Intelligent network card OVS dynamic unloading acceleration method and system

Abstract

The invention discloses an intelligent network card OVS dynamic unloading acceleration method and system, and the method comprises the steps: constructing a dynamic reconfigurable matching-action assembly line on an intelligent network card, and the assembly line is composed of a plurality of micro-engines; receiving a data packet, and dynamically selecting and assembling a micro-engine sequence required for processing the data packet according to data stream characteristics and a current network strategy; carrying out pooling and virtualization management on hardware resources of the intelligent network card, and distributing independent virtual forwarding examples for a plurality of tenants; a shared memory area is established between the host and the intelligent network card and is used for synchronizing flow table operation and statistical information; and deploying a configurable telemetering probe in the data path, and collecting and reporting network processing state information. According to the method, the problems of assembly line stiffness, resource bottleneck, large collaboration overhead and weak observability in the prior art are solved, and OVS full unloading with high performance, high resource utilization rate, multi-tenant isolation and transparent operation and maintenance is realized.

Description

Technical Field

[0001] This invention relates to a method and system for dynamic offloading and acceleration of smart network interface cards (NICs) via Offloading Virtualization (OVS), and more particularly to a method and system for dynamic offloading and acceleration of smart NICs via Offloading Virtualization (OVS) based on dynamic pipeline reconfiguration and hardware resource virtualization. Background Technology

[0002] With the widespread adoption of cloud-native, microservice, and serverless architectures, data center networks are placing higher demands on forwarding performance, resource isolation, and observability. As a core component of virtualized networks, Open vSwitch's data plane performance has become a bottleneck for the overall system. Existing OVS offloading solutions suffer from the following problems: statically fixed pipelines that struggle to adapt to dynamic policy changes; conflicts between OVS flow tables and TCAM resource constraints, leading to "rule explosion"; low coordination efficiency between the control plane and data plane, resulting in high synchronization overhead; severe resource contention in multi-tenant environments, lacking strict isolation and fair scheduling; and a significant decrease in network observability after hardware offloading.

[0003] Therefore, it is of great significance to provide a method and system for dynamic offloading acceleration of intelligent network interface cards (OVS) based on dynamic pipeline reconfiguration and hardware resource virtualization. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a method and system for dynamic offloading acceleration of intelligent network interface cards (OVS), aiming to solve the above-mentioned problems and achieve efficient collaboration between the OVS data plane and control plane.

[0005] Technical solution: The present invention provides a method for dynamic offloading and acceleration of intelligent network interface cards (OVS), comprising:

[0006] A dynamically reconfigurable matching action pipeline is constructed on the smart network interface card, the pipeline consisting of multiple micro-engines;

[0007] Receive data packets, and dynamically select and assemble the micro-engine sequence required to process the data packets based on data stream characteristics and current network policies;

[0008] The hardware resources of smart network interface cards are pooled and virtualized for management, and independent virtual forwarding instances are allocated to multiple tenants;

[0009] A shared memory area is established between the host and the smart network card for flow table operations and statistical information synchronization;

[0010] Deploy configurable telemetry probes in the data path to collect and report network processing status information.

[0011] Furthermore, the dynamically reconfigurable matching-action pipeline is controlled by a dynamic pipeline manager, which receives policy instructions from the host control plane and adjusts the connection relationships and processing logic between microengines in real time.

[0012] Furthermore, the microengine includes a programmable microengine and a fixed-function microengine, wherein the programmable microengine supports dynamic loading and configuration based on FPGA or eBPF instruction sets.

[0013] Furthermore, the hardware resources include TCAM, SRAM, counters, and status tables, and the pooling management includes abstracting the resources into a unified resource pool and allocating quota-limited virtual resources to each tenant through a management program.

[0014] Furthermore, the method also includes a two-level flow table management mechanism:

[0015] The first-level flow table is stored in TCAM and is used to accurately match high-priority flows;

[0016] The second-level flow table is stored in SRAM, using an efficient hash algorithm to store low-frequency flow table entries, and supports a flow table prefetching algorithm based on machine learning prediction, which preloads predicted flow table entries from the host into the SRAM cache.

[0017] Furthermore, the shared memory area is used to store the flow table operation queue and the statistics counter. The host control plane updates the flow table by writing to the queue, the smart network card obtains the update by polling, and the statistics are triggered for synchronization based on the change threshold or read directly by the control plane.

[0018] Furthermore, the telemetry probe collects latency, packet loss, and path tracing information during data packet processing according to user configuration; and reports the encapsulated telemetry data to an external monitoring system through a channel independent of the data plane.

[0019] The present invention discloses an intelligent network interface card (NIC) OVS dynamic offloading acceleration system, comprising:

[0020] The dynamic pipeline management module is used to control the assembly and scheduling of micro-engines;

[0021] The resource virtualization management module is used for hardware resource pooling and multi-tenant isolation.

[0022] A shared memory management module is used for data synchronization between the host and the network card.

[0023] The telemetry data acquisition module is used for the observability output of network status.

[0024] The computer device of the present invention includes one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the steps of the intelligent network interface card OVS dynamic offloading acceleration method.

[0025] The present invention discloses a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the intelligent network interface card (OVS) dynamic offloading acceleration method.

[0026] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0027] (1) This invention constructs a dynamically reconfigurable matching-action pipeline to realize on-demand assembly of data packet processing paths, overcoming the shortcomings of traditional static pipelines that are difficult to respond to network policy changes in real time, and enabling smart network cards to flexibly adapt to the frequently changing traffic demands in multi-cloud and heterogeneous cloud environments.

[0028] (2) Hardware resource virtualization and pooling management are adopted, combined with a two-level flow table mechanism to improve flow table storage efficiency. A machine learning-driven flow table prefetching algorithm is introduced to ensure stable data plane performance under high load scenarios.

[0029] (3) By establishing a shared memory channel between the host and the smart network card, efficient synchronization of flow table operations and statistical information is achieved. The control plane updates the rules by directly writing to the shared queue, and the network card obtains the updates by polling, avoiding frequent interruptions and PCIe bandwidth occupation, significantly reducing CPU overhead and communication latency, and improving the overall system throughput.

[0030] (4) This invention lays the technical foundation for realizing network self-optimization, self-repair and self-security, and has forward-looking and strategic value. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method described in this invention. Detailed Implementation

[0032] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0033] like Figure 1 As shown, this embodiment provides a method for dynamic offloading and acceleration of smart network interface cards (OVS), including:

[0034] (1) Construct a dynamically reconfigurable matching-action pipeline on the smart network card, wherein the pipeline is composed of multiple micro engines;

[0035] (2) Receive data packets and dynamically select and assemble the micro-engine sequence required to process the data packets based on the data flow characteristics and the current network policy;

[0036] (3) Pool and virtualize the hardware resources of the smart network interface card and allocate independent virtual forwarding instances to multiple tenants;

[0037] (4) Establish a shared memory area between the host and the smart network card for flow table operations and statistical information synchronization;

[0038] (5) Deploy configurable telemetry probes in the data path to collect and report network processing status information.

[0039] Specifically, it includes:

[0040] The dynamically reconfigurable matching-action pipeline is controlled by a dynamic pipeline manager (DPM), which receives policy instructions from the host control plane and adjusts the connection relationships and processing logic between microengines in real time.

[0041] The microengine includes a programmable microengine and a fixed-function microengine. The programmable microengine supports dynamic loading and configuration based on FPGA or eBPF instruction sets, while the fixed-function microengine implements NAT, VxLAN encapsulation / decapsulation, checksum calculation, etc., for example, using ASIC integrated circuits.

[0042] Dynamic Pipeline Manager (DPM) dynamically assembles processing paths as packets arrive, based on the characteristics of the data flow (such as protocol type and priority) and the current policy. For example, an L2 forwarding flow might only pass through three micro-engines, while a flow requiring Deep Packet Inspection (DPI) and firewall processing might be routed to a path with more processing units. This on-demand assembly improves hardware efficiency.

[0043] The dynamic assembly of micro-engine sequences is based on real-time data flow characteristics and dynamic network policies. As data packets enter the smart NIC, a set of the most suitable micro-engines is selected and connected in real time, forming a highly efficient and customized processing pipeline. The control plane (such as the OVS controller) distributes network policies (such as flow tables, security policies, and QoS rules) to the smart NIC. Data packets enter the smart NIC's ingress queue. The data flow feature extractor quickly parses the packet header, extracting key features: the 5-tuple (source / destination IP, port, protocol), VLAN ID, tunnel identifier, priority field, etc.

[0044] Based on the extracted features, the micro-engine combination rule table is queried, and the best matching strategy is selected. If there is no perfect match, the longest prefix matching plus the default path strategy is used. An ordered sequence of micro-engine IDs is generated.

[0045] The hardware resources include TCAM, SRAM, counters, and status tables. The pooling management includes abstracting the resources into a unified resource pool and allocating quota-limited virtual forwarding instances to each tenant through a lightweight management program.

[0046] The lightweight management program is deployed on the control plane of the smart NIC, typically running on the integrated management core of the NIC (such as an ARM Cortex-A series processor). This program abstracts the physical resources of the smart NIC (including TCAM, SRAM, counters, DMA channels, etc.) into a unified resource pool and allocates a quota-limited Virtual Forwarding Instance (VFI) to each tenant. Each VFI enjoys an independent pipeline configuration space, flow table storage area, and statistical counter group, and hardware isolation mechanisms ensure performance and security isolation between tenants.

[0047] The lightweight management program works in conjunction with the Dynamic Pipeline Manager (DPM): When the DPM needs to assemble a pipeline for a tenant, it first requests hardware resources from the lightweight management program. The management program allocates resources according to the tenant's resource quota and configures the corresponding access permissions and address mappings. After resource allocation is completed, the DPM can then connect to the microengine and configure its parameters, thereby forming a virtual forwarding path dedicated to the tenant.

[0048] The method also includes a two-level flow table management mechanism:

[0049] The first-level flow table is stored in TCAM and is used to accurately match high-priority flows;

[0050] The second-level flow table is stored in SRAM, using an efficient hash algorithm to store low-frequency flow table entries, and supports a flow table prefetching algorithm based on machine learning prediction, which preloads predicted flow table entries from the host into the SRAM cache.

[0051] The flow table prefetching algorithm based on machine learning prediction is specifically an intelligent prefetching method based on Temporal Attention Network (TA-LSTM), which includes the following steps:

[0052] Feature collection and sequence construction: The smart NIC collects features such as timestamp, five-tuple hash value, access type, packet size, and tenant identifier for each flow table entry, and constructs an access time sequence according to time windows;

[0053] Model training and deployment: The TA-LSTM model is trained on the host side using historical traffic data, and then deployed to the embedded processor or programmable logic unit of the smart network card after lightweight compression.

[0054] Real-time prediction and prefetching: The model predicts the top-K most likely flow table entries to be accessed in the future based on the current access sequence. If there is free space in the SRAM cache, it prefetches from the host memory to the SRAM cache in batches via DMA.

[0055] Feedback optimization: The system records the prefetch hit rate and fine-tunes the model parameters online accordingly to achieve adaptive traffic pattern changes.

[0056] A shared memory area is established between the host and the smart network interface card (NIC). This shared memory area stores flow table operation queues and statistical counters. The host control plane updates the flow table by writing to the queue. CRUD operations on flow table entries are performed by directly writing to the shared memory queue. The NIC obtains updates through polling instead of interrupts, significantly reducing latency and CPU overhead. The smart NIC obtains updates through polling, and statistical information is either triggered by changing thresholds or directly read by the control plane.

[0057] The telemetry probe collects information on delay, packet loss, and path tracing during data packet processing according to user configuration; it defines a standard telemetry data format and allows users to dynamically configure the types of telemetry data to be collected and the sampling frequency through the control plane, and reports the encapsulated telemetry data to the external monitoring system through a channel independent of the data plane.

[0058] The telemetry probes are non-invasive probes implanted at key nodes in the smart network interface card's (NIC) data path. These key nodes are processing stages or modules that have a decisive impact on network performance, latency, packet loss, forwarding decisions, resource utilization, and fault diagnosis. Selecting these nodes to implant non-invasive telemetry probes enables refined monitoring and improved observability of the entire data packet processing flow. Non-invasive probes are implanted at key nodes in the smart NIC data path.

[0059] This embodiment also provides a smart network interface card (NIC) OVS dynamic offloading acceleration system, including:

[0060] The dynamic pipeline management module is used to control the assembly and scheduling of micro-engines;

[0061] The resource virtualization management module is used for hardware resource pooling and multi-tenant isolation.

[0062] A shared memory management module is used for data synchronization between the host and the network card.

[0063] The telemetry data acquisition module is used for the observability output of network status.

[0064] The system described above is based on the same inventive concept as the method described above, and will not be repeated here.

[0065] This embodiment also provides a computer device, including one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the steps of the smart network interface card OVS dynamic offloading acceleration method.

[0066] This embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned smart network interface card (OVS) dynamic offloading acceleration method.

Claims

1. An intelligent network card OVS dynamic offload acceleration method, characterized in that, include: A dynamically reconfigurable matching action pipeline is constructed on the smart network interface card, the pipeline consisting of multiple micro-engines; Receive data packets, and dynamically select and assemble the micro-engine sequence required to process the data packets based on data stream characteristics and current network policies; The hardware resources of smart network interface cards are pooled and virtualized for management, and independent virtual forwarding instances are allocated to multiple tenants; A shared memory area is established between the host and the smart network card for flow table operations and statistical information synchronization; Deploy configurable telemetry probes in the data path to collect and report network processing status information. 2.The method of claim 1, wherein, The dynamically reconfigurable matching action pipeline is controlled by a dynamic pipeline manager, which receives policy instructions from the host control plane and adjusts the connection relationships and processing logic between microengines in real time. 3.The method of claim 2, wherein, The microengine includes a programmable microengine and a fixed-function microengine, wherein the programmable microengine supports dynamic loading and configuration based on FPGA or eBPF instruction sets. 4.The method of claim 1, wherein, The hardware resources include TCAM, SRAM, counters, and status tables. The pooling management includes abstracting the resources into a unified resource pool and allocating quota-limited virtual resources to each tenant through a management program.

5. The method for dynamic offloading and acceleration of an intelligent network interface card (NIC) under OVS according to claim 4, characterized in that, The method also includes a two-level flow table management mechanism: The first-level flow table is stored in TCAM and is used to accurately match high-priority flows; The second-level flow table is stored in SRAM. It uses an efficient hash algorithm to store low-frequency flow table entries and supports a flow table prefetching algorithm based on machine learning prediction, which preloads predicted flow table entries from the host into the SRAM cache.

6. The method of claim 1, wherein the method further comprises: The shared memory area is used to store the flow table operation queue and the statistics counter. The host control plane updates the flow table by writing to the queue, and the smart network card obtains the updates by polling. The statistics are triggered by the change threshold or read directly by the control plane.

7. The method of claim 1, wherein the method further comprises: The telemetry probe collects information on delay, packet loss, and path tracing during data packet processing according to user configuration; and reports the encapsulated telemetry data to an external monitoring system through a channel independent of the data plane.

8. A smart network interface card (NIC) OVS dynamic offloading acceleration system, characterized in that, include: The dynamic pipeline management module is used to control the assembly and scheduling of micro-engines; The resource virtualization management module is used for hardware resource pooling and multi-tenant isolation. A shared memory management module is used for data synchronization between the host and the network card. The telemetry data acquisition module is used for the observability output of network status.

9. A computer device, comprising: It includes one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the steps of a smart network interface card OVS dynamic offloading acceleration method as described in any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the OVS dynamic offloading acceleration method for a smart network interface card as described in any one of claims 1-7.

Images