Methods, systems, devices, and media for cross-datacenter congestion control based on congestion alert packet filtering

Abstract

This invention discloses a cross-data center congestion control method, system, device, and medium based on congestion alert packet filtering. The method includes: each data center's core layer switch is connected to a DCI switch; each data center's DCI switch is interconnected with DCI switches in other data centers via optical fiber to achieve cross-data center communication; the receiving DCI switch filters congestion alert packets intended to be transmitted back to the sending data center based on congestion information collected from the receiving data center. This invention solves the problem of abnormal speed reduction in DCQCN in cross-data center scenarios due to receiving severely delayed congestion alert packets, improving the performance of cross-data center traffic without affecting the performance of existing traffic within the data center.

Description

Technical Field

[0001] This invention relates to the field of computer network technology, and specifically to a cross-data center congestion control method, system, device, and medium based on congestion alert packet filtering. Background Technology

[0002] Current data center network congestion control algorithms include DCQCN (Datacenter Congestion Notification) and HPCC. DCQCN, proposed in 2015, is a congestion control algorithm that uses Congestion Notification Packets (CNPs) to determine congestion and adjust the data transmission rate. Due to its low deployment difficulty and simple mechanism, DCQCN is a standard congestion control algorithm supported by many RDMA network cards.

[0003] The problem with DCQCN is that the algorithm was originally designed for congestion mitigation within a single data center. In cross-data center scenarios, the feedback latency is much greater than that within a data center (1ms vs. 10us), causing the algorithm to be unable to respond to congestion in a timely manner, resulting in performance degradation.

[0004] Specifically, when the switch port buffer reaches the configured gate capacity, DCQCN uses the RED algorithm to randomly mark passing data packets with ECN (Explicit Congestion Notification) with a certain probability. Upon receiving a data packet marked with ECN, the receiving end sends a CNP to the sending end. The sending end runs a rate control algorithm, slowing down upon receiving a CNP and accelerating if no CNP is received for a continuous period or during the continuous transmission of a certain amount of data.

[0005] DCQCN has been widely deployed in data center networks of various sizes due to its simple execution logic and excellent performance at the endpoint and intranet switch sides. However, in cross-data center scenarios, simply deploying the DCQCN algorithm has not yielded satisfactory performance.

[0006] As a sender-driven congestion control algorithm that relies on congestion signals for feedback, DCQCN achieves a time delay (RTT) from congestion occurrence to feedback effectiveness. However, in cross-datacenter scenarios, the RTT of cross-domain traffic can reach hundreds of microseconds to about one millisecond, far exceeding the RTT of several to tens of microseconds for intra-datacenter traffic. This causes a series of sender-driven congestion control algorithms, including DCQCN, to fail to provide timely feedback on congestion, leading to a performance degradation in cross-domain flows. One major manifestation of this is that the rate of cross-domain flows converges to a lower value due to the impact of bursty traffic within the data center.

[0007] The following analysis covers two common scenarios.Figure 1 It refers to cross-domain data flow transmission from data center A to data center B: A burst of traffic at the transmitter (DC A) causes congestion, but the congestion signal needs to travel 1 times the RTT (Relay Time To Time) to be transmitted back to the transmitter.

[0008] A burst of traffic at the receiving end (DC B) causes congestion. Assuming the one-way delay of the long-distance link is D, the congestion signal must pass through at least D to be transmitted back to the sending end. The traffic that has been slowed down at the sending end must pass through D again to reach the congestion point.

[0009] These severely delayed congestion signals can sometimes instruct the sender to perform incorrect rate reduction, especially for small bursts of traffic, where the congestion is often resolved by the time the congestion signal is received by the sender.

[0010] like Figure 2 The data structure describes the throughput of a long, continuous flow across two remote data centers, followed by short bursts of traffic within those data centers. The CNP generated by the cross-data center traffic is only transmitted back to the sender after the two short bursts have ended. By this time, the congestion has already subsided, but the DCQCN throughput drops to zero and only slowly recovers after a long period, severely impacting the performance of cross-domain traffic. Summary of the Invention

[0011] This invention addresses the shortcomings of existing technologies by providing a cross-data center congestion control method, system, device, and medium based on congestion alert packet filtering. It solves the problem of abnormal speed reduction in DCQCN in cross-data center scenarios due to receiving severely delayed congestion alert packets, improves the performance of cross-data center traffic, and does not affect the performance of the original traffic within the data center.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: A cross-data center congestion control method based on congestion alert packet filtering is proposed. The core layer switch of each data center is connected to a DCI switch. The DCI switch of each data center is interconnected with the DCI switches of other data centers through optical fiber to realize cross-data center communication. The receiving DCI switch filters the congestion alert packets to be transmitted back to the sending data center based on the congestion information collected by the receiving data center.

[0013] To optimize the above technical solution, the specific measures also include: Furthermore, the receiving end's DCI switch filters the congestion warning packets to be transmitted back to the sending end's data center based on the congestion information it has collected from the receiving end's data center, specifically as follows: For each transmission path from the receiving DCI switch to the receiving core layer switch, after the receiving DCI switch receives the congestion alert packet for the first time on the transmission path, it starts to count the proportion r of ECN echo in the ACKs received in each RTT. At the same time, when updating the proportion r during each RTT, it records the number of times the proportion r continuously decreases, and intercepts all congestion alert packets during the RTT within n DCs. If the receiving DCI switch still receives a congestion alert packet after n RTTs, the receiving DCI switch filters the congestion alert packet based on the ratio r and the number of times the ratio r continuously decreases, c.

[0014] Furthermore, the receiving DCI switch filters congestion warning packets based on the ratio r and the number of consecutive decreases of the ratio r c, specifically as follows: If the proportion r of ECN echo in the ACK of the previous RTT exceeds the ECN proportion threshold p, and the number of times the proportion r decreases continuously c is less than the number threshold t, then the congestion warning packet is forwarded to the sending data center normally. After receiving the congestion warning packet, the sending data center reduces the speed of the cross-data center flow; otherwise, the congestion warning packet is blocked. If the receiving DCI switch does not receive a congestion alert packet for k consecutive RTTs on the transmission path (k>n), clear all records and restore the state to that before the first congestion alert packet was received.

[0015] Furthermore, the formula for calculating the proportion r of ECN echo in the ACK is as follows:

[0016] In the formula, This indicates the number of ACK packets with the ECN echo flag received by the receiving DCI switch within one RTT. This represents the total number of ACK packets received by the receiving DCI switch.

[0017] Furthermore, the transmission path from the receiving DCI switch to the receiving core layer switch is distinguished by a tuple, which consists of the receiving IP address and the receiving DCI switch transmission port number portnum. Maintain a state table for all transmission paths, which records the congestion status of each transmission path, as well as the proportion r of ECN echo in the ACK and the number of times the proportion r continuously decreases, c. After receiving the ACK and congestion warning packet, obtain the receiver's IP address from the packet header, and recalculate the receiver's DCI switch transmission port number portnum based on the packet header information to determine the transmission path number, and modify the corresponding content in the state table.

[0018] This invention also proposes a cross-data center congestion control system based on congestion alert packet filtering. The core layer switch of each data center is connected to a DCI switch. The DCI switch of each data center is interconnected with the DCI switches of other data centers through optical fiber to realize cross-data center communication. The receiving DCI switch filters the congestion alert packets to be transmitted back to the sending data center based on the congestion information it collects from the receiving data center.

[0019] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the cross-data center congestion control method based on congestion alert packet filtering as described above.

[0020] The present invention also proposes a computer-readable storage medium storing a computer program that causes a computer to execute the cross-datacenter congestion control method based on congestion alert packet filtering as described above.

[0021] The beneficial effects of this invention are: DCQCN improves the throughput of cross-data center traffic and enhances the efficiency of cross-data center data transmission.

[0022] Taking into account the degree of remote congestion, congestion alert packets are filtered based on the collected congestion information, resulting in a shorter average flow completion time for cross-data center traffic compared to traditional DCQCN.

[0023] This invention is based on the existing DCQCN algorithm and does not require special functional support from the end-side RDMA network card. Attached Figure Description

[0024] Figure 1 This is a diagram illustrating congestion in cross-data center communication.

[0025] Figure 2 This is a throughput curve using the traditional DCQCN algorithm.

[0026] Figure 3 It is a topology diagram of cross-data center communication.

[0027] Figure 4 This is a flowchart of the cross-data center congestion control method based on congestion alert packet filtering proposed in this invention.

[0028] Figure 5 This is a throughput curve of the cross-data center congestion control method based on congestion alert packet filtering proposed in this invention.

[0029] Figure 6This is a comparison chart of the average flow completion time of the present invention and the DCQCN algorithm in cross-data center traffic. Detailed Implementation

[0030] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0031] Example 1 This invention proposes a cross-datacenter congestion control method based on congestion alert packet filtering. The topology of cross-datacenter communication is as follows: Figure 3 Each data center (assuming a spine-leaf architecture) has its core switch connected to a DCI (Data Center Interconnect Switch). Each data center's DCI switch is interconnected with DCI switches in other data centers via fiber optic cables, enabling cross-data center communication. For a cross-data center flow, the data center where the sending end is located is called the near-end data center, and the data center where the receiving end is located is called the far-end data center. The receiving end's DCI switch filters congestion warning packets intended for transmission back to the sending data center based on the congestion information it collects from the receiving data center. The process for filtering congestion warning packets is as follows: Figure 4 As shown, it includes the following steps: For each transmission path from the receiving DCI switch to the receiving core layer switch, after the receiving DCI switch receives the first Congestion Notification Packet (CNP) on the transmission path, the path is classified as slightly congested. The switch then begins to calculate the proportion *r* of ECN echoes in the ACKs received within each RTT (round-trip time). Simultaneously, when updating the proportion *r* during each RTT, the switch records the number of consecutive decreases *c* of the proportion *r*. All Congestion Notification Packets are intercepted within the RTTs of *n* core layers. The formula for calculating the proportion *r* of ECN echoes in the ACKs is as follows:

[0032] In the formula, This indicates the number of ACK packets with the ECN echo flag received by the receiving DCI switch within one RTT. This represents the total number of ACK packets received by the receiving DCI switch.

[0033] If the receiving DCI switch still receives congestion warning packets after n RTTs, the path is considered severely congested. The receiving DCI switch filters the congestion warning packets based on the ratio r and the number of consecutive decreases of the ratio r, specifically: If the proportion r of ECN echo in the ACK of the previous RTT exceeds the ECN proportion threshold p, and the number of times the proportion r decreases continuously c is less than the number threshold t, then it is determined that the congestion has not been relieved, and the congestion warning packet is forwarded to the sending data center normally. After receiving the congestion warning packet, the sending data center reduces the speed of the cross-data center flow. Otherwise, it is determined that the congestion has been temporarily relieved, and the congestion warning packet is intercepted to avoid unnecessary speed reduction.

[0034] If the receiving DCI switch does not receive a congestion alert packet for k consecutive RTTs on the transmission path (k>n), clear all records and restore the state to that before the first congestion alert packet was received.

[0035] Since this method uses the transmission path as the control granularity, path differentiation is necessary. The transmission path from the receiving DCI switch to the receiving core layer switch is differentiated using a tuple, which consists of the receiving IP address and the receiving DCI switch's transmission port number (portnum). A state table is maintained for all transmission paths, recording the congestion state of each path, the proportion *r* of the ECN echo in the ACK, and the number of consecutive decreases in proportion *r*, *c*. Upon receiving the ACK and congestion alert packets, the receiver's IP address is obtained from the packet header, and the receiver's DCI switch transmission port number *portnum* is recalculated based on the header information to determine the transmission path number, and the corresponding content in the state table is modified. Simultaneously, one or two timers are maintained for each path in non-initial states (all states except before the first CNP reception) for timing logic within the algorithm.

[0036] Example 2 This invention proposes a cross-data center congestion control system based on congestion alert packet filtering, corresponding to the method in Embodiment 1. The core layer switch of each data center is connected to a DCI switch, and the DCI switch of each data center is interconnected with the DCI switches of other data centers through optical fiber to realize cross-data center communication. The receiving DCI switch filters the congestion alert packets to be transmitted back to the sending data center based on the congestion information it collects from the receiving data center.

[0037] The implementation methods of each module and its function in the system are completely consistent with the steps of the method in Implementation Example 1, so they will not be repeated here.

[0038] Example 3 This invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the cross-data center congestion control method based on congestion alert packet filtering as described in Embodiment 1.

[0039] Example 4 This invention proposes a computer-readable storage medium storing a computer program that causes a computer to execute a cross-data center congestion control method based on congestion alert packet filtering as described in Embodiment 1.

[0040] This application tested the impact of the conventional DCQCN algorithm and the present invention on cross-data center traffic throughput on the NS3 network emulator.

[0041] The traditional DCQCN process is as follows: 1. When congestion occurs in the remote data center, the remote DCI switch receives the CNP from the receiving end and forwards it to the near end.

[0042] 2. After passing through the long-distance link and the near-end data center, the CNP is received by the sender. The sender reduces the traffic speed according to the DCQCN logic. At this point, at least one long-distance link delay D has elapsed since the congestion occurred, and the reduced traffic takes at least another one long-distance link delay D to reach the congestion point. Assuming the long-distance link is 100km, the reduced traffic will reach the congestion point 1ms after the congestion occurs, but the congestion may have already been alleviated by the reduced speed of other traffic in the remote data center, which has a response speed of tens of microseconds. Thus, this cross-data center flow loses a significant portion of throughput, i.e., the sender's abnormal speed reduction is caused by the CNP generated at the remote end not being able to be fed back in time.

[0043] The throughput curve of the cross-datacenter congestion control method based on congestion alert packet filtering proposed in this invention is as follows: Figure 5 As shown.

[0044] Cross-data center traffic only experienced a brief drop in throughput due to switch congestion and queuing caused by short bursts of traffic within the remote data center, and then quickly recovered. It did not cause the abnormal speed reduction caused by CNP transmission back to the sending end as in the DCQCN algorithm.

[0045] Furthermore, this application sets high traffic loads for both intra-datacenter and inter-datacenter traffic on NS3, and compares the average flow completion time (slowdown is the normalized flow completion time) of the present invention and the DCQCN algorithm for inter-datacenter traffic. The comparison results are as follows: Figure 6 As shown.

[0046] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, and portable compact disc read-only memory (CD). ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0047] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0048] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A cross-data center congestion control method based on congestion alert packet filtering, wherein the core layer switch of each data center is connected to a DCI switch, and the DCI switch of each data center is interconnected with the DCI switches of other data centers via optical fiber to achieve cross-data center communication, characterized in that: The receiving DCI switch filters congestion alert packets intended to be transmitted back to the sending data center based on the congestion information it collects from the receiving data center.

2. The cross-datacenter congestion control method based on congestion alert packet filtering as described in claim 1, characterized in that, The receiving end's DCI switch filters the congestion warning packets to be transmitted back to the sending end's data center based on the congestion information it collects from the receiving end's data center. Specifically, the filtering process is as follows: For each transmission path from the receiving DCI switch to the receiving core layer switch, after the receiving DCI switch receives the congestion alert packet for the first time on the transmission path, it starts to count the proportion r of ECN echo in the ACKs received in each RTT. At the same time, when updating the proportion r during each RTT, it records the number of times the proportion r continuously decreases, and intercepts all congestion alert packets during the RTT within n DCs. If the receiving DCI switch still receives a congestion alert packet after n RTTs, the receiving DCI switch filters the congestion alert packet based on the ratio r and the number of times the ratio r continuously decreases, c.

3. The cross-datacenter congestion control method based on congestion alert packet filtering as described in claim 2, characterized in that, The receiving DCI switch filters congestion warning packets based on a ratio r and the number of consecutive decreases of the ratio r, specifically as follows: If the proportion r of ECN echo in the ACK of the previous RTT exceeds the ECN proportion threshold p, and the number of times the proportion r decreases continuously c is less than the number threshold t, then the congestion warning packet is forwarded to the sending data center normally. After receiving the congestion warning packet, the sending data center reduces the speed of the cross-data center flow; otherwise, the congestion warning packet is blocked. If the receiving DCI switch does not receive a congestion alert packet for k consecutive RTTs on the transmission path (k>n), clear all records and restore the state to that before the first congestion alert packet was received.

4. The cross-datacenter congestion control method based on congestion alert packet filtering as described in claim 2, characterized in that, The formula for calculating the proportion r of ECN echo in the ACK is as follows: In the formula, This indicates the number of ACK packets with the ECN echo flag received by the receiving DCI switch within one RTT. This represents the total number of ACK packets received by the receiving DCI switch.

5. The cross-datacenter congestion control method based on congestion alert packet filtering as described in claim 2, characterized in that, The transmission path from the receiving DCI switch to the receiving core layer switch is distinguished by a tuple, which consists of the receiving IP address and the receiving DCI switch's transmission port number (portnum). Maintain a state table for all transmission paths, which records the congestion status of each transmission path, as well as the proportion r of ECNecho in the ACK and the number of times the proportion r continuously decreases, c. After receiving the ACK and congestion warning packet, obtain the receiver's IP address from the packet header, and recalculate the receiver's DCI switch transmission port number portnum based on the packet header information to determine the transmission path number, and modify the corresponding content in the state table.

6. A cross-data center congestion control system based on congestion alert packet filtering, wherein the core layer switch of each data center is connected to a DCI switch, and the DCI switch of each data center is interconnected with the DCI switches of other data centers via optical fiber to achieve cross-data center communication, characterized in that: The receiving DCI switch filters congestion alert packets intended to be transmitted back to the sending data center based on the congestion information it collects from the receiving data center.

7. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the cross-datacenter congestion control method based on congestion alert packet filtering as described in any one of claims 1-5.

8. A computer-readable storage medium storing a computer program, characterized in that, The computer program causes the computer to execute the cross-datacenter congestion control method based on congestion alert packet filtering as described in any one of claims 1-5.

Images