Cross-data center RDMA network congestion control method and device based on in-band telemetering

Abstract

The invention discloses a cross-data center RDMA network congestion control method and device based on in-band telemetering, and belongs to the technical field of network transmission control. The method comprises the following steps: sending a first cross-data center flow to an upstream DCI switch at an initial rate; according to a Switch-INT message returned by an upstream DCI switch, a first fair rate is calculated, a second cross-data center flow is sent to the upstream DCI switch at the first fair rate, an ACK message of a downstream data center responding to the first cross-data center flow is obtained, and the ACK message comprises a suggested rate; and generating a final rate in combination with the first fair rate and the suggested rate, and sending a third cross-data center flow to an upstream DCI switch at the final rate. According to the method, the network sensing capability of the cross-data center flow can be improved, the fairness of the flow in the data center and the cross-data center flow is improved, and the transmission delay of the datagram is optimized.

Description

Technical Field

[0001] This invention relates to the field of network transmission control technology, and in particular to a method and apparatus for cross-data center RDMA network congestion control based on in-band telemetry. Background Technology

[0002] In recent years, with the continuous growth in demand for data center networks, more and more service providers and cloud network providers are choosing to deploy their businesses across multiple data centers. These data centers are located in different regions and even countries, bringing many advantages to cross-data center applications: users in different regions can enjoy a faster access experience, distributed deployment helps protect user data privacy, and cross-data center networks enhance data disaster recovery capabilities and overall network reliability.

[0003] Remote Direct Memory Access (RDMA), as an emerging network technology, is gradually becoming a standard in data center networks due to its high throughput, low latency, and ultra-low CPU load. However, to effectively utilize RDMA technology across data center networks, the underlying network must support lossless transmission. Currently, Priority Flow Control (PFC) is commonly used to ensure lossless transmission at switch ports, but its use can lead to head-up congestion and PFC pause frame storms. Therefore, when deploying an RDMA network, a suitable congestion control algorithm is essential to avoid triggering PFC.

[0004] The design of congestion control protocols has always been a key focus of data center network research. Early TCP protocol improvements were based on TCP principles, adjusting the transmission rate through packet loss signals within the network. However, by the time the sender detects packet loss, severe congestion has often already occurred within the network. Therefore, in addition to using packet loss signals, data centers can utilize other signals to indicate congestion, such as the switch's ECN (Congestion Notification Signal), end-to-end RTT (Round-Trip Time) delay, and INT (In-Band Telemetry) technology. These signals allow the sender to adjust its rate more quickly, thereby reducing packet queuing. Based on these signals, several efficient congestion control protocols have emerged, such as DCQCN, TIMELY, and HPCC. Furthermore, there are also some switch-based congestion control algorithms, such as BFC and RoCC.

[0005] End-to-end congestion control protocols operate within a control loop from sender to receiver and back to sender, forming the entire RTT (Round-Trip Time). For cross-data center networks, the feedback loop is longer, and the RTT increases accordingly, requiring the sender to wait for a longer RTT before responding to congestion. During this time, the sender continues to send data packets, leading to queue backlog. Furthermore, in scenarios where long-distance traffic (cross-data center traffic) and short-distance traffic (intra-data center traffic) compete, different traffic types have different RTTs, causing inconsistencies in congestion response times and resulting in unfair bandwidth allocation. For example, when applications communicate between data centers, existing RDMA technology faces challenges in achieving high throughput over long-distance links. Typically, intra-data center link latency is 10 to 50 microseconds, while inter-data center latency is 1 to 3 milliseconds. However, current congestion control algorithms primarily consider intra-data center transmission demands, causing the available bandwidth of intra-data center flows to be squeezed when long-RTT cross-data center flows exist, leading to unfair bandwidth allocation.

[0006] Existing switch-based congestion control protocols typically require deployment on most switches in a data center network. For example, RoCC and BFC algorithms need to be deployed on all data center switches, which significantly increases the difficulty of deployment and hinders incremental deployment in existing data centers. Furthermore, a single switch has limited ability to assess the overall network congestion level, and its congestion control effectiveness for congestion originating from different senders is also relatively limited.

[0007] To address the issue of unfair data flow across data centers, the industry has adopted various technical measures to mitigate the impact, but these measures still have limitations. For example, Annual uses ECN to provide timely feedback on congestion within the data center, thereby controlling the flow rate of the WAN within the data center. However, WANs have many uncontrollable factors, such as long-distance and cross-regional communication links, which may reduce the optimization effect of dual-loop systems. Traffic fluctuations and latency jitter in WANs can still lead to performance deviations from expectations.

[0008] However, the aforementioned technical methods do not solve the fairness issue of cross-data center flows. The coarse-grained information feedback makes it impossible for terminals to accurately perceive the network status of data centers. In particular, when the upstream data center where the sender is located is not congested, but the downstream data center where the receiver is located is congested, how to adjust the rate of cross-data center flows to adapt to the available bandwidth of the two data centers?

[0009] In-band Network Telemetry (INT) is a network monitoring and troubleshooting technique that allows network devices to carry telemetry information about the network status during normal packet forwarding. This telemetry information is transmitted along with the data packets, eliminating the need for additional probe packets or out-of-band communication, hence the name "in-band telemetry." In traditional network monitoring, network status information is typically collected through separate probe packets or external tools, while INT directly embeds network status information into the header (or trailer) of user data packets. This approach allows network operators to understand in real time the path, latency, packet loss rate, queue information, and other data traversed by each data packet within the network. Summary of the Invention

[0010] This invention proposes a congestion control method and apparatus for cross-data center RDMA networks based on in-band telemetry, in order to eliminate or improve one or more defects existing in the prior art. It solves the problems of excessively long feedback loops in end-to-end congestion control protocols in cross-data center network scenarios, which easily leads to queue accumulation, and unfair competition between cross-data center flows and intra-data center flows; and the problems of switch-based congestion control protocols requiring large-scale modifications to switches and having mediocre control effects.

[0011] To achieve the above objectives, the technical solution of the present invention includes the following:

[0012] A congestion control method for cross-datacenter RDMA networks based on in-band telemetry, applied to an upstream data center, the method comprising:

[0013] Send the first cross-datacenter flow to the upstream DCI switch at the initial rate;

[0014] Based on the Switch-INT message returned by the upstream DCI switch, calculate the first fair rate and send the second cross-data center flow to the upstream DCI switch at the first fair rate; wherein, the Switch-INT message is generated based on the INT information of the switches along the route from the upstream data center to the upstream DCI switch carried in the first cross-data center flow.

[0015] Obtain the ACK packet from the downstream data center in response to the first cross-data center flow, the ACK packet containing the suggested rate;

[0016] The first fair rate and the proposed rate are combined to generate a final rate, and a third cross-datacenter flow is sent to the upstream DCI switch at the final rate.

[0017] Furthermore, the process of generating the ACK message includes:

[0018] The upstream DCI switch resets the header information of the first cross-data center flow based on the INT information of the switches along the way from the upstream data center to the upstream DCI switch, and sends the first cross-data center flow with the reset header information to the downstream DCI switch at the initial rate.

[0019] The downstream DCI switch will reset the header information of the first cross-data center flow and forward it to the downstream data center at the initial rate;

[0020] The downstream DCI switch will insert the INT information it possesses into the first cross-data center stream of the reset header information at the switches along the way to the downstream data center.

[0021] The downstream data center generates an ACK packet based on the INT information of the switches encountered along the way from the downstream DCI switch to the downstream data center.

[0022] Furthermore, the ACK message also includes a second fair rate, which is generated by the downstream data center based on the INT information of the switches along the way from the downstream DCI switch to the downstream data center and by applying a preset congestion control algorithm. The preset congestion control algorithm includes either the HPCC algorithm or the PowerTCP algorithm.

[0023] Furthermore, after the upstream DCI switch receives the third cross-datacenter flow, it also includes:

[0024] The upstream DCI switch sends a third cross-data center flow to the downstream DCI switch at the final rate.

[0025] After receiving the third cross-data center flow, the downstream DCI switch sends the third cross-data center flow to the downstream data center at the second fair rate.

[0026] Furthermore, when the downstream DCI switch sends the third cross-data center flow to the downstream data center at the second fair rate, it places data packets with the same destination IP server into the same virtual output queue.

[0027] Furthermore, the process of generating the proposed rate includes:

[0028] The downstream data center obtains the predicted rate based on the suggested rate in the previous ACK packet; wherein the initial value of the predicted rate is the second fair rate;

[0029] The predicted queue transmission delay is evaluated based on the predicted rate;

[0030] The predicted queue transmission delay is compared with the set queue delay threshold;

[0031] When the predicted queue transmission delay exceeds the set queuing delay threshold, a suggested rate is calculated in conjunction with the specified time length.

[0032] Furthermore, the method also includes:

[0033] The upstream data center generates a data packet limit, and when the amount of data to be sent exceeds the data packet limit, it sends the third cross-data center flow to the upstream DCI switch based on the data packet limit; wherein, the data packet limit is obtained based on the bandwidth and latency of the long-distance link between the upstream switch and the downstream switch.

[0034] A congestion control device for cross-data center RDMA networks based on in-band telemetry, the device comprising:

[0035] The acquisition module is used to acquire the Switch-INT message returned by the upstream DCI switch and the ACK message of the downstream data center in response to the first cross-data center flow; wherein, the Switch-INT message is generated based on the INT information of the switches along the way from the upstream data center to the upstream DCI switch carried in the first cross-data center flow, and the ACK message contains the suggested rate and the second fair rate.

[0036] The calculation module is used to calculate a first fair rate based on the Switch-INT message, and to generate a final rate by combining the first fair rate and the proposed rate;

[0037] The sending module is configured to send a first cross-data center flow to the upstream DCI switch at an initial rate, send a second cross-data center flow to the upstream DCI switch at the first fair rate, and send a third cross-data center flow to the upstream DCI switch at the final rate.

[0038] An electronic device includes: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the cross-data center RDMA network congestion control method based on in-band telemetry as described above.

[0039] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the cross-datacenter RDMA network congestion control method based on in-band telemetry as described above.

[0040] A cross-datacenter RDMA network congestion control system based on in-band telemetry, the system comprising:

[0041] An upstream data center is used to send a first cross-data center flow to an upstream DCI switch at an initial rate; calculate a first fair rate based on the Switch-INT message returned by the upstream DCI switch, and send a second cross-data center flow to the upstream DCI switch at the first fair rate; wherein, the Switch-INT message is generated based on the INT information of the switches along the path from the upstream data center to the upstream DCI switch carried in the first cross-data center flow; obtain the ACK message of the downstream data center responding to the first cross-data center flow; generate a final rate by combining the first fair rate and the suggested rate, and send a third cross-data center flow to the upstream DCI switch at the final rate;

[0042] The upstream DCI switch is used to generate and return the Switch-INT message to the upstream data center; and to send the first cross-data center flow, the second cross-data center flow and the third cross-data center flow to the downstream DCI switch at the initial rate, the first fair rate and the final rate, respectively.

[0043] Downstream data center, used to generate ACK packets in response to the first cross-data center flow; wherein the ACK packets contain a suggested rate and a second fair rate;

[0044] Downstream DCI switches are used to send the first cross-data center flow to the downstream data center at an initial rate, and to send the second cross-data center flow and the cross-data center flow to the downstream data center at a second fair rate.

[0045] Compared with the prior art, the present invention has the following effects and advantages:

[0046] (1) The control loop for cross-data center transmission is subdivided from end to end into near-source feedback, receiver-driven and switch-assisted, thereby improving the network's perception of cross-data center flows;

[0047] (2) Improved fairness of intra-datacenter and cross-datacenter flows;

[0048] (3) Reduced the queuing time of cross-data center flows at the switch and optimized the transmission latency of data packets. Attached Figure Description

[0049] Figure 1 This is a flowchart of cross-data center data transfer.

[0050] Figure 2 It is a Switch-INT message format.

[0051] Figure 3 This is a schematic diagram of the DCI switch framework. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0053] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0054] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0055] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0056] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0057] It should be emphasized here that the step markers mentioned below are not a limitation on the order of the steps, but should be understood as meaning that the steps can be executed in the order mentioned in the embodiments, or in a different order than in the embodiments, or several steps can be executed simultaneously.

[0058] The cross-datacenter RDMA network congestion control method based on in-band telemetry of the present invention is executed on a cross-datacenter network, which includes at least an upstream datacenter, an upstream DCI switch, a downstream DCI switch, and a downstream datacenter connected sequentially. Typically, the workflow of this cross-datacenter network is as follows: the sender in the upstream datacenter organizes the data to be sent into data packets and sends them to the upstream DCI switch; the upstream DCI switch forwards the data packets to the downstream DCI switch; the downstream DCI switch sends the data packets to the downstream datacenter; and the receiver in the downstream datacenter receives and processes the data packets. However, due to the difference in fair rates between the upstream and downstream datacenters, when the upstream fair rate is greater than the downstream fair rate, some data will accumulate on the downstream DCI switch. When the downstream DCI switch has too much buffered data, it will affect the end-to-end latency of the data packets. Therefore, it is necessary to optimize the end-to-end latency of the cross-datacenter flow.

[0059] Figure 1The following is a flowchart of the cross-data center transmission process of the system of the present invention, which mainly includes the following steps.

[0060] Step 1: The upstream data center organizes the data to be sent into a first cross-data center flow and sends it to the upstream DCI switch at an initial rate. After receiving the first cross-data center flow, the upstream DCI switch reassembles the data according to the INT header information in the first cross-data center flow to form a Switch-INT message. The upstream DCI switch then returns the Switch-INT message to the upstream data center and sends the first cross-data center flow with the reset INT header information to the downstream DCI switch at the initial rate.

[0061] The upstream data center first organizes the generated data packets, generates the first cross-data center flow, and sends the first cross-data center flow to the upstream DCI switch at an initial rate.

[0062] This invention designs a "near-source feedback" mechanism for cross-data center flows. Data centers are typically connected via leased lines, which in turn connect to the internal networks of the data centers using DCI switches. When cross-data center traffic travels from the upstream data center to the upstream DCI switch, each switch inserts its relevant INT (Intent to Data) information into the data packet. When the upstream DCI switch receives the cross-data center data packet, it copies the INT header information and, based on... Figure 2 After being reassembled into a Switch-INT message, this message is sent to the sender in the upstream data center. The Switch-INT message contains INT information along the path. Simultaneously, the upstream DCI switch clears the INT information from the first cross-data center flow before sending it to the downstream DCI switch.

[0063] In a preferred embodiment, the upstream data center sets a preset limit on the number of data packets it generates when generating cross-data center flows. This preset limit is twice the latency-bandwidth product, which is calculated based on the bandwidth and latency of the long-distance link between the upstream and downstream switches.

[0064] Step 2: The downstream DCI switch sends the first cross-data center flow to the downstream data center at the initial rate; the downstream data center generates an ACK packet based on the INT header information in the first cross-data center flow at this time, and returns the ACK packet to the upstream data center via the downstream DCI switch and the upstream DCI switch.

[0065] This invention designs a "receiver-driven" approach for cross-data center flows. When a cross-data center flow reaches a downstream DCI switch and is then sent to the downstream data center, the switch within the downstream data center re-inserts INT information. The downstream data center receiver generates an ACK packet based on the INT information. This packet includes a suggested rate and the fair rate (second fair rate) for the flow within the downstream data center. After the downstream data center receiver returns this ACK packet to the downstream DCI switch, the downstream DCI switch, based on the second fair rate in the received ACK packet, sends the cross-data center flow back to the downstream data center. Simultaneously, the downstream DCI switch also sends the ACK packet to the upstream DCI switch, ensuring that the upstream data center receives the ACK packet.

[0066] In a preferred embodiment, downstream switches place packets destined for the same IP server into the same virtual output queue to separate traffic from different downstream data centers and avoid head congestion problems caused by different flows destined for the same IP server being in the same virtual output queue.

[0067] In a preferred embodiment, the second fair rate is obtained based on a preset congestion control algorithm, which includes, but is not limited to, the HPCC algorithm and the PowerTCP algorithm.

[0068] The proposed rate generation is based on the "buffer queue management function" of the cross-data center flow designed in this invention. Due to the difference in fair rates between upstream and downstream data centers in the cross-data center flow, when the upstream fair rate is greater than the downstream fair rate, some data will accumulate on the downstream DCI switch. When the downstream DCI switch has too much buffered data, it will affect the end-to-end latency of the data packets. Therefore, it is necessary to optimize the end-to-end latency of the cross-data center flow. This algorithm aims to optimize cross-data center network traffic management by dynamically adjusting the sending rate through a series of steps to reduce queuing latency. When the system reaches a steady state, the proposed rate fed back by the receiver through ACK is the sending rate of the sender. Therefore, when the system reaches a steady state, the downstream DCI switch can know the sending rate of the sender in advance. The DCI switch calculates the average receiving rate by predicting the future receiving rate and monitors the current sending rate in real time. Based on the predicted receiving rate, this algorithm evaluates the predicted queue transmission latency and compares it with the set queuing latency threshold. When the queue latency exceeds the threshold, the algorithm calculates the target rate to control the queue latency to the threshold within a specified time. Ultimately, by adjusting the transmission rate, smooth traffic management is achieved, significantly improving data transmission efficiency and adapting to constantly changing network conditions.

[0069] In a preferred embodiment, the process of generating the suggested rate includes: obtaining a predicted rate based on the suggested rate in the previous ACK packet; wherein the initial value of the predicted rate is a second fair rate; evaluating the predicted queue transmission delay based on the predicted rate; comparing the predicted queue transmission delay with a set queuing delay threshold; and calculating the current suggested rate in combination with a specified time length when the predicted queue transmission delay exceeds the set queuing delay threshold.

[0070] Step 3: The upstream data center calculates the first fair rate based on the Switch-INT message, and before receiving the ACK message, sends the second cross-data center flow to the upstream DCI switch at the first fair rate.

[0071] When the sending end receives a Switch-INT message, it applies a congestion control algorithm based on the INT information to calculate the fair rate (first fair rate) of the flow in the upstream data center, and sends the data according to the calculated first fair rate.

[0072] Prioritize that the upstream and downstream data centers use the same congestion control algorithm.

[0073] Step 4: After receiving the ACK message, the upstream data center will combine the first fair rate and the suggested rate to generate the final rate, and send the third cross-data center flow to the upstream DCI switch at the final rate.

[0074] As described above, this invention, based on practical experience, observes that network conditions undergo a process of stabilization from instability. During this process, the final rate will gradually converge to the fair rate within the downstream data center. In other words, before the network stabilizes, this invention dynamically adjusts the transmission rate of the upstream data center to prevent congestion; after the network stabilizes, the suggested rate issued by the downstream data center becomes the fair rate within the downstream data center, and the upstream data center will increasingly consider this suggested rate when initiating cross-data center flows.

[0075] This method is executed across a data center network, such as Figure 3 As shown, a cross-data center network includes at least an upstream data center, an upstream DCI switch, a downstream DCI switch, and a downstream data center connected in sequence. The upstream and downstream data centers, from top to bottom, consist of a core layer and an access layer connected in sequence. The core layer primarily provides high-capacity bandwidth. The access layer, for convenient server connection, is typically placed at the top of the rack and is also known as a ToR switch.

[0076] Upstream DCI switches interconnect with the core switches in the core layer of the upstream data center for data transmission and communication; downstream DCI switches interconnect with the core switches in the core layer of the downstream data center for data transmission and communication. The upstream and downstream DCI switches are connected via long-distance optical fibers and use wavelength division multiplexing (WDM) technology to improve data transmission speed.

[0077] Under normal circumstances, the upstream and downstream data centers use the same congestion control algorithm by default. In steps 6 and 7, to better adapt to the congestion control algorithms within the upstream and downstream data centers, the downstream data center uses the same preset congestion control algorithm as the upstream and downstream data centers when generating arrival notification messages. In some embodiments, the preset congestion control algorithm includes, but is not limited to, the HPCC algorithm and the PowerTCP algorithm.

[0078] In summary, this invention designs a "buffer queue management function" for cross-data center flows. Due to the difference in fair rates between upstream and downstream data centers, when the upstream fair rate is greater than the downstream fair rate, some data accumulates on the downstream DCI switch. When the downstream DCI switch has too much buffered data, it affects the end-to-end latency of data packets. Therefore, it is necessary to optimize the end-to-end latency of cross-data center flows. This invention aims to optimize cross-data center network traffic management by dynamically adjusting the sending rate through a series of steps to reduce queuing latency. When the system reaches a steady state, the suggested rate fed back by the receiver via ACK is the sending rate of the sender. Therefore, when the system reaches a steady state, the downstream DCI switch can know the sending rate of the sender in advance. The downstream DCI switch calculates the average receiving rate by predicting future receiving rates and monitors the current sending rate in real time. Based on the predicted receiving rate, this algorithm evaluates the predicted queue transmission latency and compares it with a set queuing latency threshold. When the queue latency exceeds this threshold, the algorithm calculates the target rate to control the queue latency to the threshold within a specified time. Ultimately, by adjusting the transmission rate, smooth traffic management is achieved, significantly improving data transmission efficiency and adapting to constantly changing network conditions.

[0079] Although the present invention has been disclosed above with reference to embodiments, it is not intended to limit the present invention. Appropriate modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention should be covered within the protection scope of the present invention, which is defined by the claims.

Claims

1. A congestion control method for cross-data center RDMA networks based on in-band telemetry, characterized in that, The method, applied to an upstream data center in a cross-data center RDMA network, includes: Send the first cross-datacenter flow to the upstream DCI switch at the initial rate; Based on the Switch-INT message returned by the upstream DCI switch, calculate the first fair rate and send the second cross-data center flow to the upstream DCI switch at the first fair rate; wherein, the Switch-INT message is generated based on the INT information of the switches along the route from the upstream data center to the upstream DCI switch carried in the first cross-data center flow. Obtain the ACK packet from the downstream data center in response to the first cross-data center flow, the ACK packet containing the suggested rate; The first fair rate and the proposed rate are combined to generate a final rate, and a third cross-datacenter flow is sent to the upstream DCI switch at the final rate.

2. The method according to claim 1, characterized in that, The process of generating the ACK message includes: The upstream DCI switch resets the header information of the first cross-data center flow based on the INT information of the switches along the way from the upstream data center to the upstream DCI switch, and sends the first cross-data center flow with the reset header information to the downstream DCI switch at the initial rate. The downstream DCI switch will reset the header information of the first cross-data center flow and forward it to the downstream data center at the initial rate; The downstream DCI switch will insert the INT information it possesses into the first cross-data center stream of the reset header information at the switches along the way to the downstream data center. The downstream data center generates an ACK packet based on the INT information of the switches encountered along the way from the downstream DCI switch to the downstream data center.

3. The method according to claim 2, characterized in that, The ACK message also includes a second fair rate, which is generated by the downstream data center based on the INT information of the switches along the way from the downstream DCI switch to the downstream data center and by applying a preset congestion control algorithm. The preset congestion control algorithm includes either the HPCC algorithm or the PowerTCP algorithm.

4. The method according to claim 3, characterized in that, After the upstream DCI switch receives the third cross-datacenter flow, it also includes: The upstream DCI switch sends a third cross-data center flow to the downstream DCI switch at the final rate. After receiving the third cross-data center flow, the downstream DCI switch sends the third cross-data center flow to the downstream data center at the second fair rate.

5. The method according to claim 4, characterized in that, When the downstream DCI switch sends the third cross-data center flow to the downstream data center at the second fair rate, it puts data packets with the same destination IP server into the same virtual output queue.

6. The method according to claim 3, characterized in that, The process of generating the proposed rate includes: The downstream data center obtains the predicted rate based on the suggested rate in the previous ACK packet; wherein the initial value of the predicted rate is the second fair rate; The predicted queue transmission delay is evaluated based on the predicted rate; The predicted queue transmission delay is compared with the set queue delay threshold; When the predicted queue transmission delay exceeds the set queuing delay threshold, the current recommended rate is calculated in conjunction with the specified time length.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes: The upstream data center generates a data packet limit, and when the amount of data to be sent exceeds the data packet limit, it sends the third cross-data center flow to the upstream DCI switch based on the data packet limit; wherein, the data packet limit is obtained based on the bandwidth and latency of the long-distance link between the upstream switch and the downstream switch.

8. A congestion control device for cross-data center RDMA networks based on in-band telemetry, characterized in that, The device includes: The acquisition module is used to acquire the Switch-INT message returned by the upstream DCI switch and the ACK message of the downstream data center in response to the first cross-data center flow; wherein, the Switch-INT message is generated based on the INT information of the switches along the way from the upstream data center to the upstream DCI switch carried in the first cross-data center flow, and the ACK message contains the suggested rate and the second fair rate. The calculation module is used to calculate a first fair rate based on the Switch-INT message, and to generate a final rate by combining the first fair rate and the proposed rate; The sending module is configured to send a first cross-datacenter flow to the upstream DCI switch at an initial rate, send a second cross-datacenter flow to the upstream DCI switch at the first fair rate, and send a third cross-datacenter flow to the upstream DCI switch at the final rate before the cross-datacenter RDMA network stabilizes.

9. An electronic device, characterized in that, The electronic device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the cross-data center RDMA network congestion control method based on in-band telemetry as described in any one of claims 1 to 7.

10. A congestion control system for cross-data center RDMA networks based on in-band telemetry, characterized in that, The system includes: An upstream data center is used to send a first cross-data center flow to an upstream DCI switch at an initial rate; calculate a first fair rate based on the Switch-INT message returned by the upstream DCI switch, and send a second cross-data center flow to the upstream DCI switch at the first fair rate; wherein, the Switch-INT message is generated based on the INT information of the switches along the path from the upstream data center to the upstream DCI switch carried in the first cross-data center flow; obtain the ACK message of the downstream data center responding to the first cross-data center flow; generate a final rate by combining the first fair rate and the suggested rate, and send a third cross-data center flow to the upstream DCI switch at the final rate; The upstream DCI switch is used to generate and return the Switch-INT message to the upstream data center; and to send the first cross-data center flow, the second cross-data center flow and the third cross-data center flow to the downstream DCI switch at the initial rate, the first fair rate and the final rate, respectively. Downstream data center, used to generate ACK packets in response to the first cross-data center flow; wherein the ACK packets contain a suggested rate and a second fair rate; Downstream DCI switches are used to send the first cross-data center flow to the downstream data center at an initial rate, and to send the second cross-data center flow and the cross-data center flow to the downstream data center at a second fair rate.

Images