiple performance indicators and easily leading to an imbalance where optimizing one indicator results in the deterioration of others.

[0009] To address this, this invention proposes a dynamic adjustment method for FDP SSD recycling unit granularity for mixed loads. Addressing the shortcomings of existing technologies, such as poor adaptability of static granularity configuration and insufficient multi-dimensional index collaborative optimization capabilities, this method first extracts and analyzes load characteristics through simulation, and then models the impact of RU granularity size on performance. Finally, it constructs three sub-modules: a load-aware dynamic adjustment method for RU granularity, triggering rules, hierarchical RU granularity selection set construction, and a multi-index weighted decision algorithm. This enables the RU granularity to adaptively adjust with load characteristics, solving the problems of adaptability failure and WAF spikes caused by fixed RU granularity under varying load scenarios.

[0010] A search revealed application publication number CN121070287A, which discloses a scheduling method, apparatus, data processor, and device for NVMe solid-state drives (SSDs), relating to the field of data processor technology. The method includes: receiving a request to create a virtual block device (VPD); allocating a target NVMe SSD that meets the requirements; if the target NVMe SSD supports FreeDepth Deployment (FDP) functionality, selecting multiple placement identifiers to be allocated from the target NVMe SSD's placement identifier list; and creating a VPD that supports FDP functionality based on the multiple placement identifiers and logical block address ranges; if the target NVMe SSD does not support FDP functionality, creating a VPD that does not support FDP functionality. Through this method, the data processor transparently maps and exposes the FDP capabilities of the physical NVMe SSD on the DPU side to the front-end VPD, enabling the host side to correctly use FDP technology on the VPD, achieving data placement control consistent with the physical disk's capabilities.

[0011] Based on the above analysis, the two solutions differ fundamentally in their technical focus and approach. The core of the comparative document lies in transparently transmitting the physical FDP SSD placement capability to the host through the virtualization layer, addressing control plane coordination and ecosystem compatibility issues. Its optimization effect highly depends on the accuracy of the host file system's data classification strategy, and the granularity of the recycling units (RUs) within the device remains a factory-fixed static configuration, unable to dynamically change with load. When the host classification strategy is inaccurate or the business load fluctuates drastically, the FDP optimization effect under this solution will severely degrade, even causing WAF spikes and performance fluctuations. In contrast, this invention focuses on adaptive data plane optimization, fundamentally breaking through the architectural limitations of fixed RU granularity. This invention first avoids the robustness defects of host-dependent strategies by extracting I / O load characteristics in real time at the device end. Then, it constructs load-aware dynamic adjustment trigger rules and a hierarchical RU granularity selection set, enabling the system to perceive the phased changes in mixed loads and proactively trigger re-decision. Finally, it introduces a multi-index weighted decision algorithm to perform collaborative quantitative evaluation among mutually constraining indicators such as WAF, GC migration volume, parallelism, and tail latency, selecting the optimal granularity configuration under the current load characteristics. Thus, this invention solves the technical problems of static granularity configuration failing to adapt to mixed load scenarios and the imbalance of multi-dimensional performance indicators, improving the performance stability and lifespan of FDP SSDs in dynamic environments. Summary of the Invention

[0012] This invention addresses the problems of existing technologies being unable to adapt to FDP architectures, failing to consider the significant differences in fixed RU granularity strategies under different loads, and evaluating the impact of RU granularity on performance based on only a single-dimensional indicator. It proposes a method for dynamically adjusting the granularity of FDP SSD recycling units for mixed loads. The technical solution of this invention is as follows:

[0013] A dynamic adjustment method for FDP SSD recycling unit granularity for mixed loads is proposed, which is divided into two collaborative layers: a data acquisition and extraction analysis layer and a modeling and execution layer. The data acquisition and extraction analysis layer is used to build an FDP simulation environment, extract multi-dimensional load features, quantify the coupling relationship between RU granularity and multiple performance indicators, and output a load-RU granularity adaptation scoring model. The modeling and execution layer, based on the scoring model, achieves adaptive dynamic switching of RG-level RU granularity through load-aware granularity adjustment trigger rules, hierarchical RU granularity selection set, and multi-indicator weighted decision algorithm.

[0014] Furthermore, the data acquisition and extraction analysis layer execution steps include:

[0015] S1. Based on the simulation platform, build an FDP SSD simulation environment, using only the RU granularity as a single controllable variable, to reproduce RG / RUH data isolation, RU-level GC, and space management mechanisms;

[0016] S2. Extract four core load characteristics: data popularity, number of concurrent streams, access skew, and invalidation intensity, and complete the standardized classification of load scenarios;

[0017] S3. Quantify the variation patterns of WAF, latency, and throughput under various load conditions, including hot skew, cold-hot balanced weak skew, and high invalidation of cold data, with different RU granularities.

[0018] S4. Construct a load characteristic and RU granularity adaptation scoring model, and output the optimal RU granularity, granularity adaptation range and boundary value for each load scenario as the input basis for modeling and execution layers.

[0019] Furthermore, the modeling and execution layer comprises three main sub-modules: a load-aware granularity adjustment triggering sub-module, a hierarchical RU granularity selection set construction sub-module, and a multi-index weighted decision algorithm sub-module;

[0020] The load sensing granularity adjustment triggering submodule is set with multiple constraints, and the granularity adjustment process can only be entered if all of them are met.

[0021] The hierarchical RU granularity selection set construction submodule defines the effective upper and lower limits of RU granularity and divides the candidate intervals of fine, medium and coarse granularity.

[0022] The multi-index weighted decision algorithm submodule allocates dynamic weights based on the analytic hierarchy process, calculates the comprehensive adaptation score of each candidate granularity and adds stability verification, and selects the optimal RU granularity to complete the configuration update within RG.

[0023] Furthermore, the multiple constraints of the load-aware granularity adjustment triggering submodule include:

[0024] Trigger threshold; Set a load change rate threshold so that RU granular adjustment is only initiated when the load characteristics change significantly beyond the threshold, thus filtering out small load fluctuations and avoiding performance degradation caused by frequent adjustments.

[0025] RG-level isolation; RU granularity adjustment must follow the RG grouping isolation rules, that is, cross-RG operations are strictly prohibited to prevent damage to the RG-level GC isolation capability;

[0026] Delaying RU granularity changes: First, during GC operation, the RU adjustment operation corresponding to RG needs to be delayed until the GC process ends; second, when the device is in a high-load environment, RU granularity adjustment needs to be temporarily suspended.

[0027] Furthermore, the execution process of the hierarchical RU granularity selection set construction submodule includes:

[0028] 1) Define the RU granularity boundary; all granularities within the interval are integer multiples of the page size. Experimentally, find the minimum effective granularity that enables WAF optimization as the lower bound R. min And the maximum effective granularity for maximizing parallelism is bounded by R. max R min and R max These are the upper and lower limits of granularity set based on experimental measured data and simulation platform parameters, respectively.

[0029] 2) Construct a three-level hierarchical RU granularity selection set for load adaptation; while ensuring that the granularity range is within the defined RU granularity boundaries, divide the RU interval into fine-grained, medium-grained, and coarse-grained intervals.

[0030] The fine-grained RU interval range is proposed to be [R min 0.5×R basic ], R basic This refers to the default settings adopted by most emulators based on research;

[0031] The proposed range for the medium-grained RU interval is [0.5×R]. basic , 2×R basic ];

[0032] The range of the coarse-grained RU interval is tentatively defined as [2×R]. basic , R max ].

[0033] Furthermore, the execution steps of the multi-index weighted decision algorithm submodule include:

[0034] WAF, I / O latency, and throughput are selected as the three core evaluation indicators. WAF and latency are negative indicators, while throughput is a positive indicator.

[0035] Based on the current load characteristics, the Analytic Hierarchy Process (AHP) is used to dynamically allocate the weights of each indicator, abandoning the fixed static weights.

[0036] The three types of indicators are normalized and mapped to the [0,1] interval. The comprehensive adaptation score Si for each candidate granularity is calculated using the adaptation scoring formula.

[0037] The granularity with the highest overall score is selected as the initial optimal granularity, and a stability check is performed: if the difference between the optimal granularity and the current RG granularity score exceeds the threshold Δ... stable If an adjustment is needed, then the original granularity configuration is maintained; otherwise, it is retained.

[0038] Furthermore, the comprehensive adaptation score calculation formula is as follows:

[0039]

[0040] In the formula, S i The comprehensive fit score representing the granularity of the i-th candidate; w j Let x be the dynamic weight of the j-th indicator at this moment; i,j norm It is the normalized value of the j-th index for the i-th candidate granularity.

[0041] Furthermore, the complete scheduling process at the modeling and execution layer includes:

[0042] Step S301: Receive the load characteristic data and load characteristic RU granularity adaptation scoring model results output by the data acquisition and extraction analysis layer as the core input for this granularity adjustment;

[0043] Step S302: Execute RU granularity change trigger rule verification and simultaneously complete constraint condition judgment. If the constraint condition is met, proceed to step S303; otherwise, proceed to step S309.

[0044] Step S303: Based on the output of the adaptive scoring model, define the RU granularity boundary and determine the minimum effective granularity lower bound R that can achieve WAF optimization. min The upper bound of the maximum effective granularity R for maximizing parallelism max and the reference granularity R basic All candidate granularities are integer multiples of the page size;

[0045] Step S304: Based on the defined granularity boundaries, construct a three-level hierarchical RU granularity candidate selection set for load adaptation;

[0046] Step S305: Activate the multi-index weighted decision algorithm to extract multi-dimensional core performance indicators;

[0047] Step S306: Based on the current load characteristics, the analytic hierarchy process is used to complete the dynamic weight allocation of load perception and output the dynamic weight values ​​of each indicator that are adapted to the current load scenario; then, all indicators are normalized to eliminate dimensional differences.

[0048] Step S307: Based on the normalized index value and dynamic weight, calculate the comprehensive suitability score of each candidate granularity, complete the candidate granularity score ranking, and select the candidate granularity with the highest comprehensive score as the preliminary optimal granularity.

[0049] Step S308: Perform stability verification, i.e., determine whether the score difference between the preliminary optimal granularity and the current RG RU granularity exceeds the preset stability threshold Δ. stable If the threshold is exceeded, proceed to step S310; otherwise, proceed to step S309.

[0050] Step S309: Keep the current RU granularity configuration of RG unchanged, end the current adjustment process, and wait for the next trigger cycle;

[0051] Step S310: Perform dynamic adjustment of RU granularity within the target RG, synchronously complete metadata and data migration, and achieve optimal granularity adaptation.

[0052] The advantages and beneficial effects of this invention are as follows:

[0053] The innovation of this invention lies in constructing a closed-loop collaborative mechanism consisting of a load-aware triggering module, a hierarchical RU granularity selection set, and a multi-index weighted decision algorithm. This solution improves system robustness by extracting I / O load characteristics in real time at the device end, eliminating the dependence on the correctness of host-side data classification strategies. By setting a load drift threshold triggering mechanism, it effectively avoids performance fluctuations caused by frequent adjustments while sensitively responding to load changes. Through the multi-index weighted decision algorithm, it incorporates mutually constraining performance indicators such as WAF, GC migration volume, channel parallel utilization, and I / O tail latency into a unified quantitative model, achieving a Pareto optimal balance of multi-dimensional performance and providing technical support for the large-scale deployment of FDP SSDs in dynamic environments. Attached Figure Description

[0054] Figure 1 This invention provides a preferred embodiment of the technology roadmap for the optimization scheme of recycling unit granularity size based on FDP SSD architecture;

[0055] Figure 2 A schematic diagram of the modeling and execution layer based on the FDP SSD architecture;

[0056] Figure 3 Modeling and execution layer flowchart based on FDP SSD architecture. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0058] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0059] The present invention provides a dynamic adjustment method for the granularity of FDP SSD recycling units for mixed workloads, which deeply considers the characteristics of FDP SSD architecture and adaptively adjusts the RU granularity size according to the mixed workload to achieve optimization. Specifically, a two-layer module is constructed to realize the function of dynamic adjustment of RU granularity, including a data acquisition and extraction analysis layer and a modeling and execution layer.

[0060] 1. Data Acquisition, Extraction, and Analysis Layer. This layer first extracts and analyzes the coupling relationship between load characteristics and RU granularity, clarifying the limitations of existing fixed RU granularity schemes. Subsequently, based on this analysis, a load-aware dynamic adjustment method for RU granularity is designed for the modeling and execution layers.

[0061] Specifically, this layer comprises four main steps. First, based on the open-source FEMU simulation platform and existing FDP simulation simulators such as FDPVirt and WARP, an FDP SSD simulation experimental environment is built. Then, through detailed simulation experiments and quantitative analysis, the relationship between load access characteristics and RU granularity under the FDP architecture is clarified. The impact of different RU granularities on FDP SSD performance under different load scenarios is quantified. Finally, a load characteristic and RU granularity adaptation scoring model is constructed, laying the foundation for the design of a load-aware dynamic adjustment method for RU granularity. In this step, only RU granularity is treated as the sole controllable variable, eliminating the interference of other hardware factors on performance results. The core mechanisms of RG / RUH data isolation, independent GC, and RU-level space management under the FDP architecture are reproduced, providing a reproducible and verifiable experimental basis for subsequent load characteristic extraction and performance impact quantification.

[0062] Secondly, load access features are extracted. At this level, four core load features are extracted, including data popularity, concurrent flow, access skew, and invalidation intensity, and standardized load classification is completed.

[0063] Furthermore, the impact of different RU granularities on the performance of FDP SSDs under different load scenarios is quantified, including various scenarios such as hot data with strong skew load, mixed hot and cold data with weak skew load, and cold data with high invalidation data stream load.

[0064] Finally, an RU granularity adaptation scoring model is generated. This model is the final output of the entire load characteristic and RU granularity relationship analysis method module. The output includes the optimal RU granularity, adaptation granularity range and boundary value for different load scenarios, providing core quantitative basis for the subsequent load-aware RU granularity dynamic adjustment method module.

[0065] 2. Modeling and Execution Layer. This layer designs a load-aware dynamic adjustment method for RU granularity. Specifically, it includes three main sub-modules: trigger rules for load-aware dynamic adjustment of RU granularity, construction of a hierarchical RU granularity selection set, and design of a multi-index weighted decision algorithm. This achieves adaptive optimization of RU granularity as load characteristics change, aiming to solve the problems of adaptability failure and WAF spikes caused by fixed RU granularity in scenarios with changing load.

[0066] 1) Load-Aware RU Granularity Dynamic Adjustment Trigger Submodule Design. This submodule aims to design a trigger that initiates the RU granularity adjustment process only when specific conditions are met. When load characteristics change, the original RU configuration may lose its FDP optimization advantages due to incompatibility with the current load, and may even cause device performance to degrade to the level of traditional SSDs. Small load fluctuations do not require changes to the RU configuration, avoiding performance degradation caused by frequent adjustments. Therefore, a load change rate threshold is set to filter out small load fluctuations, and only load changes above the threshold will consider adjusting the RU granularity. At the same time, RU granularity adjustments must be strictly performed according to RG group isolation. Adjusting RU granularity across RGs will break RG-level GC isolation, leading to the mixing of data from different lifecycles, a surge in the amount of effective GC data migration, and ultimately causing WAF spikes. Considering that during GC operations on an FDP SSD, the RU is the smallest indivisible unit, adjusting the RU granularity at this time will cause GC execution failure and data consistency risks. Therefore, the RG RU granularity adjustment behavior should be delayed until GC is completed. In addition, when the device is under high load, the data migration and metadata reconfiguration operations involved in RU granularity adjustment will compete with business I / O and GC operations for hardware resources, exacerbating performance interference and latency impact. Therefore, the RU granularity adjustment behavior of the RG should also be delayed at this time.

[0067] 2) Design of the Hierarchical RU Granularity Selection Set Construction Submodule. This submodule aims to provide a candidate range of RU granularities with clear load adaptation and performance boundaries for subsequent multi-index decision-making algorithms. By extracting and analyzing load characteristics from the data acquisition and analysis layer and the output of the RU granularity adaptation scoring model, the RU granularity boundaries are further defined, and then a three-level hierarchical RU granularity selection set for load adaptation is constructed. Specifically, the effective upper and lower limits of RU granularity are defined based on measured and simulated parameters to avoid performance and WAF degradation caused by granularity that is too large or too small. Within the effective boundaries, a three-level load adaptation RU granularity selection set of fine, medium, and coarse is constructed, corresponding to three core requirements: WAF optimization, WAF and parallelism balancing, and throughput maximization, respectively, to achieve differentiated adaptation under different load scenarios.

[0068] 3) Multi-index weighted decision-making algorithm design submodule. This module is based on a three-level RU granularity candidate set, using WAF, latency, and throughput as core evaluation indicators. It achieves dynamic weight allocation based on load awareness through the analytic hierarchy process (AHP), and calculates the comprehensive suitability score for each candidate granularity after index normalization. A stability verification mechanism is also added, triggering adjustment only when the score difference between the optimal candidate granularity and the currently used granularity exceeds a threshold, ultimately achieving accurate and stable decision-making for the optimal RU granularity under the current load.

[0069] Figure 1 This invention presents a technical roadmap for a recycling unit granularity balancing optimization scheme based on the FDP SSD architecture. Based on statistical analysis of the multi-dimensional impact of load access characteristics and FDP SSD device-level RU granularity, this invention designs a recycling unit granularity balancing optimization scheme, aiming to achieve a dynamic balance between the advantages of fine-grained space management and coarse-grained parallel efficiency, simultaneously improving the performance and reliability of FDP SSDs. The specific technical route mainly consists of two steps. First, the coupling relationship between load characteristics and RU granularity is extracted and analyzed to clarify the limitations of existing fixed RU granularity schemes. Second, based on the above analysis, a load-aware dynamic adjustment method for RU granularity is designed to achieve adaptive execution and optimization of RU granularity.

[0070] Figure 2 This diagram illustrates the modeling and execution layer based on the FDP SSD architecture. This layer comprises three main sub-modules: trigger rules for dynamic adjustment of RU granularity around load awareness, construction of a hierarchical RU granularity selection set, and design of a multi-metric weighted decision algorithm.

[0071] First, design the trigger rules for load-aware RU-level dynamic adjustment. To ensure that the dynamic adjustment of the RU granularity of FDP SSDs can both leverage the optimization benefits of FDP technology and avoid the performance and reliability risks associated with the adjustment, the specific rules include:

[0072] 1) Trigger threshold. Set a load change rate threshold so that RU granular adjustment is only initiated when the load characteristics change significantly beyond the threshold. This filters out small load fluctuations and avoids performance degradation caused by frequent adjustments.

[0073] 2) RG-level isolation. RU granularity adjustments must follow the RG grouping isolation rules, i.e., cross-RG operations are strictly prohibited to prevent damage to RG-level GC isolation capabilities and avoid WAF spikes caused by the mixing of data from different lifecycles and a surge in the amount of effective GC data migration.

[0074] 3) Delay RU granularity changes. First, since the RU is the smallest indivisible unit for GC operations on FDP SSDs, the RU adjustment operation corresponding to the RG needs to be delayed during GC operation until the GC process is completed to avoid GC execution failure and data consistency risks. Second, when the device is under high load, RU granularity adjustments need to be temporarily suspended to avoid data migration, metadata reconfiguration operations, business I / O, and GC processes competing for hardware resources, which would exacerbate performance interference and access latency degradation.

[0075] Secondly, we design a hierarchical RU selection set construction module. This submodule consists of two steps.

[0076] 1) Define the RU granularity boundary. All granularities within the interval are integer multiples of the page size. Through experiments, find the minimum effective granularity that enables WAF optimization as the lower bound R. min This avoids problems such as excessively small granularity leading to excessively high erase frequency and complete loss of hardware parallelism; and the maximum effective granularity for maximizing parallelism is defined by an upper bound R. max This avoids problems such as excessive forced mixing of hot and cold data, a surge in internal fragmentation, and performance degradation caused by excessively large granularity. R min and R max These are the upper and lower limits of granularity set based on experimental data and simulation platform parameters, respectively.

[0077] 2) Construct a three-level hierarchical RU granularity selection set for load adaptation. While ensuring that the granularity range is within the defined RU granularity boundaries, divide the RU intervals into fine-grained, medium-grained, and coarse-grained RU intervals.

[0078] The fine-grained RU interval range is proposed to be [R min 0.5×R basic ], R basic Based on research, the default settings adopted by most simulators are usually 256MB. Fine-grained RU can achieve more precise isolation between hot and cold data. For highly concentrated hot data and skewed loads, it can accurately isolate a very small amount of hot data, minimize the amount of effective data moved by GC, and achieve a WAF close to the ideal value. For scenarios with high invalidation of cold data, it can effectively isolate high invalidation flow, suppress noise RUH and improve the overall WAF.

[0079] The proposed range for the medium-grained RU interval is [0.5×R]. basic , 2×R basic Medium-grained RUs can achieve a balance between WAF control and hardware parallelism. For mixed loads with balanced distribution of hot and cold data and low access skew, they can avoid the parallelism loss caused by fine-grained RUs and the excessive WAF caused by the mixing of hot and cold data caused by coarse-grained RUs.

[0080] The range of the coarse-grained RU interval is tentatively defined as [2×R]. basic , R max The core advantage of coarse-grained RU lies in its ability to fully leverage parallel performance. Coarse-grained RU has two main advantages: first, a large-size RU can be mapped to more flash memory channels and chips simultaneously, enabling cross-channel parallel read and write operations and maximizing the utilization of hardware parallel resources; second, a large-size RU can accommodate more sequentially written data, reducing the number of GC triggers and the frequency of erase operations, thus lowering GC overhead. Therefore, it can significantly improve throughput.

[0081] Finally, a multi-index weighted decision-making algorithm module is designed. Based on the three-level hierarchical RU granularity candidate set constructed above, a multi-index weighted decision-making algorithm is further designed to select the optimal RU granularity under the current load scenario from the candidate set, and complete the final decision of dynamic adjustment.

[0082] 1) Multi-dimensional Core Performance Indicator Considerations. This section proposes using three main metrics—WAF, latency, and throughput—to evaluate the effectiveness of RU granularity configuration. First, WAF is the most crucial evaluation metric; the optimization effect of all RU granularity configurations is judged primarily by the reduction in WAF. Second, latency is equally important as WAF in system performance and is also a core QoS evaluation metric for enterprise-grade SSDs. Finally, throughput is considered as a relatively low-weighted evaluation metric to balance WAF and performance considerations.

[0083] 2) Load-aware dynamic weight allocation. Based on the load characteristic analysis data provided by the data acquisition and extraction analysis layer, the Analytic Hierarchy Process (AHP) will be used for analysis to complete the dynamic weight allocation for load adaptation, discarding static weights. The final output will be the adapted weight values ​​for different load scenarios.

[0084] 3) Indicator Normalization and Comprehensive Suitability Score. Since different indicators have different orders of magnitude and positive / negative weights, direct weighted calculation is not possible. Therefore, all indicators are normalized, mapping them to the [0,1] interval. WAF and latency are used as negative indicators, and throughput as a positive indicator. Subsequently, based on the normalized indicator values ​​and the load-aware dynamic weights, a comprehensive suitability score is calculated for each candidate granularity. A higher score indicates better overall suitability for that granularity under the current load. The expression is:

[0085]

[0086] In the formula, S i The comprehensive fit score representing the granularity of the i-th candidate; w j Let x be the dynamic weight of the j-th indicator at this moment; i,j norm It is the normalized value of the j-th index for the i-th candidate granularity.

[0087] Final decision and stability verification. The candidate granularity with the highest overall adaptability score is selected as the initial optimal granularity. To avoid frequent RU granularity adjustments due to small load fluctuations, a stability verification threshold Δ is set. stable Adjustments are only performed if, under the load characteristic-RU granularity adaptation scoring model, the difference between the initial optimal granularity and the RU granularity score currently used by the RG exceeds this threshold; otherwise, the current configuration remains unchanged to avoid performance fluctuations caused by frequent adjustments. Finally, the RU granularity is adjusted to achieve the optimal result.

[0088] Figure 3 This is a flowchart of the modeling and execution layer based on the FDP SSD architecture. The process covers three core sub-modules: trigger rule validation, hierarchical RU granularity selection set construction, and multi-index weighted decision algorithm, as detailed below:

[0089] In step S301, this process is started, and the load characteristic data and the load characteristic RU granularity adaptation scoring model results output by the data acquisition and extraction analysis layer are received as the core input for this granularity adjustment.

[0090] In step S302, the RU granularity change trigger rule verification is performed, and the constraint condition judgment is completed simultaneously. If the constraint condition is met, step S303 is executed; otherwise, step S309 is executed.

[0091] In step S303, based on the output of the adaptation scoring model, the RU granularity boundary is defined, and the minimum effective granularity lower bound R that can achieve WAF optimization is determined. min The upper bound of the maximum effective granularity R for maximizing parallelism max and the reference granularity R basicAll candidate granularities are integer multiples of the page size.

[0092] In step S304, a three-level hierarchical RU granularity candidate selection set for load adaptation is constructed based on the defined granularity boundaries.

[0093] In step S305, a multi-index weighted decision algorithm is started to extract multi-dimensional core performance indicators.

[0094] In step S306, based on the current load characteristics, the hierarchical analysis method is used to complete the dynamic weight allocation of load perception and output the dynamic weight values ​​of each indicator that are adapted to the current load scenario; then, all indicators are normalized to eliminate dimensional differences.

[0095] In step S307, based on the normalized index value and dynamic weight, the comprehensive suitability score of each candidate granularity is calculated, the candidate granularity scores are sorted, and the candidate granularity with the highest comprehensive score is selected as the preliminary optimal granularity.

[0096] In step S308, a stability check is performed, which determines whether the score difference between the preliminary optimal granularity and the current RG RU granularity exceeds a preset stability threshold Δ. stable If the threshold is exceeded, proceed to step S310; otherwise, proceed to step S309.

[0097] In step S309, the RU granularity configuration of the current RG remains unchanged, the current adjustment process ends, and the process waits for the next trigger cycle.

[0098] In step S310, dynamic adjustment of RU granularity is performed within the target RG, and metadata and data migration are completed synchronously to achieve optimal granularity adaptation.

[0099] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0100] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0101] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A method for dynamically adjusting the granularity of FDP SSD recycling units for mixed loads, characterized in that, The system is divided into two collaborative layers: a data acquisition and extraction analysis layer and a modeling and execution layer. The data acquisition and extraction analysis layer is used to build the FDP simulation environment, extract multi-dimensional load features, quantify the coupling relationship between RU granularity and multiple performance indicators, and output a load-RU granularity adaptation scoring model. The modeling and execution layer, based on the scoring model, achieves adaptive dynamic switching of RG-level RU granularity through load-aware granularity adjustment trigger rules, hierarchical RU granularity selection set, and multi-indicator weighted decision algorithm.

2. The method for dynamically adjusting the granularity of FDP SSD recycling units for mixed loads according to claim 1, characterized in that, The data acquisition, extraction, and analysis layer execution steps include: S1. Based on the simulation platform, build an FDP SSD simulation environment, using only the RU granularity as a single controllable variable, to reproduce RG / RUH data isolation, RU-level GC, and space management mechanisms. S2. Extract four core load characteristics: data popularity, number of concurrent streams, access skew, and invalidation intensity, and complete the standardized classification of load scenarios; S3. Quantify the variation patterns of WAF, latency, and throughput under various load conditions, including hot skew, cold and hot balanced weak skew, and high invalidation of cold data, with different RU granularities. S4. Construct a load characteristic and RU granularity adaptation scoring model, and output the optimal RU granularity, granularity adaptation range and boundary value for each load scenario as the input basis for modeling and execution layers.

3. The method for dynamically adjusting the granularity of FDP SSD recycling units for mixed loads according to claim 1, characterized in that, The modeling and execution layer comprises three main sub-modules: a load-aware granularity adjustment triggering sub-module, a hierarchical RU granularity selection set construction sub-module, and a multi-index weighted decision algorithm sub-module. The load sensing granularity adjustment triggering submodule is set with multiple constraints, and the granularity adjustment process can only be entered if all of them are met. The hierarchical RU granularity selection set construction submodule defines the effective upper and lower limits of RU granularity and divides the candidate intervals of fine, medium and coarse granularity. The multi-index weighted decision algorithm submodule allocates dynamic weights based on the analytic hierarchy process, calculates the comprehensive adaptation score of each candidate granularity and adds stability verification, and selects the optimal RU granularity to complete the configuration update within the RG.

4. The hybrid load oriented FDP SSD recycling unit granularity dynamic adjustment method according to claim 3, characterized in that, The multiple constraints of the load-aware granularity adjustment triggering submodule include: 1) Trigger threshold; Set a load change rate threshold so that RU granular adjustment is only initiated when the load characteristics change significantly beyond the threshold, thus filtering out small load fluctuations and avoiding performance degradation caused by frequent adjustments; 2) RG-level isolation; RU granularity adjustment must follow the RG grouping isolation rules, that is, cross-RG operations are strictly prohibited to prevent damage to the RG-level GC isolation capability; 3) Delay the execution of RU granularity changes; Firstly, during GC operation, the RU adjustment operation of the corresponding RG needs to be delayed until the GC process ends; Secondly, when the device is in a high-load environment, the RU granularity adjustment needs to be temporarily suspended.

5. The hybrid load oriented FDP SSD recycling unit granularity dynamic adjustment method according to claim 3, characterized in that, The execution process of the hierarchical RU granularity selection set construction submodule includes: 1) Define the RU granularity boundary; all granularities within the interval are integer multiples of the page size. Experimentally, find the minimum effective granularity that enables WAF optimization as the lower bound R. min And the maximum effective granularity for maximizing parallelism is bounded by R. max R min and R max These are the upper and lower limits of granularity set based on experimental measured data and simulation platform parameters, respectively. 2) Construct a three-level hierarchical RU granularity selection set for load adaptation; while ensuring that the granularity range is within the defined RU granularity boundaries, divide the RU interval into fine-grained, medium-grained, and coarse-grained intervals. The fine-grained RU interval range is proposed to be [R min 0.5×R basic ], R basic This refers to the default settings adopted by most emulators based on research; The proposed range for the medium-grained RU interval is [0.5×R]. basic , 2×R basic ]; The range of the coarse-grained RU interval is tentatively defined as [2×R]. basic , R max ].

6. The method for dynamically adjusting the granularity of FDP SSD recycling units for mixed loads according to claim 3, characterized in that, The execution steps of the multi-index weighted decision algorithm submodule include: WAF, I / O latency, and throughput are selected as the three core evaluation indicators. WAF and latency are negative indicators, while throughput is a positive indicator. Based on the current load characteristics, the Analytic Hierarchy Process (AHP) is used to dynamically allocate the weights of each indicator, abandoning the fixed static weights. The three types of indicators are normalized and mapped to the [0,1] interval. The comprehensive adaptation score Si of each candidate granularity is calculated by the adaptation scoring formula. The granularity with the highest overall score is selected as the initial optimal granularity, and a stability check is performed: if the difference between the optimal granularity and the score of the currently used granularity in RG exceeds the threshold Δ stable If an adjustment is needed, then the original granularity configuration is maintained; otherwise, it is retained.

7. The method for dynamically adjusting the granularity of FDP SSD recycling units for mixed loads according to claim 6, characterized in that, The formula for calculating the overall compatibility score is: ； In the formula, S i The comprehensive fit score representing the granularity of the i-th candidate; w j Let x be the dynamic weight of the j-th indicator at this moment; i,j norm It is the normalized value of the j-th index for the i-th candidate granularity.

8. The method for dynamically adjusting the granularity of FDP SSD recycling units for mixed loads according to claim 1, characterized in that, The complete scheduling process of the modeling and execution layer includes: Step S301: Receive the load characteristic data and load characteristic RU granularity adaptation scoring model results output by the data acquisition and extraction analysis layer as the core input for this granularity adjustment; Step S302: Execute RU granularity change trigger rule verification and simultaneously complete constraint condition judgment. If the constraint condition is met, proceed to step S303; otherwise, proceed to step S309. Step S303: Based on the output of the adaptive scoring model, define the RU granularity boundary and determine the minimum effective granularity lower bound R that can achieve WAF optimization. min The upper bound of the maximum effective granularity R for maximizing parallelism max and the reference granularity R basic All candidate granularities are integer multiples of the page size; Step S304: Based on the defined granularity boundaries, construct a three-level hierarchical RU granularity candidate selection set for load adaptation; Step S305: Activate the multi-index weighted decision algorithm to extract multi-dimensional core performance indicators; Step S306: Based on the current load characteristics, the analytic hierarchy process is used to complete the dynamic weight allocation of load perception and output the dynamic weight values ​​of each indicator that are adapted to the current load scenario; then, all indicators are normalized to eliminate dimensional differences. Step S307: Based on the normalized index value and dynamic weight, calculate the comprehensive suitability score of each candidate granularity, complete the candidate granularity score ranking, and select the candidate granularity with the highest comprehensive score as the preliminary optimal granularity. Step S308: Perform stability verification, i.e., determine whether the score difference between the preliminary optimal granularity and the current RG RU granularity exceeds the preset stability threshold Δ. stable If the threshold is exceeded, proceed to step S310; otherwise, proceed to step S309. Step S309: Keep the current RU granularity configuration of RG unchanged, end the current adjustment process, and wait for the next trigger cycle; Step S310: Perform dynamic adjustment of RU granularity within the target RG, synchronously complete metadata and data migration, and achieve optimal granularity adaptation.

Images