A microservice mesh system that provides cloud computing services

By isolating sandbox proxy instances by tenant and using zero-copy delivery paths in the borderless vehicle service mesh architecture, the problems of resource preemption and cascading avalanche are solved, achieving efficient traffic processing and congestion prediction, and improving the stability and performance of the system.

CN123123418APending Publication Date: 2026-11-17XINJIANG XINGYUHUI INFORMATION TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611499005.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-24
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Under the borderless vehicle service mesh architecture, the existing solution fails to effectively achieve traffic partitioning and isolation by tenant, resulting in resource preemption and cascading avalanche. Furthermore, there is significant data copying and context switching overhead during traffic offloading, making it impossible to perceive the true congestion status of deep downstream links.

Method used

By dynamically creating sandbox proxy instances per tenant, combined with zero-copy transmission paths and backpressure awareness and prediction modules, traffic isolation by tenant and fault domain isolation are achieved. Furthermore, by embedding comprehensive link load confidence information in the packet header, the congestion trend of deep downstream links can be perceived for proactive processing.

Benefits of technology

It achieves application-layer traffic resource isolation between multiple tenants, reduces data copying and context switching overhead, avoids cascading avalanches, and improves system processing efficiency and stability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123418A_ABST
    Figure CN123123418A_ABST
Patent Text Reader

Abstract

This invention discloses a microservice mesh system for providing cloud computing services. The invention relates to the field of cloud-native microservice mesh technology. The system includes a control plane module and a data plane module deployed on each computing node. The data plane module parses data packets using protocol probes mounted on the network interface card (NIC) driver layer, distinguishing between kernel-direct traffic and complex application-layer traffic. For traffic requiring offloading, it dynamically creates sandbox proxy instances constrained by control group resource limits for each tenant. Data transfer between the NIC and the sandbox proxy is achieved based on zero-copy memory mapping. Simultaneously, a link-wide load confidence field is embedded in the packet header. The ingress proxy of the call chain performs congestion trend prediction and performs tiered processing. This invention achieves strong multi-tenant isolation, reduces traffic offloading overhead, and suppresses the risk of cascading avalanches.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of cloud-native microservice mesh technology, specifically to a microservice mesh system that provides cloud computing services. Background Technology

[0002] With the development of cloud computing, microservice architecture has been widely used in various cloud-native business systems due to its characteristics such as agile development and module decoupling. In order to separate business code from service governance logic such as service discovery, traffic control, security authentication, and observability, service mesh technology has emerged. Its core function is to transparently take over the communication traffic between services and execute corresponding governance strategies without intruding on business code.

[0003] Currently, the mainstream solutions in the industry generally adopt the sidecar mode: the control plane dynamically distributes configurations such as routing, circuit breaking, and authentication to a dedicated proxy process injected into each business container group through the discovery service protocol. The data plane is intercepted and processed by the proxy process, which handles all incoming and outgoing traffic of the corresponding container group. In this mode, each business container group needs to run an independent proxy process. As the number of container groups increases, the computing resources occupied by the proxy process itself and the deployment and maintenance costs increase accordingly.

[0004] To reduce the resource overhead of the sidecar mode, the industry has further proposed a sidecar-free architecture. A representative solution deploys a unified shared agent at the compute node level and uses extended Berkeley packet filter technology mounted on hooks in the kernel network stack to intercept and accelerate traffic forwarding. All tenant business container groups running on this node share the same user-space agent process to handle application layer traffic involving complex business semantics, thus eliminating the overhead of injecting a separate agent process for each business container group. At the same time, in this solution, the agent of each compute node usually only performs rate limiting and circuit breaking policies based on local indicators such as local queue length and local error rate. Summary of the Invention

[0005] Existing edgeless service mesh solutions deploy a unified user-space proxy process on the same computing node to handle application layer traffic from all tenants. This proxy process lacks hard isolation of resources and fault domains for different tenants. When a tenant experiences sudden traffic spikes or abnormal requests, it can consume processing resources from other tenants, leading to resource contention. Furthermore, if this shared proxy process malfunctions, it will affect the business traffic of all tenants on that node. Additionally, when traffic is offloaded to the user-space proxy for processing, data packets undergo multiple copies and context switches between the network interface card, kernel network protocol stack, and the user-space proxy, resulting in significant processing latency and processor resource overhead. Moreover, in existing solutions, the rate limiting and circuit breaking strategies of each node proxy are based solely on local indicators such as local queue length and local error rate, failing to perceive the true congestion status of deep downstream service links, easily triggering cascading avalanche phenomena.

[0006] The technical problems to be solved by this invention are: how to achieve high-performance traffic forwarding by utilizing node-level shared kernel acceleration capabilities under a borderless vehicle service mesh architecture, while providing strict isolation for application layer traffic involving complex business semantics by tenant, how to reduce the overhead of kernel-mode and user-mode data copying and context switching during traffic offloading to user-mode proxies, and how to enable proxies at all levels in the call chain to perceive the real load status of deep downstream links, and to implement probabilistic dropping or degradation processing in advance at the traffic source to suppress the occurrence of cascading avalanches.

[0007] To address the aforementioned technical problems, this invention provides a microservice mesh system for providing cloud computing services, including a control plane module and a data plane module deployed on each computing node of a cloud-native cluster. The data plane module includes a protocol probe and decision engine module, a sandbox proxy pool manager, a zero-copy delivery module, and a backpressure awareness and prediction module. The protocol probe and decision engine module is mounted on the network interface card (NIC) driver layer and parses data packets arriving at the node's NIC. Based on the parsing results, the data packets are classified into a first type of data packet that does not involve complex application layer features and a second type of data packet that does involve complex application layer features. The first type of data packet is directly forwarded in kernel mode after querying the local forwarding table. For the second type of data packet, the corresponding tenant identifier is extracted. The sandbox proxy pool manager queries the tenant security context based on the tenant identifier and dynamically creates or reuses dedicated security contexts according to the tenant dimension. Sandbox proxy instances are configured, and each sandbox proxy instance is allocated an independent control group resource limit and an independent user-space and kernel-space shared memory region to achieve strong isolation between multi-tenants. The zero-copy delivery module establishes a zero-copy data frame delivery path between the network card and the sandbox proxy instance based on a filling ring queue, a receiving ring queue, a sending ring queue, and a completion ring queue. This allows data packets that need to be offloaded to be processed in user space without having to go through multiple copies in the traditional kernel network protocol stack. The backpressure perception and prediction module embeds the link comprehensive load confidence and backpressure signaling fields in the packet header. The agents at each level of the call chain read and update these fields, and the processing nodes at the call chain entry point perform trend prediction based on the historical sequence of these fields. At the traffic source, requests with a high predicted congestion probability are probabilistically dropped or rejected at the source.

[0008] It should be noted that existing edgeless service mesh solutions generally believe that node-level shared user-space agents can eliminate the resource overhead of injecting agents into container groups per container in the edge car mode. There is no need to perform hard isolation of the shared agent by tenant resources and fault domains, nor is there a need to be aware of the real congestion status of deep downstream links. This invention breaks through the above conventional understanding by dynamically creating independent sandbox agent instances constrained by control group resource quotas by tenant, and coordinating them with zero-copy delivery paths based on circular queues and deep congestion prediction based on the comprehensive load confidence of packet header links. Without introducing the resource overhead of the edge car mode, it simultaneously solves the three types of technical problems that originally needed to be addressed separately: resource preemption, data copy overhead, and cascading avalanche, achieving non-obvious technical effects.

[0009] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0010] Firstly, by dynamically creating dedicated sandbox proxy instances for each tenant and constraining them with control group resource limits, the advantages of node-level resource sharing are retained while resource isolation and fault domain isolation of application layer traffic processing among multiple tenants are achieved, avoiding the impact of abnormal traffic from a single tenant on resource preemption of other tenants.

[0011] Secondly, by using a zero-copy data transmission path based on filling the ring queue, receiving the ring queue, sending the ring queue, and completing the ring queue, data packets can bypass the traditional kernel network protocol stack and be mapped and processed directly in the user space and kernel space shared memory area of ​​the sandbox proxy instance, reducing the data copy and context switching overhead during traffic offloading.

[0012] Third, by progressively embedding and updating the link comprehensive load confidence information in the message header, the processing node on the call chain entry side can detect the congestion trend of deep downstream links in advance and implement hierarchical processing at the traffic source without relying on centralized topology management components, thereby reducing the possibility of cascading avalanches caused by local rate limiting failures. Attached Figure Description

[0013] Figure 1 This is a diagram illustrating the overall architecture of the microservice mesh system of this invention.

[0014] Figure 2 This is a flowchart of the data processing of the data plane module of the present invention;

[0015] Figure 3 This is a flowchart of the back pressure sensing and prediction processing of the present invention. Detailed Implementation

[0016] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0017] The microservice mesh system for providing cloud computing services provided by this invention is deployed in each computing node of a cloud-native cluster in the form of container network interface plugins and daemon sets, and is divided into two parts: control plane and data plane.

[0018] The control plane module is located on the management side of the cluster. Internally, it includes a configuration generation unit, a tenant security context storage unit, and a backpressure prediction model distribution unit. The configuration generation unit generates protocol identification rules for use by the data plane modules of each compute node, based on the governance policies configured by operations personnel. These rules include a protocol feature rule table for determining packet types. The tenant security context storage unit centrally maintains the tenant security context for each tenant. The tenant security context includes a tenant identifier, the corresponding resource quota information, and the corresponding isolation policy information. The isolation policy information includes the upper limit of control group resource quotas allowed to be allocated to the sandbox proxy instance corresponding to the tenant. The backpressure prediction model distribution unit distributes the parameters required for the local prediction model to the processing nodes at the call chain entry point. These parameters include weight coefficients, smoothing coefficients, and judgment thresholds. The control plane module connects to the data plane modules deployed on each compute node through a management channel, distributing the aforementioned protocol identification rules, tenant security contexts, and backpressure prediction model parameters to the corresponding data plane modules.

[0019] The data plane module is deployed on each compute node, and from the network interface card (NIC) hardware layer upwards, it includes: a protocol probe and decision engine module mounted on the NIC driver layer, which contains high-speed data path processing hooks; a user-space and kernel-space shared memory area provided by the zero-copy transfer module, which is surrounded by a fill ring queue, a receive ring queue, a send ring queue, and a completion ring queue; a user-space sandbox proxy pool managed by the sandbox proxy pool manager, in which multiple sandbox proxy instances belonging to different tenants run in parallel, and each sandbox proxy instance is subject to its own independent control group resource quota constraints; each compute node also runs multiple service container groups hosted on that node; traffic sent and received by each service container group is intercepted by the NIC and protocol probe and decision engine module of that node, and flows to the kernel-space direct path or the user-space sandbox proxy path according to the judgment result, and finally flows out of the compute node; the cross-node service call traffic between nodes and between nodes and external gateways carries backpressure signaling message identifiers.

[0020] The protocol probe and decision engine module is mounted on the high-speed data path processing hook point of the network card driver layer. When a data packet arrives at the node network card, the module loads the corresponding parsing program to parse the data packet in real time, extracting the five-tuple information and application layer feature bytes of the data packet. The five-tuple information includes the source IP address, source port, destination IP address, destination port, and protocol type. The application layer feature bytes include the transport layer security protocol handshake identifier field and the hypertext transfer protocol frame header identifier field. According to the protocol feature rule table issued by the control plane module, the protocol probe and decision engine module matches and determines the extracted five-tuple information and application layer feature bytes: if the determination result is that the data packet belongs to the transport layer regular traffic or a simple message of fixed length, and does not involve application layer semantic processing requirements such as circuit breaking, multi-tenant authentication, and retry, then the data packet is marked as kernel direct conversion type. If the determination result is that the data packet hits the application layer complex feature, then the data packet is marked as user-space offload type, and the corresponding tenant identifier field is extracted from the data packet.

[0021] For packets marked as kernel direct forwarding type, the protocol probe and decision engine modules directly query the local forwarding table to obtain the target forwarding path, and redirect the packet to the corresponding network interface through the channel provided by the zero-copy delivery module to complete the forwarding. This process does not require any user-space proxy, achieving zero-stagnation forwarding of simple traffic.

[0022] For packets marked as user-mode offload type, the protocol probe and decision engine module extracts the corresponding tenant identifier field from the packet and passes the tenant identifier to the sandbox proxy pool manager. The sandbox proxy pool manager queries the tenant security context cache stored in the kernel-mode mapping table based on the received tenant identifier to obtain the resource isolation policy corresponding to the tenant and the current sandbox proxy instance status.

[0023] The sandbox proxy pool manager determines whether the sandbox proxy instance corresponding to the tenant is alive. If the sandbox proxy instance is alive, the memory page corresponding to the data packet is directly mapped to the user-mode and kernel-mode shared memory region of the sandbox proxy instance by filling the circular queue and receiving the circular queue. If the sandbox proxy instance corresponding to the tenant does not exist or has been reclaimed, the sandbox proxy pool manager dynamically creates a new sandbox proxy instance according to the tenant's resource isolation policy, allocates the corresponding control group resource limit to the instance, and allocates an independent user-mode and kernel-mode shared memory region for it. Then, the memory page corresponding to the data packet is mapped to the user-mode and kernel-mode shared memory region of the newly created sandbox proxy instance. Different tenants have different sandbox proxy instances. Each instance runs within its own independent control group resource boundary. When a sandbox proxy instance of one tenant experiences an anomaly (such as memory overflow or crash), it will not affect the normal operation of the corresponding sandbox proxy instances of other tenants. This achieves resource isolation and fault domain isolation for application layer traffic processing among multiple tenants, avoiding resource preemption and isolation failure issues caused by shared user-space proxies. After the data packet is mapped to the corresponding sandbox proxy instance, the instance completes application layer semantic processing in user space. The application layer semantic processing includes circuit breaker determination, multi-tenant authentication verification, and retry policy execution. After the sandbox proxy instance completes the processing, it writes the processing result back in a zero-copy manner through the sending ring queue and the completion ring queue. The protocol probe and decision engine module then completes the subsequent forwarding operations.

[0024] The zero-copy delivery module establishes a zero-copy data transfer path between the network interface card (NIC) and the user-space sandbox proxy instance based on a shared memory region between user space and kernel space (i.e., a shared memory region used for zero-copy data frame transfer between user space and kernel space, hereinafter referred to as the shared memory region). Four types of circular queues are set up around this shared memory region: the NIC driver obtains free memory block descriptors from the filling circular queues, writes the received data frames into the corresponding memory blocks in the shared memory region, and then notifies the user-space sandbox proxy instance through the receiving circular queue. After the user-space sandbox proxy instance completes its processing, it submits the memory block descriptors to be sent through the sending circular queue. After the NIC driver completes data transmission... By completing the circular queue, the corresponding memory block descriptor is returned so that the memory block can be reused by subsequent data frames. Throughout the data transmission process, the data frame is always kept in the shared memory area. The user-mode sandbox agent instance directly performs read and write operations on the shared memory area without going through multiple data copies and kernel-user mode context switches of the traditional kernel network protocol stack. This reduces the processing latency and processor resource overhead introduced by traffic offloading, so that the introduction of dynamic sandbox agents will not become a performance bottleneck of the system. The shared memory area corresponding to each sandbox agent instance is independent of each other and is located within the control group resource boundary of the corresponding tenant, so as to reflect the hard isolation relationship of resources.

[0025] This invention redefines the header structure of inter-microservice communication messages. In addition to the conventional field area (including request identifier, tenant identifier, and original business fields), a backpressure extension field area is added. This backpressure extension field area is further subdivided into a link comprehensive load confidence field, a local timestamp field, and a backpressure signaling flag field. The link comprehensive load confidence field is used to carry load status information obtained by downstream service nodes through hierarchical aggregation. The local timestamp field is used to identify the collection time of the load status information, allowing the receiving end to determine the timeliness of the information. The backpressure signaling flag field is used to indicate whether the message carries valid backpressure information.

[0026] After completing request processing, the data plane proxy of a downstream service node located deep in the call chain collects the current load metrics of its node. These load metrics include the current number of requests in the queue, the current processor utilization, and the average response latency within the current statistical window. Based on these load metrics, the node's load confidence is calculated using the following method:

[0027] ;

[0028] in, This indicates the load confidence level of this node. Indicates the current number of requests in the queue. This indicates the preset maximum queue capacity of this node. This represents the current processor utilization rate, with values ​​ranging from greater than 0 to less than 1. This represents the average response time within the current statistical window. This indicates the preset service level agreed latency threshold. , , These are the weight coefficients corresponding to the queue backlog percentage, processor utilization, and latency percentage, respectively, and they satisfy the following conditions: In a specific example, it can be set , , .

[0029] The data plane proxy further weights and fuses the load confidence of this node with the deep load confidence already carried in the forwarded packets from downstream nodes to obtain the comprehensive link load confidence to be implanted in this response packet. The calculation method is as follows:

[0030] ;

[0031] in, This indicates the overall link load confidence level written to the packet header this time. This indicates the confidence level of the deep payload already carried in the downstream forwarded message. This is the local weighting coefficient, with a value range greater than 0 and less than 1. In a specific example, it can be set to... After the calculated link load confidence is written into the corresponding field of the message's custom header, the message is passed upstream along the call chain. Each intermediate agent that the message passes through repeats the above collection, calculation and fusion processing to achieve the aggregation and propagation of deep load status along the call chain.

[0032] The processing node (i.e., gateway or first-layer proxy) on the ingress side of the call chain continuously collects the comprehensive load confidence sequence of the links arriving within the historical time window, and calculates the exponentially weighted moving average and the first-order difference value of the sequence in turn. The calculation methods are as follows:

[0033] ;

[0034] ;

[0035] in, This represents the exponentially weighted moving average calculated at time t. This represents the overall link load confidence level arriving at time t. This is a smoothing coefficient, with a value range greater than 0 and less than 1. In a specific example, it can be set to... , The first-order difference of the link's overall load confidence between time t and time t-1 is used to reflect the rate of change in congestion trends.

[0036] The processing node on the ingress side of the calling chain inputs the above exponentially weighted moving average and first-order difference value into the local prediction model to calculate the predicted congestion probability within the short-term time window. The calculation method is as follows:

[0037] ;

[0038] in, This represents the predicted congestion probability calculated at time t. This represents the Sigmoid normalization function, used to map the result of a linear combination to an interval greater than 0 and less than 1. , These are the weighting coefficients corresponding to the exponentially weighted moving average term and the first-order difference term, respectively.

[0039] The processing node on the ingress side of the call chain will predict the congestion probability and compare it with a preset first threshold. and the preset second threshold ( Less than The following comparison is performed: If the predicted congestion probability is less than the first threshold, the current load of the deep link is determined to be normal, and new arrival requests are allowed to pass normally. If the predicted congestion probability is greater than or equal to the first threshold and less than the second threshold, the drop probability is calculated based on the predicted congestion probability. The calculation method is as follows:

[0040] ;

[0041] in, This represents the probability of dropping newly arriving requests. Based on this probability, probabilistic dropping or downgrading is performed on new arrival requests. If the predicted congestion probability is greater than or equal to the second threshold, it is determined that cascading congestion is about to occur in the deep link. The new arrival request is then directly rejected at the source and a downgraded response is triggered to prevent the request from continuing to propagate into the deep link. In a specific example, a first threshold can be set. Second threshold It should be noted that the above-mentioned back pressure perception and prediction method can achieve forward perception of deep link congestion trends without relying on any centralized topology management components, relying only on the historical load information carried in the packet header. Agents at all levels do not need to know the global service call topology relationship to collaboratively complete the source suppression of cascading avalanches.

[0042] To facilitate understanding of the overall working process of this invention, a specific application scenario is used for illustration. Suppose a computing node on a cloud computing platform runs multiple service container groups belonging to tenant A and tenant B. When a request message from a service container group of tenant A arrives at the network interface card (NIC) of this node, the protocol probe and decision engine module parses the message, identifies that it carries a Hypertext Transfer Protocol (HTTP) frame header identifier field, determines that the message hits the application layer complex feature, marks it as a user-mode offload type, and extracts the corresponding tenant identifier as tenant A. The sandbox proxy pool manager queries the tenant security context based on this tenant identifier. If it determines that the sandbox proxy instance corresponding to tenant A has not yet been created, it creates the corresponding sandbox proxy instance according to tenant A's resource isolation policy, allocates the corresponding control group resource quota and an independent user-mode and kernel-mode shared memory region to it, and then maps the memory page corresponding to the request message to the sandbox proxy instance through the filling and receiving circular queues. After the sandbox proxy instance completes application layer processing such as authentication and verification, it is then sent through the sending circular queue and completed... The circular queue is written back and forwarded to the target service in a zero-copy manner. Meanwhile, if the message sent by another business container group of tenant B is regular transport layer traffic that does not involve application layer semantic processing, the message is directly marked as kernel direct forwarding type by the protocol probe and decision engine module. After querying the local forwarding table, it is directly forwarded without going through any user-space proxy. The handling of burst traffic on the tenant A side will not occupy the sandbox proxy resources of tenant B, nor will it affect the forwarding efficiency of tenant B's regular traffic. During the process of the request being passed along the call chain to the deep service and receiving a response, each downstream proxy calculates and merges the load confidence of its node in turn, writes the link comprehensive load confidence into the response message header and passes it upstream level by level. The gateway on the call chain entry side calculates the exponentially weighted moving average and the first-order difference value based on the historical sequence to obtain the predicted congestion probability. Based on the interval in which the predicted congestion probability is located, it performs allow, probabilistic drop or source rejection processing on subsequent new arrival requests, thereby performing proactive regulation of upstream traffic before congestion occurs in the deep link.

[0043] In another embodiment of the present invention, for deployment environments with smart NIC hardware acceleration capabilities, the protocol sniffing and partial forwarding decision logic executed in the kernel mode of the protocol probe and decision engine module can be moved down to the programmable data plane of the smart NIC for execution. The NIC hardware can directly complete traffic classification and map the data frames that need to be offloaded to the user-space and kernel-space shared memory area of ​​the user-space sandbox proxy instance through a hardware queue, thereby further reducing the involvement of the host processor. This implementation can be used as an equivalent alternative to the aforementioned kernel-space-based processing method.

[0044] In another embodiment of the present invention, for the local prediction model, in addition to using the aforementioned exponentially weighted moving average combined with first-order difference, a graph neural network model trained based on the service dependency topology graph structure can also be used. The link comprehensive load confidence sequence reported by agents at all levels and the service call topology relationship are used as inputs to the graph neural network model, and the output is the predicted congestion probability. This implementation is suitable for scenarios with relatively abundant computing resources and high requirements for prediction accuracy, and can be used as an equivalent alternative to the aforementioned local prediction model.

[0045] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A microservice mesh system for providing cloud computing services, characterized in that, This includes a control plane module and a data plane module deployed on each compute node of the cloud-native cluster; The data plane module includes a protocol probe and decision engine module, a sandbox proxy pool manager, a zero-copy transfer module, and a back pressure sensing and prediction module. The protocol probe and decision engine module is used to parse data packets arriving at the node's network card through an extended Berkeley packet filter hook mounted on the network card driver layer, obtain parsing results, and divide the data packets into a first type of data packet and a second type of data packet based on the parsing results. The first type of data packet is a data packet that does not involve complex application layer features, and the second type of data packet is a data packet that hits complex application layer features. For the first type of data packet, the local forwarding table is queried to obtain the forwarding path and forwarding is completed. For the second type of data packet, the tenant identifier corresponding to the data packet is extracted. The sandbox proxy pool manager is used to query the tenant security context based on the tenant identifier, obtain the liveness status of the sandbox proxy instance corresponding to the tenant identifier, and map the second type of data packet to the sandbox proxy instance when the sandbox proxy instance is live. When the sandbox proxy instance is not live, a new sandbox proxy instance is created based on the resource isolation policy corresponding to the tenant identifier, and a control group resource quota and an independent user-mode and kernel-mode shared memory area are allocated to the newly created sandbox proxy instance. The second type of data packet is then mapped to the newly created sandbox proxy instance. The sandbox proxy instance is used to perform application-layer semantic processing on the data packets mapped to it in user mode. The zero-copy delivery module is used to establish a zero-copy delivery path for data frames between the node network card and the sandbox proxy instance based on filling the ring queue, receiving the ring queue, sending the ring queue, and completing the ring queue. The back pressure sensing and prediction module is used to calculate the load confidence of the node based on the load index of the node, weight and fuse the load confidence of the node with the load confidence already carried in the message to obtain the link comprehensive load confidence, write the link comprehensive load confidence into the corresponding field of the message header and make the message carrying the link comprehensive load confidence be passed upstream along the call chain. The control plane module is used to generate protocol identification rules, maintain tenant security context, and distribute backpressure prediction model parameters to the data plane module of each computing node.

2. The system according to claim 1, characterized in that, When the protocol probe and decision engine module parses the data packets, it extracts the five-tuple information and application layer feature bytes. The five-tuple information includes the source IP address, source port, destination IP address, destination port, and protocol type. The application layer feature bytes include the transport layer security protocol handshake identifier field and the hypertext transfer protocol frame header identifier field. Based on a preset protocol feature rule table, the protocol probe and decision engine module determines data packets that do not involve circuit breaking, multi-tenant authentication, and retry processing requirements as the first type of data packets, and determines data packets that match the complex application layer features as the second type of data packets.

3. The system according to claim 1, characterized in that, The zero-copy delivery module is specifically used to: map the memory page corresponding to the second type of data packet to the user-space and kernel-space shared memory area of ​​the corresponding sandbox proxy instance through the filling ring queue and the receiving ring queue; after the sandbox proxy instance completes the application layer semantic processing, write the processing result back in a zero-copy manner through the sending ring queue and the completing ring queue, and the subsequent forwarding is completed by the protocol probe and decision engine module.

4. The system according to claim 1, characterized in that, Sandbox agent instances corresponding to different tenant identifiers are allocated independent user-mode and kernel-mode shared memory regions, and the resource usage of each sandbox agent instance is constrained by the resource quota of its respective control group.

5. The system according to claim 1, characterized in that, The load metrics of this node include the current backlog of requests, the current processor utilization, and the average response latency within the current statistical window. The backpressure perception and prediction module calculates the load confidence of this node by multiplying the ratio of the current backlog of requests to the preset queue capacity limit, the ratio of the current processor utilization and the ratio of the average response latency within the current statistical window to the preset service level agreed latency threshold by the corresponding weight coefficients, and then summing them to obtain the load confidence of this node. The sum of the three weight coefficients is 1.

6. The system according to claim 5, characterized in that, The back pressure sensing and prediction module is also used to perform weighted fusion of the load confidence of this node with the deep load confidence carried in the downstream packets to obtain the comprehensive load confidence of the link. The value range of the local weight coefficient is greater than 0 and less than 1. The processing node on the link ingress side is called to calculate the exponentially weighted moving average and the first difference value of the comprehensive load confidence sequence of the link arriving within the historical time window. The exponentially weighted moving average and the first difference value are input into the local prediction model to calculate the predicted congestion probability.

7. The system according to claim 6, characterized in that, The back pressure sensing and prediction module compares the predicted congestion probability with a preset first threshold and a preset second threshold, wherein the first threshold is less than the second threshold. When the predicted congestion probability is less than the first threshold, the new arrival request is allowed to proceed. When the predicted congestion probability is greater than or equal to the first threshold and less than the second threshold, the drop probability is calculated based on the predicted congestion probability, and the new arrival request is probabilistically dropped based on the drop probability. When the predicted congestion probability is greater than or equal to the second threshold, the new arrival request is rejected at the source and a degradation response is triggered.

8. A microservice mesh data processing method for providing cloud computing services, applied to the system described in claim 1, characterized in that, include: When a data packet arrives at the node's network interface card (NIC), the packet is parsed using the extended Berkeley packet filter hook mounted on the NIC driver layer to obtain the parsing result. Based on the analysis results, data packets that do not involve complex application layer features are marked as first-class data packets, and data packets that match complex application layer features are marked as second-class data packets. For the first type of data packet, the local forwarding table is queried to obtain the forwarding path, and the data packet is redirected to the corresponding network interface to complete the forwarding; For the second type of data packet, extract the tenant identifier corresponding to the data packet, query the tenant security context based on the tenant identifier, and obtain the liveness status of the sandbox proxy instance corresponding to the tenant identifier; When the sandbox proxy instance is alive, the memory page corresponding to the data packet is mapped to the user-mode and kernel-mode shared memory region of the sandbox proxy instance by filling the circular queue and receiving the circular queue. When the sandbox proxy instance is not alive, a new sandbox proxy instance is created according to the resource isolation policy corresponding to the tenant identifier. The newly created sandbox proxy instance is allocated a control group resource quota and an independent user-mode and kernel-mode shared memory region. The memory page corresponding to the data packet is mapped to the newly created sandbox proxy instance by filling the circular queue and receiving the circular queue. The sandbox proxy instance that obtains the data packet from the mapping completes application layer semantic processing in user space. The application layer semantic processing includes circuit breaker determination, multi-tenant authentication verification, and retry policy execution. The processing results are written back in a zero-copy manner by sending to and completing the circular queue, thus completing subsequent forwarding.

9. The method according to claim 8, characterized in that, When parsing the data packet, the five-tuple information and application layer feature bytes of the data packet are extracted. The five-tuple information includes the source IP address, source port, destination IP address, destination port and protocol type. The extracted five-tuple information and application layer feature bytes are matched and judged according to the preset protocol feature rule table to obtain the parsing result.

10. A backpressure control method for suppressing cascading congestion in a microservice mesh, characterized in that, include: The data plane proxy of the downstream service node collects the current load metrics of the node, including the current number of requests in the queue, the current processor utilization, and the average response latency within the current statistical window. Calculate the load confidence level of this node based on the aforementioned load metrics; The load confidence of this node is weighted and fused with the deep load confidence already carried in the downstream packets to obtain the overall load confidence of the link. Write the link comprehensive load confidence score into the corresponding field of the message custom header, so that the message carrying the link comprehensive load confidence score is passed upstream along the call chain. Each intermediate agent through which the message passes repeatedly performs the processing of collecting load indicators, calculating load confidence scores and weighted fusion. The processing node on the chain entry side continuously collects the comprehensive load confidence of the links arriving within the historical time window, and calculates the exponentially weighted moving average and first-order difference value for the collected sequences; The exponentially weighted moving average and the first-order difference are input into the local prediction model to calculate the predicted congestion probability. The predicted congestion probability is compared with a preset first threshold and a preset second threshold, wherein the first threshold is less than the second threshold. When the predicted congestion probability is less than the first threshold, the new arrival request is allowed to proceed. When the predicted congestion probability is greater than or equal to the first threshold and less than the second threshold, the drop probability is calculated based on the predicted congestion probability. The new arrival request is probabilistically dropped based on the drop probability. When the predicted congestion probability is greater than or equal to the second threshold, the new arrival request is rejected at the source and a downgrade response is triggered.