NVMe storage array virtualization device based on FPGA

By implementing functional modules such as the SR-IOV virtualization module in the FPGA, the performance bottleneck of traditional software in solving high-concurrency, low-latency NVMe storage access is solved. This enables low-latency, high-throughput multi-tenant storage services and service quality isolation, improves system resource isolation, enhances the practical application of the system, and achieves service quality isolation in a low-latency multi-tenant environment, thereby improving system resource utilization and resource isolation capabilities.

CN121858027APending Publication Date: 2026-04-14CHINA SHIPBUILDING IND CORP NO 723 RESEARCH INSTITUTE
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202511869670.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-12-11
Publication Date
2026-04-14

AI Technical Summary

Technical Problem

Traditional software-layer storage virtualization solutions suffer from performance bottlenecks when handling high-concurrency, low-latency NVMe storage access. They exhibit high I/O path latency and high CPU resource consumption, making it difficult to meet the efficiency and scalability requirements of cloud computing and high-performance computing scenarios for storage systems. In particular, service quality is difficult to guarantee in multi-tenant environments.

Method used

An FPGA-based NVMe storage array virtualization device is adopted. By implementing the SR-IOV virtualization module, virtual address mapping unit, resource pooling processing module and dynamically reconfigurable functional module in the FPGA hardware, the entire process of storage virtualization is completed, realizing low-latency, high-throughput multi-tenant storage service and service quality isolation.

Benefits of technology

Significantly reduces I/O path latency, approaching the native performance level of physical SSDs, achieves hardware-level resource isolation, improves system throughput and fault tolerance, saves host computing resources, and increases resource utilization and system energy efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121858027A_ABST
    Figure CN121858027A_ABST
Patent Text Reader

Abstract

The invention discloses an NVMe storage array virtualization device based on an FPGA (Field Programmable Gate Array), which is characterized in that single I / O (Input / Output) virtualization (SR-IOV) is realized in the FPGA, and a plurality of NVMe virtual functions which can be directly accessed by a virtual machine are externally provided; mapping from a virtual volume logic address to a physical storage address is completed in hardware, striping management and RAID verification calculation are carried out on a plurality of NVMe solid state disks at the rear end, and storage resource pooling is achieved; the partial reconfigurable characteristic of the FPGA is utilized, I / O scheduling, data compression and prefetching strategies are dynamically adjusted, and the adaptability and performance efficiency of the system are improved. Hardware unloading of the whole I / O path is completed in the FPGA, intervention of a host CPU is not needed, delay is remarkably reduced, the throughput capacity is improved, service quality isolation and elastic expansion in a multi-tenant environment are supported, and an effective technical path is provided for constructing a next-generation high-performance and low-delay storage infrastructure.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to high-performance storage system technology, and more specifically to an FPGA-based NVMe storage array virtualization device. Background Technology

[0002] With the rapid development of big data and cloud computing technologies, data volume is exploding, placing unprecedented demands on the performance, scale, and efficiency of storage systems. NVMe (Non-Volatile Memory Express) storage arrays, with their low latency and high throughput, are rapidly becoming a core storage solution in data centers and high-performance computing. Compared to traditional SATA and SAS interfaces, the NVMe protocol improves the access efficiency of solid-state drives (SSDs) by supporting deep queues up to 64K, multi-core parallel processing capabilities, and a streamlined command set. Test data shows that NVMe SSDs can achieve 4K random read / write latency as low as tens of microseconds, outperforming traditional interface devices. This performance advantage is crucial for handling I / O-intensive tasks such as real-time analytics, high-frequency trading, and artificial intelligence training. Meanwhile, the dynamic, elastic scaling, and multi-tenant sharing characteristics of cloud computing environments urgently require that underlying storage resources be efficiently pooled, flexibly scheduled, and allocated on demand.

[0003] Storage virtualization technology is a key path to meeting this need. Its core value lies in abstracting physical storage devices into a unified logical resource pool, shielding users from the heterogeneity and complexity of the underlying hardware, and providing transparent and flexible storage services. Through the virtualization layer, administrators can dynamically create, delete, expand, or migrate virtual storage volumes, achieving on-demand allocation and efficient utilization of resources. This capability is particularly important in large-scale cloud data centers, supporting the rapid deployment and migration of virtual machines (VMs) or containers, ensuring strong isolation of storage resources between different tenants or applications, and ultimately improving the overall infrastructure's resource utilization and operational efficiency. Traditional software-layer storage virtualization solutions (such as Logical Volume Manager (LVM) based on the host operating system or Hypervisor-based virtual storage layers) often face performance bottlenecks when handling high-concurrency, low-latency NVMe storage access. The complexity of the software stack, frequent system calls, and the overhead of CPU interrupt handling introduce additional latency and consume valuable host computing resources. Research indicates that under peak loads, the CPU resources consumed by the software virtualization layer itself can sometimes reach as high as 30%, which not only limits the full utilization of storage performance but also squeezes out computing power intended for business applications.

[0004] Faced with the performance constraints of software solutions, seeking hardware-level acceleration has become an inevitable trend for optimizing storage virtualization performance. Field-Programmable Gate Arrays (FPGAs) demonstrate enormous potential in this field due to their unique hardware architecture. FPGAs possess extremely high parallel processing capabilities. Internally, they consist of numerous programmable logic units (such as LUTs and flip-flops), dedicated hardware modules (such as Block RAM and DSPs), and high-speed interconnect interfaces (such as PCIe), supporting the embedding of complex processing logic (such as NVMe command parsing, address mapping, and queue scheduling) into the hardware pipeline. This parallel execution mechanism can handle multiple I / O requests simultaneously, improving throughput. FPGAs can achieve extremely low processing latency. By directly processing critical operations on the I / O path through hardware logic, FPGAs can bypass the overhead of context switching and memory copying caused by the operating system kernel and software driver stack, significantly shortening the data transmission path and achieving near-hardware nanosecond (ns) or microsecond (µs) response times. FPGAs offer high flexibility and reconfigurability. Its hardware logic can be updated or optimized after deployment through programming (such as using VHDL / Verilog or HLS), allowing it to adapt to protocol evolution (such as updates to the NVMe standard) or meet the customized needs of specific application scenarios (such as scheduling algorithms for specific workloads, data compression / encryption) more quickly compared to application-specific integrated circuits (ASICs). FPGAs typically have a higher performance per watt than CPUs, with power consumption mainly related to the active logic resources, providing superior energy efficiency when efficiently processing specific tasks.

[0005] Virtualizing NVMe storage arrays based on FPGAs has profound theoretical value and broad application prospects. From a theoretical perspective, this research represents a significant exploration of the intersection of computer architecture and storage systems. It aims to construct a high-performance, low-overhead, and scalable hardware virtualization abstraction layer. The core challenge lies in efficiently and transparently mapping and presenting the multi-queue parallelism and namespace resources of physical NVMe devices to upper-layer systems or tenants, while ensuring resource isolation, QoS (Quality of Service) guarantees, and data consistency through hardware mechanisms. This involves a deep understanding of the NVMe protocol stack, high-speed hardware pipeline design, hardware implementation of resource scheduling algorithms, and innovation in efficient collaboration mechanisms between FPGAs, host systems, and back-end storage arrays. Overcoming key technical difficulties, such as implementing complex virtualization management logic within the limited on-chip resources of an FPGA, designing a low-latency, high-parallel adaptive scheduler, and constructing efficient address translation and caching mechanisms, will provide important theoretical support and paradigm references for the architectural design of next-generation intelligent storage systems.

[0006] From an application perspective, FPGA-based NVMe storage virtualization solutions are expected to play a revolutionary role in several key areas. In cloud data centers, it can provide virtual block storage services with near-physical device performance for massive numbers of virtual machines or containers, significantly reducing the performance loss and CPU overhead caused by storage virtualization, and improving the overall resource utilization efficiency and economy of the data center. Alibaba Cloud's research shows that optimizing NVMe storage infrastructure is one of its key technological directions supporting ultra-large-scale deployments. In the field of high-performance computing (HPC), this solution can meet the demanding requirements of scientific computing, simulation, and other applications for high-speed reading and writing of massive amounts of data, accelerating data processing processes and shortening task execution time. In artificial intelligence (AI) and machine learning (ML) scenarios, it can reduce the waiting time for loading large-scale training datasets and improve the utilization of computing resources such as GPUs, thereby accelerating model training iterations. At edge computing nodes, the combination of the low power consumption and customizability of FPGAs with the high performance of NVMe can provide powerful local storage acceleration capabilities for real-time data processing and analysis at the edge. This technology also naturally aligns with the development trend of "computational storage". FPGAs can perform near-data processing tasks (such as filtering, transformation, and aggregation) at the data storage location, further reducing data movement, releasing host load, and bringing a performance leap to applications such as database analysis and stream processing. Summary of the Invention

[0007] The purpose of this invention is to provide an FPGA-based NVMe storage array virtualization device to overcome the performance bottlenecks of traditional software-layer storage virtualization schemes when handling high-concurrency, low-latency NVMe storage access. Specifically, these bottlenecks manifest as high I / O path latency, high CPU resource consumption, and difficulty in guaranteeing service quality in multi-tenant environments, failing to meet the stringent requirements of cloud computing and high-performance computing scenarios for the efficiency and scalability of storage systems.

[0008] The technical solution for achieving the objective of this invention is as follows: an FPGA-based NVMe storage array virtualization device, comprising a field-programmable gate array (FPGA) chip, multiple physical NVMe storage devices connected to the FPGA, and an interface module for implementing PCIe communication, wherein the FPGA is configured to execute control logic for storage virtualization, characterized in that:

[0009] The FPGA integrates an SR-IOV virtualization module, which presents the device to the outside world as a standard NVMe endpoint that supports multiple virtual functions (VFs). Each virtual function corresponds to an independent virtual storage volume and can be directly accessed by different computing instances on the host side.

[0010] The FPGA is also equipped with a virtual address mapping unit, which is used to realize the spatial translation of the logical block address (LBA) of the virtual storage volume to the address of the back-end physical NVMe storage device;

[0011] The FPGA further includes a resource pooling processing module, which is used to perform striped distribution management of multiple physical NVMe storage devices and synchronously perform RAID verification calculations during data writing to provide redundancy protection; the resource pooling processing module establishes parallel data paths with each physical NVMe storage device to realize the splitting and aggregation of I / O requests;

[0012] The FPGA also includes a dynamically reconfigurable functional module area, which is used to switch I / O scheduling strategies, data compression modes or prefetch algorithms online according to real-time workload changes.

[0013] Among them, the SR-IOV virtualization module, virtual address mapping unit, resource pooling processing module and dynamically reconfigurable functional module area work together to complete the entire process of storage virtualization at the hardware level, thereby achieving low latency, high throughput multi-tenant storage service and service quality isolation.

[0014] Furthermore, the dynamically reconfigurable functional module area includes multiple pre-configured logical functional sub-modules, each of which includes at least an adaptive I / O scheduler, a lightweight data compression engine, and an intelligent prefetch controller based on access pattern recognition. The FPGA dynamically loads the corresponding logical functional sub-modules according to the received performance monitoring feedback information or policy instructions issued by the host to optimize the I / O processing efficiency under the current workload.

[0015] Furthermore, the intelligent prefetch controller is configured to periodically collect the I / O access trajectory of each virtual storage volume, and use a finite state machine model to classify and identify the access patterns. When a sequential or step-size regular access is detected, the prefetch process is automatically started, and the predicted subsequent data blocks are loaded from the back-end physical NVMe storage device to the FPGA on-chip cache or external high-speed DDR memory.

[0016] Furthermore, the SR-IOV virtualization module also integrates a QoS control unit, which is used to independently configure the bandwidth limit, IOPS quota and burst traffic window for each virtual function (VF), and to measure and regulate the I / O requests of each virtual function in real time through the token bucket algorithm, so as to prevent resource preemption and realize service quality isolation between multi-tenants.

[0017] Furthermore, the resource pooling processing module also includes bad block management and wear leveling logic, which is used to monitor the health status of each physical NVMe storage device, avoid marked failed storage areas during data writing, and dynamically adjust the data mapping position according to the erase and write count distribution of each storage area to extend the overall lifespan of the storage array.

[0018] Furthermore, the virtual address mapping unit adopts a hierarchical mapping structure, including a high-frequency access page table cache located in the FPGA on-chip BRAM and a complete mapping table located in external DDR memory. When a cache miss occurs, adjacent mapping entries are asynchronously prefetched through the background DMA engine to improve the hit rate of subsequent queries.

[0019] Furthermore, the FPGA connects to multiple physical NVMe storage devices via PCIe switching and connects to the host system as an NVMe device endpoint, forming a "one virtual machine with multiple pipes and bidirectional direct connection" topology. In this topology, all I / O data flows are directly transmitted between the host system and the physical NVMe storage devices under the control of the FPGA chip, without passing through the host CPU or memory, thus achieving a completely offloaded end-to-end hardware virtualization path.

[0020] Furthermore, the virtual address mapping unit adopts a page-based mapping structure, with each page being 4KB in size. The mapping table entries are stored in the on-chip BRAM of the FPGA chip, and the mapping table supports the simultaneous operation of no less than 64 virtual memory volumes.

[0021] Furthermore, the resource pooling processing module divides all physical NVMe storage devices into fixed-size data blocks and distributes them in a striped manner according to RAID 5; when writing data, the processing module automatically splits the data and sends it in parallel to multiple physical NVMe storage devices.

[0022] Furthermore, the SR-IOV virtualization module is instantiated in the FPGA chip as 4 to 32 independent virtual functions. Each virtual function has an independent submission queue and completion queue, and communicates with the host system through the MSI-X interrupt mechanism.

[0023] Compared with existing technologies, the significant advantages of this invention are: 1) By migrating all the core virtualization functions of the NVMe storage array to the FPGA hardware, the performance overhead of context switching, memory copying, and interrupt handling brought by the traditional software stack is fundamentally bypassed, significantly reducing I / O path latency and achieving microsecond-level response time, approaching the native performance level of physical SSDs. 2) The SR-IOV mechanism enables multiple virtual machines or containers to directly and exclusively access their respective virtual storage volumes, achieving hardware-level resource isolation and pass-through access, effectively avoiding the "neighbor noise" problem and ensuring the quality of service (QoS) in a multi-tenant environment. 3) By integrating virtual address mapping, RAID checksum calculation, striping management, and other functions into the FPGA, unified pooling and protocol offloading of backend storage resources are achieved, improving system throughput and fault tolerance. 4) The introduction of partially reconfigurable technology allows I / O scheduling, data compression, and prefetching strategies to be dynamically adjusted according to the load, enhancing the system's adaptability and flexibility. 5) In addition, the entire data path is decoupled from the host CPU and memory, significantly saving host computing resources and freeing up more computing power for business processing. This solution combines high performance, low latency, high energy efficiency, and strong scalability, providing a practical technical path for building next-generation cloud data centers, high-performance computing, and edge intelligent storage infrastructure. Attached Figure Description

[0024] Figure 1 This is a diagram illustrating the overall architecture of an FPGA-based NVMe storage array virtualization system.

[0025] Figure 2 Block diagram of FPGA internal functional module integration and interaction relationship

[0026] Figure 3 I / O data path and timing flowchart

[0027] Figure 4 This is a state machine diagram for dynamic reconfiguration and intelligent prefetch control logic. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0029] An FPGA-based NVMe storage array virtualization device includes a field-programmable gate array (FPGA) chip, multiple physical NVMe storage devices connected to the FPGA, and an interface module for implementing PCIe communication. The FPGA is configured to execute control logic for storage virtualization. The FPGA integrates an SR-IOV virtualization module to present the device externally as a standard NVMe endpoint supporting multiple virtual functions (VFs), each VF corresponding to an independent virtual storage volume and directly accessible by different computing instances on the host side. The FPGA also includes a virtual address mapping unit for spatial translation from virtual volume logical block addresses (LBAs) to addresses of the backend physical NVMe storage devices. The FPGA further... The system includes a resource pooling processing module for striped distribution management of multiple physical NVMe storage devices and synchronously performs RAID verification calculations during data writing to provide redundancy protection. The resource pooling processing module establishes parallel data paths with each physical NVMe storage device to achieve I / O request distribution and aggregation. The FPGA also includes a dynamically reconfigurable functional module area for online switching of I / O scheduling strategies, data compression modes, or prefetch algorithms based on real-time workload changes. The SR-IOV virtualization module, virtual address mapping unit, resource pooling processing module, and dynamically reconfigurable functional module area work together to complete the entire storage virtualization process at the hardware level, thereby achieving low-latency, high-throughput multi-tenant storage services and service quality isolation.

[0030] The technical challenge of this solution lies in the fact that traditional software-layer virtualization architectures rely on the host CPU to handle command parsing, address mapping, and queue scheduling in the I / O path, resulting in high I / O latency and high CPU resource consumption. This makes it difficult to meet the demands of high-performance cloud environments for low-latency, high-concurrency storage access. Especially in multi-tenant shared scenarios, resource contention among virtual machines intensifies, making it difficult to guarantee Quality of Service (QoS).

[0031] The specific implementation is as follows: A Xilinx Virtex UltraScale+ series FPGA chip is used as the core processing platform. This chip possesses abundant programmable logic resources (approximately 9 million LUTs), high-speed Block RAM (BRAM), DSP, and high-speed serial (32.75 Gb / s Serdes) channels, supporting multiple PCIe Gen3 x16 interfaces to meet the requirements of high-bandwidth, low-latency data interaction. The FPGA connects 4 to 8 commercial NVMe solid-state drives via an onboard PCIe switching chip, using M.2 or U.2 interfaces to form a back-end storage array. Internally, the FPGA uses HDL (Verilog / VHDL) to construct a complete NVMe HostController logic, supporting the standard NVMe 1.4 protocol stack. The SR-IOV virtualization module is instantiated in the FPGA as 4-32 independent virtual functions (VFs). Each VF has its own Submission Queue (SQ) and Completion Queue (CQ) and communicates with the host system through the MSI-X interrupt mechanism. Each Virtual Container (VF) is bound to a virtual storage volume with a configurable capacity of 100GB to 128TB, which is allocated by the administrator to different virtual machines or containers through a management interface. The virtual address mapping unit adopts a page-based mapping structure, with each page being 4KB in size. The mapping table entries are stored in the on-chip BRAM of the FPGA, supporting up to 64 virtual volumes running simultaneously. The resource pooling processing module divides all physical NVMe devices into fixed-size data blocks (chunksize of 512kB), which are striped and distributed according to RAID5. When writing data, it is automatically split and sent in parallel to multiple SSDs, and a dedicated XOR engine is used to generate checksum information in real time. All operations are completed in the FPGA pipeline without host intervention. The PCIe switching chip (such as Broadcom PLX PEX8748) is used to expand the connection of multiple NVMe devices to form a star topology. The dynamically reconfigurable functional module area is located in the Partial Reconfiguration Region of the FPGA. It initially loads the basic I / O scheduler (round-robin mechanism) and can subsequently switch to a priority queue or shortest seek first strategy according to the load type. The entire system is deployed within a standard PCIe card form factor, powered by a dual 12V / 3.3V power input, with power consumption controlled below 50W. In practical applications, the FPGA model can also be an Intel or Lattice model with equivalent performance resources, and the NVMe SSD can be replaced with other compatible devices; this application does not limit this.

[0032] Working principle: When the host initiates an I / O request, the PCIe interface receives the NVMe command. The SR-IOV module identifies the associated Virtual Function (VF) and performs LBA translation on the corresponding Virtual Address Mapping Unit to obtain the target physical SSD number and offset address. The resource pooling module writes data in parallel to multiple SSDs according to striping rules, while simultaneously initiating RAID checksum calculations. Read operations aggregate the data blocks returned by each disk. The dynamic reconfiguration module continuously monitors the queue depth, response time, and access mode of each VF. When a load change is detected, it triggers a partial bitstream update and loads new scheduling or compression logic. All processes are completed in the FPGA's internal hardware pipeline, without passing through host memory or the operating system kernel.

[0033] Technical benefits: It achieves full hardware offloading of the I / O path, reduces end-to-end latency, and improves upon traditional software virtualization solutions; it supports direct access from up to 32 virtual machines, and achieves hardware-level isolation between virtual volumes to effectively prevent resource contention; the system throughput can reach over 8GB / s, while significantly saving host CPU resources and significantly improving the overall energy efficiency and resource utilization of the data center.

[0034] An FPGA-based NVMe storage array virtualization device includes a field-programmable gate array (FPGA) chip, multiple physical NVMe storage devices connected to the FPGA, and an interface module for implementing PCIe communication. The FPGA is configured to execute control logic for storage virtualization. The FPGA integrates an SR-IOV virtualization module to present the device externally as a standard NVMe endpoint supporting multiple virtual functions (VFs), each VF corresponding to an independent virtual storage volume and directly accessible by different computing instances on the host side. The FPGA also includes a virtual address mapping unit for spatial translation from virtual volume logical block addresses (LBAs) to addresses of the backend physical NVMe storage devices. The FPGA further... The system includes a resource pooling module for striped distribution management of multiple physical NVMe storage devices and synchronously performs RAID verification calculations during data writing to provide redundancy protection. The resource pooling module establishes parallel data paths with each physical NVMe storage device to achieve I / O request distribution and aggregation. The FPGA also includes a dynamically reconfigurable functional module area for online switching of I / O scheduling strategies, data compression modes, or prefetch algorithms based on real-time workload changes. The SR-IOV virtualization module, virtual address mapping unit, resource pooling module, and dynamically reconfigurable functional module area work together to complete the entire storage virtualization process at the hardware level, thereby achieving low-latency, high-throughput multi-tenant storage services and service quality isolation. Preferably, the dynamically reconfigurable functional module area includes multiple pre-configured logical functional sub-modules, each including at least an adaptive I / O scheduler, a lightweight data compression engine, and an intelligent prefetch controller based on access pattern recognition. The FPGA dynamically loads the corresponding sub-modules based on received performance monitoring feedback or policy instructions issued by the host to optimize I / O processing efficiency under the current workload.

[0035] The solution further addresses the problem that static, fixed I / O processing strategies cannot adapt to the dynamically changing workload characteristics in a cloud computing environment (such as bursty read-intensive operations and write amplification), leading to uneven resource utilization and affecting overall performance.

[0036] The specific implementation involves dividing the FPGA chip into a dynamically reconfigurable region occupying approximately 20% of the total area. This region is pre-programmed with bitstream images of three functional sub-modules: an adaptive I / O scheduler employing a weighted fair queue (WFQ) algorithm, with weights dynamically adjusted based on historical response times; a lightweight data compression engine based on the LZ4 compression protocol, with compression units controlled by a state machine, supporting real-time compression of data blocks up to 64KB, achieving an average compression ratio of 2:1; and an intelligent prefetch controller with a built-in finite state machine model for analyzing the locality and regularity of I / O sequences. These three sub-modules exist as independent IP cores and can be loaded on demand via the ICAP (Internal Configuration Access Port) interface. During system operation, the FPGA's internal performance monitoring module collects the IOPS, latency, queue depth, and read / write ratio of each VF every 10ms and uploads this data to the management agent via the AXI bus. When three consecutive samples show that the read operation ratio exceeds 80% and the sequentiality is higher than 70%, the prefetch controller is triggered to load the data. If frequent small-block random writes are detected, the compression engine is activated to reduce the actual write volume. All reconstruction processes are completed within 200μs without affecting ongoing I / O services. The compression engine operates at a frequency of 300MHz, the prefetch controller cache depth is 128 entries, and the scheduler supports up to 8 priority queues. In practical applications, the compression algorithm can also be replaced with Zstandard or Snappy hardware implementations; this embodiment does not limit this.

[0037] Working principle: The dynamic reconfiguration mechanism drives the switching of functional modules through runtime performance feedback, enabling the FPGA to automatically match the optimal processing strategy according to different application scenarios. For example, in database query scenarios, intelligent prefetching is enabled to preload index pages; in log writing scenarios, compression is enabled to reduce write amplification; and under mixed loads, an adaptive scheduler is enabled to balance the service time of various requests.

[0038] Technical effects: It improves the system's intelligence level and resource adaptability, reduces the average I / O response time, and improves the effective bandwidth utilization. Especially in typical AI training dataset reading scenarios, the high prefetch hit rate significantly reduces the number of repeated accesses to the backend SSD.

[0039] An FPGA-based NVMe storage array virtualization device includes a field-programmable gate array (FPGA) chip, multiple physical NVMe storage devices connected to the FPGA, and an interface module for implementing PCIe communication. The FPGA is configured to execute control logic for storage virtualization. The FPGA integrates an SR-IOV virtualization module to present the device externally as a standard NVMe endpoint supporting multiple virtual functions (VFs), each VF corresponding to an independent virtual storage volume and directly accessible by different computing instances on the host side. The FPGA also includes a virtual address mapping unit for spatial translation from virtual volume logical block addresses (LBAs) to addresses of the backend physical NVMe storage devices. The FPGA further... The system includes a resource pooling processing module for striped distribution management of multiple physical NVMe storage devices and synchronously performs RAID checksum calculations during data writing to provide redundancy protection. The resource pooling processing module establishes parallel data paths with each physical NVMe storage device to achieve I / O request distribution and aggregation. The FPGA also includes a dynamically reconfigurable functional module area for online switching of I / O scheduling strategies, data compression modes, or prefetch algorithms based on real-time workload changes. The SR-IOV virtualization module, virtual address mapping unit, resource pooling processing module, and dynamically reconfigurable functional module area work together to complete the entire storage virtualization process at the hardware level, thereby achieving low-latency, high-throughput multi-tenant storage services and service quality isolation. Preferably, the intelligent prefetch controller is configured to periodically collect the I / O access trajectory of each virtual volume and classify and identify access patterns using a finite state machine model. When sequential or step-size regular access is detected, the prefetch process is automatically initiated, and the predicted data blocks are loaded from the backend physical NVMe storage devices into the FPGA on-chip cache or external high-speed DDR memory.

[0040] The problem addressed by this solution is that existing prefetching mechanisms rely heavily on software statistical analysis, resulting in delayed responses and a lack of fine-grained identification of access patterns, which can easily lead to invalid prefetching, wasted bandwidth, and increased latency.

[0041] Implementation details: The intelligent prefetch controller is integrated into the FPGA logic and comprises two core components: an access trajectory collector and a pattern recognition state machine. The trajectory collector records the starting LBA and length of the last 10 read requests for each virtual volume every 1ms, forming an access sequence. The pattern recognition state machine includes four states: "Initial," "Sequential," "Jump," and "Random." It determines the access pattern by comparing whether the difference between adjacent LBAs is constant or linear. If three consecutive LBA increments are the same, it is determined to be a sequential access, entering the "Sequential" state and initiating the prefetch process; if the difference changes in a fixed step size, it is determined to be a step-by-step access, entering the "Jump" state. The prefetch controller generates prefetch commands for several subsequent 4KB data blocks based on the recognition results and submits them to the backend NVMe queue via the DMA engine. Prefetched data is temporarily stored in the FPGA's onboard DDR4-3200 memory (16GB capacity) or preferentially written to the on-chip BRAM (4MB capacity) for extremely high-frequency data access. The prefetch window size is dynamically adjusted, ranging from 1 to 16 pages, depending on the current link bandwidth and cache utilization. The controller is equipped with a fallback mechanism; if two consecutive prefetches fail, prefetching is paused and the access mode is re-evaluated. In practical applications, DDR memory can also be replaced with DDR5 or HBM2e depending on the FPGA selection; this application embodiment does not limit this.

[0042] Working principle: The prefetch controller continuously monitors the behavior characteristics of I / O streams and uses a finite state machine to achieve fast mode switching. After confirming a regular access trend, it actively pulls data and keeps it in the near-computation cache before it is actually requested, thereby greatly shortening the response time of subsequent read operations.

[0043] Technical effects: It enables hardware-level real-time prefetching decision-making, avoids the latency caused by software polling, and improves prefetching accuracy and overall system read throughput.

[0044] An FPGA-based NVMe storage array virtualization device includes a field-programmable gate array (FPGA) chip, multiple physical NVMe storage devices connected to the FPGA, and an interface module for implementing PCIe communication. The FPGA is configured to execute control logic for storage virtualization. The FPGA integrates an SR-IOV virtualization module to present the device externally as a standard NVMe endpoint supporting multiple virtual functions (VFs), each VF corresponding to an independent virtual storage volume and directly accessible by different computing instances on the host side. The FPGA also includes a virtual address mapping unit for spatial translation from virtual volume logical block addresses (LBAs) to addresses of the backend physical NVMe storage devices. The FPGA further... The system includes a resource pooling module for striped distribution management of multiple physical NVMe storage devices and synchronously performs RAID checksum calculations during data writing to provide redundancy protection. The resource pooling module establishes parallel data paths with each physical NVMe storage device to achieve I / O request distribution and aggregation. The FPGA also includes a dynamically reconfigurable functional module area for online switching of I / O scheduling strategies, data compression modes, or prefetch algorithms based on real-time workload changes. The SR-IOV virtualization module, virtual address mapping unit, resource pooling module, and dynamically reconfigurable functional module area work collaboratively to complete the entire storage virtualization process at the hardware level, thereby achieving low-latency, high-throughput multi-tenant storage services and service quality isolation. Preferably, the SR-IOV virtualization module also integrates a QoS control unit for independently configuring bandwidth limits, IOPS quotas, and burst traffic windows for each virtual function (VF), and using a token bucket algorithm to measure and regulate I / O requests for each VF in real time to prevent resource preemption and achieve service quality isolation between multi-tenants.

[0045] The technical problem this solution addresses is that in a multi-tenant environment, the high I / O load of a virtual machine may crowd out the storage resources of other tenants, leading to a decline in service quality, i.e., the "neighbor noise" problem.

[0046] Implementation details: The QoS control unit is integrated within the SR-IOV virtualization module. Each VF is configured with an independent token bucket counter. The bucket capacity represents the maximum allowed burst request count (default 1024 commands), and the filling rate is a preset IOPS limit (configurable within the range of 1k~100k). Whenever an I / O request is received, one token is deducted from the corresponding VF's token bucket. If there are insufficient tokens, the request is temporarily stored in a waiting queue until tokens are replenished. Bandwidth limiting is achieved by counting the number of bytes transmitted per unit time, with thresholds set via registers. The control unit refreshes the status of each bucket every microsecond to ensure precise control. Administrators can dynamically modify the QoS parameters of each VF via the PCIe configuration space or a dedicated management channel (such as an I²C interface connected to the BMC). The token bucket algorithm uses a dual-rate model (single-rate dual-bucket), supporting both committed rate and peak rate constraints. All counters and status registers are located in the FPGA Block RAM, with an access latency of less than 10ns. In practical applications, QoS policies can also be issued via an external microcontroller; this embodiment does not limit this.

[0047] How it works: The QoS control unit implements admission control at the I / O entry point, forcing each VF to comply with the predetermined resource quota. Requests exceeding the limit will be delayed instead of being executed immediately, thereby ensuring the stable response of critical business virtual machines.

[0048] Technical results: It achieves microsecond-level resource monitoring accuracy, significantly reduces latency jitter of critical applications in multi-tenant scenarios, effectively prevents abnormal load from affecting other tenants, and meets enterprise-level SLA requirements.

[0049] An FPGA-based NVMe storage array virtualization device includes a field-programmable gate array (FPGA) chip, multiple physical NVMe storage devices connected to the FPGA, and an interface module for implementing PCIe communication. The FPGA is configured to execute control logic for storage virtualization. The FPGA integrates an SR-IOV virtualization module to present the device externally as a standard NVMe endpoint supporting multiple virtual functions (VFs), each VF corresponding to an independent virtual storage volume and directly accessible by different computing instances on the host side. The FPGA also includes a virtual address mapping unit for spatial translation from virtual volume logical block addresses (LBAs) to addresses of the backend physical NVMe storage devices. The FPGA further... The system includes a resource pooling module for striped distribution management of multiple physical NVMe storage devices and synchronously performs RAID checksum calculations during data writing to provide redundancy protection. The resource pooling module establishes parallel data paths with each physical NVMe storage device to achieve I / O request distribution and aggregation. The FPGA also includes a dynamically reconfigurable functional module area for online switching of I / O scheduling strategies, data compression modes, or prefetch algorithms based on real-time workload changes. The SR-IOV virtualization module, virtual address mapping unit, resource pooling module, and dynamically reconfigurable functional module area work together to complete the entire storage virtualization process at the hardware level, thereby achieving low-latency, high-throughput multi-tenant storage services and service quality isolation. Preferably, the resource pooling module also includes bad block management and wear leveling logic to monitor the health status of each physical NVMe storage device, avoid marked failed storage areas during data writing, and dynamically adjust the data mapping position based on the erase / write count distribution to extend the overall lifespan of the storage array.

[0050] The problem this solution addresses is that commercial NVMe SSDs have a limited lifespan, and long-term uneven writing can cause some blocks to fail prematurely, affecting the reliability of the storage array.

[0051] Specific implementation: Bad block management and wear leveling logic are integrated into the resource pooling module. SMART information for each SSD is periodically retrieved via NVMe Admin commands, with a focus on the "Percentage Used," "Media and Data Integrity Errors," and "Critical Warning" fields. When the cumulative erase count of a NAND block exceeds a preset threshold (e.g., 3000 times), it is added to the "about to retire" list. If an uncorrectable ECC error occurs, it is marked as a permanent bad block and access is prohibited. The wear leveling algorithm employs a global dynamic strategy, prioritizing newly written data to physical blocks with low erase counts, while maintaining a persistent bad block mapping table stored in the SSD's reserved namespace. Data redirection is transparently handled by the FPGA's internal address translation layer, unnoticed by upper-layer applications. Health monitoring occurs every minute, and abnormal events are reported to the host management system via AER (Asynchronous Event Request). All metadata is protected using CRC32 checksums to ensure persistent consistency. In practical applications, the wear leveling strategy can also be adjusted to a static hot / cold data separation mode; this embodiment does not limit this approach.

[0052] How it works: This logic checks the health status of the target physical block before each write operation. If a high-risk area is found, it automatically redirects to a healthy block and dynamically updates the mapping relationship to achieve cross-device-level lifetime optimization.

[0053] Technical benefits: It extends the overall lifespan of the storage array, significantly reduces the risk of data loss due to SSD failure, and improves system availability and maintenance cycle.

[0054] An FPGA-based NVMe storage array virtualization device includes a field-programmable gate array (FPGA) chip, multiple physical NVMe storage devices connected to the FPGA, and an interface module for implementing PCIe communication. The FPGA is configured to execute control logic for storage virtualization. The FPGA integrates an SR-IOV virtualization module to present the device externally as a standard NVMe endpoint supporting multiple virtual functions (VFs), each VF corresponding to an independent virtual storage volume and directly accessible by different computing instances on the host side. The FPGA also includes a virtual address mapping unit for spatial translation from virtual volume logical block addresses (LBAs) to addresses of the backend physical NVMe storage devices. The FPGA further... The system includes a resource pooling module for striped distribution management of multiple physical NVMe storage devices and synchronously performs RAID checksum calculations during data writing to provide redundancy protection. The resource pooling module establishes parallel data paths with each physical NVMe storage device to achieve I / O request distribution and aggregation. The FPGA also includes a dynamically reconfigurable functional module area for online switching of I / O scheduling strategies, data compression modes, or prefetch algorithms based on real-time workload changes. The SR-IOV virtualization module, virtual address mapping unit, resource pooling module, and dynamically reconfigurable functional module area work together to complete the entire storage virtualization process at the hardware level, thereby achieving low-latency, high-throughput multi-tenant storage services and service quality isolation. Preferably, the virtual address mapping unit adopts a hierarchical mapping structure, including a high-frequency access page table cache located in the FPGA's on-chip BRAM and a complete mapping table located in external DDR memory. When a cache miss occurs, an asynchronous prefetch of adjacent mapping entries is performed through a background DMA engine to improve the hit rate of subsequent queries.

[0055] The technical problem that this solution addresses is that in a large-scale virtual volume environment, the complete address mapping table is huge (up to GB level) and cannot be entirely stored on the FPGA chip, resulting in increased latency due to frequent access to external memory.

[0056] Specific implementation: The virtual address mapping unit adopts a two-level cache structure. The first-level cache consists of 4MB of BlockRAM within the FPGA, organized as an 8-way set-associative structure with 65,536 TLB entries and an LRU replacement algorithm. The second-level storage is a complete mapping table in external DDR4 memory, managed in 1MB paging units. When a read / write request is received, the corresponding LBA mapping entry is first searched in the BRAM cache. If a match is found, the physical address is returned directly. If a match is not found, the AXI DMA controller loads 16 adjacent mapping entries from the page containing the LBA from the DDR and fills them into the cache in batches. The background prefetch thread scans the cache access log every 5ms, identifies hotspot areas, and actively loads surrounding mapping entries. The mapping entry format includes the virtual LBA, physical SSD ID, device offset, validity flag, and timestamp fields. Address lookup is accelerated using hash indexes, with an average lookup time in the nanosecond range. In practical applications, DDR can also be replaced with DDR5 or HBM2e depending on the FPGA selection to further improve bandwidth; this embodiment does not limit this.

[0057] How it works: By using caching and prefetching mechanisms to reduce the frequency of access to slow external memory, and by leveraging the principle of locality of reference to preload potentially needed mapped data, it maintains a high-efficiency address translation speed.

[0058] Technical results: The mapping query hit rate is improved to over 95%, and the average address translation latency is controlled within 100 ns, effectively supporting the high-performance concurrent access requirements of large-scale virtual volumes (>100).

[0059] An FPGA-based NVMe storage array virtualization device includes a field-programmable gate array (FPGA) chip, multiple physical NVMe storage devices connected to the FPGA, and an interface module for implementing PCIe communication. The FPGA is configured to execute control logic for storage virtualization. The FPGA integrates an SR-IOV virtualization module to present the device externally as a standard NVMe endpoint supporting multiple virtual functions (VFs), each VF corresponding to an independent virtual storage volume and directly accessible by different computing instances on the host side. The FPGA also includes a virtual address mapping unit for spatial translation from virtual volume logical block addresses (LBAs) to addresses of the backend physical NVMe storage devices. The FPGA further... The FPGA includes a resource pooling processing module for striped distribution management of multiple physical NVMe storage devices and synchronously performs RAID checksum calculations during data writing to provide redundancy protection. The resource pooling processing module establishes parallel data paths with each physical NVMe storage device to achieve I / O request distribution and aggregation. The FPGA also includes a dynamically reconfigurable functional module area for online switching of I / O scheduling strategies, data compression modes, or prefetch algorithms based on real-time workload changes. The SR-IOV virtualization module, virtual address mapping unit, resource pooling processing module, and dynamically reconfigurable functional module area work collaboratively to complete the entire storage virtualization process at the hardware level, thereby achieving low-latency, high-throughput multi-tenant storage services and service quality isolation. Preferably, the FPGA connects to multiple physical NVMe storage devices via PCIe switching and acts as a device endpoint connected to the host system, forming a "one virtual machine, multiple pipes, bidirectional direct connection" topology. All I / O data flows do not pass through the host CPU or memory, achieving a completely offloaded end-to-end hardware virtualization path.

[0060] The technical problem this solution addresses is that in traditional architectures, I / O data needs to be transferred through host memory or scheduled by the CPU, which increases path latency and resource consumption, making it impossible to achieve true hardware pass-through.

[0061] Specific implementation: The FPGA chip connects to the host motherboard slot via a PCIe x16 interface, simulating an NVMe endpoint device. Simultaneously, a PCIe switch chip connects the FPGA chip and eight NVMe SSDs. The switch and FPGA chip are interconnected via PCIe x16, and the switch and each SSD are interconnected via PCIe x4 channels. The FPGA integrates a PCS / PMA module to complete the physical layer interface, and the TLP routing logic is responsible for forwarding transaction layer packets between host requests and backend devices. All data transfer is directly completed by the DMA engine within the FPGA, following the path "host <-> FPGA <-> switch <-> SSD," without passing through the CPU or system RAM. The board uses independent DDR memory chips as an intermediate buffer, used only for temporarily storing temporary metadata and prefetched content; user data can be optionally cached. After system power-on, the FPGA loads the bitstream and initializes the NVMe controller, enumerating multiple VF devices to the host. The BIOS or Hypervisor can directly recognize and assign them to VMs. In practical applications, the PCIe switch can also be replaced with a multi-port repeater or a direct-connect topology; this embodiment does not limit this.

[0062] Working principle: The FPGA acts as an intermediate hub, taking over all I / O routing and protocol conversion tasks. The host is regarded as directly connected to a high-performance NVMe device, while the back-end real storage is uniformly managed and scheduled by the FPGA, forming a transparent virtualization path.

[0063] Technical benefits: Completely eliminates the need for the host CPU to participate in I / O processing, reduces end-to-end latency to the limit of pure hardware, enhances system scalability, and supports hot-swapping and online expansion.

[0064] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0065] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An FPGA-based NVMe storage array virtualization device, comprising a field-programmable gate array (FPGA) chip, multiple physical NVMe storage devices connected to the FPGA, and an interface module for implementing PCIe communication, wherein the FPGA is configured to execute control logic for storage virtualization, characterized in that: The FPGA integrates an SR-IOV virtualization module, which presents the device to the outside world as a standard NVMe endpoint that supports multiple virtual functions (VFs). Each virtual function corresponds to an independent virtual storage volume and can be directly accessed by different computing instances on the host side. The FPGA is also equipped with a virtual address mapping unit, which is used to realize the spatial translation of the logical block address (LBA) of the virtual storage volume to the address of the back-end physical NVMe storage device; The FPGA further includes a resource pooling processing module, which is used to perform striped distribution management of multiple physical NVMe storage devices and synchronously perform RAID verification calculations during data writing to provide redundancy protection; the resource pooling processing module establishes parallel data paths with each physical NVMe storage device to realize the splitting and aggregation of I / O requests; The FPGA also includes a dynamically reconfigurable functional module area, which is used to switch I / O scheduling strategies, data compression modes or prefetch algorithms online according to real-time workload changes. Among them, the SR-IOV virtualization module, virtual address mapping unit, resource pooling processing module and dynamically reconfigurable functional module area work together to complete the entire process of storage virtualization at the hardware level, thereby achieving low latency, high throughput multi-tenant storage service and service quality isolation.

2. The FPGA-based NVMe storage array virtualization device according to claim 1, characterized in that: The dynamically reconfigurable functional module area includes multiple pre-configured logical functional sub-modules, and the logical functional sub-modules include at least an adaptive I / O scheduler, a lightweight data compression engine, and an intelligent prefetch controller based on access pattern recognition. The FPGA dynamically loads corresponding logic function sub-modules based on the received performance monitoring feedback information or policy instructions issued by the host to optimize I / O processing efficiency under the current workload.

3. The FPGA-based NVMe storage array virtualization device according to claim 2, characterized in that: The intelligent prefetch controller is configured to periodically collect the I / O access trajectory of each virtual storage volume, and use a finite state machine model to classify and identify the access patterns. When a sequential or step-size regular access is detected, the prefetch process is automatically started, and the predicted subsequent data blocks are loaded from the back-end physical NVMe storage device to the FPGA on-chip cache or external high-speed DDR memory.

4. The FPGA-based NVMe storage array virtualization device according to claim 1, characterized in that: The SR-IOV virtualization module also integrates a QoS control unit, which is used to independently configure the bandwidth limit, IOPS quota and burst traffic window for each virtual function (VF), and to measure and regulate the I / O requests of each virtual function in real time through the token bucket algorithm to prevent resource preemption and achieve service quality isolation between multi-tenants.

5. The FPGA-based NVMe storage array virtualization device according to claim 1, characterized in that: The resource pooling processing module also includes bad block management and wear leveling logic, which is used to monitor the health status of each physical NVMe storage device, avoid marked failed storage areas during data writing, and dynamically adjust the data mapping position according to the erase and write count distribution of each storage area to extend the overall lifespan of the storage array.

6. The FPGA-based NVMe storage array virtualization device according to claim 1, characterized in that: The virtual address mapping unit adopts a hierarchical mapping structure, including a high-frequency access page table cache located in the FPGA on-chip BRAM and a complete mapping table located in external DDR memory. When a cache miss occurs, adjacent mapping entries are asynchronously prefetched through the background DMA engine to improve the hit rate of subsequent queries.

7. The FPGA-based NVMe storage array virtualization device according to claim 1, characterized in that: The FPGA connects to multiple physical NVMe storage devices via PCIe switching and connects to the host system as an NVMe device endpoint, forming a "one virtual machine with multiple pipes and bidirectional direct connection" topology. Under the control of the FPGA chip, all I / O data streams are directly transmitted between the host system and the physical NVMe storage devices without passing through the host CPU or memory, achieving a completely offloaded end-to-end hardware virtualization path.

8. The FPGA-based NVMe storage array virtualization device according to claim 1, characterized in that: The virtual address mapping unit adopts a page-based mapping structure, with each page being 4KB in size. The mapping table entries are stored in the on-chip BRAM of the FPGA chip, and the mapping table supports the simultaneous operation of no less than 64 virtual memory volumes.

9. The FPGA-based NVMe storage array virtualization device according to claim 1, characterized in that: The resource pooling processing module divides all physical NVMe storage devices into fixed-size data blocks and distributes them in a striped manner according to RAID 5. When writing data, the processing module automatically splits the data and sends it in parallel to multiple physical NVMe storage devices.

10. The FPGA-based NVMe storage array virtualization device according to claim 1, characterized in that: The SR-IOV virtualization module is instantiated in the FPGA chip as 4 to 32 independent virtual functions. Each virtual function has an independent submission queue and completion queue, and communicates with the host system through the MSI-X interrupt mechanism.

Citation Information

Cited By

  • Heterogeneous storage acceleration system and method for edge ai computing

    CN122131988A