Transmission control method based on flow and delay perception in data center network

Abstract

The invention provides a transmission control method based on flow and delay perception in a data center network, which comprises the following steps of: delay measurement: calculating round-trip delay and minimum queuing delay of a data packet without queuing delay to obtain a delay measurement result; transmission is paused, and the sender of the data packet adopts a corresponding strategy to execute transmission pause operation of the data flow according to the delay measurement result and different flow types. According to the method, a pause transmission technology is adopted, the data packets can be effectively prevented from being subjected to large queuing delay, meanwhile, short flow transmission is preferentially carried out, the overall performance of the network is remarkably improved, meanwhile, the method can be flexibly integrated to an existing timeout-free transmission control protocol, queuing backlog is rapidly eliminated, the link utilization rate is improved, and therefore the network performance is further improved.

Description

Technical Field

[0001] The present invention relates to the field of data transmission technology, and in particular to a transmission control method based on flow and delay perception in a data center network. Background Art

[0002] In modern data center networks, traffic microbursts and many-to-one communication patterns are very common. Figure 1 It shows that servers in the same switch and servers in different switches send data to the same receiver. However, when high-concurrency burst flows reach shallow-buffered switches, it is very easy to cause serious queue backlogs in the switch queue, resulting in large queue delays, and even a large number of data packets being discarded and timed out due to switch buffer overflow, reducing network performance. In general, existing solutions to the timeout problem can be divided into the following two categories: host-based solutions and switch-based solutions. These solutions all have certain limitations and cannot achieve good network performance. The details are as follows:

[0003] (1) Limitations of host-based solutions: In data center networks, researchers have designed a variety of transmission protocols such as T-RACKs and AP to eliminate timeouts and improve link utilization. However, these host-based protocols cannot effectively control queue length and obtain ultra-low latency, and cannot achieve the optimal flow completion time of data flows.

[0004] (2) Limitations of switch-based solutions: In order to avoid timeouts and achieve high link utilization, switch-based solutions (such as PLATO and CP) usually use packet labeling mechanisms or payload shearing techniques to keep the TCPACK clock alive or accurately distinguish packet loss. Therefore, the sender can continue to transmit data and avoid wasting bandwidth resources, thereby achieving better network performance. However, these switch-based solutions usually have problems with excessive queue length or large amounts of packet loss. The root cause is that these solutions can only be effective when the switch cache queue length reaches a large length (close to buffer overflow). Compared with host-based solutions, switch-based solutions have greater queue delays. Summary of the invention

[0005] The purpose of the present invention is to provide a transmission control method based on flow and delay perception in a data center network, which eliminates queue backlogs and reduces queue delays by combining flow type and network congestion level, while ensuring that short flows are transmitted first, thereby maximizing the transmission performance of the network.

[0006] To achieve the above-mentioned object of the invention, the first aspect of the present invention provides a transmission control method based on flow and delay perception in a data center network, the method comprising:

[0007] Delay measurement: calculate the round-trip delay and minimum queuing delay of the data packet without queuing delay to obtain the delay measurement result;

[0008] Pause transmission,The sender of the data packet adopts corresponding strategies to execute the suspension transmission operation of the data flow according to the delay measurement results and different traffic types.

[0009] Furthermore, the round-trip time of the data packet without queuing delay is calculated. Specifically, the sender of the data packet continuously measures the round-trip delay RTT experienced by each data packet in the transmission path, and the minimum value of all measured RTTs is taken as the round-trip delay without queuing delay.

[0010] Furthermore, the minimum queuing delay of the data packet is calculated, including:

[0011] S101. Calculate the queuing delay T experienced by each data packet q , the calculation formula is:

[0012] T q =RTT rt -RTT min

[0013] In the above formula, RTT rt Indicates the RTT value measured in real time, RTT min Indicates the minimum measured RTT value;

[0014] S102, using the weighted moving average method to tolerate noise and calculate the final packet queuing delay T q , the calculation formula is as follows:

[0015] T q =(1-α)×T q +α×T p

[0016] In the above formula, α represents the weight for calculating the queuing delay, which ranges from 0 to 1. p Indicates the queuing delay experienced by the previous data packet;

[0017] S103: The queuing delay T experienced by the previous data packet p Update to T q , and according to T q For the minimum queuing delay T min to update.

[0018] Furthermore, the sender of the data packet adopts a corresponding strategy to perform a data flow transmission suspension operation according to the delay measurement result, including:

[0019] Determine whether the minimum queuing delay is 0, and analyze the traffic type, wherein the traffic type includes long flows and short flows;

[0020] If the traffic type is long flow and the minimum queuing delay is greater than 0, the packet transmission operation is immediately suspended for a period equal to the measured minimum queuing delay;

[0021] If the traffic type is short flow and the minimum queuing delay is not 0, then 1 / n e The probability of deciding whether to perform the transmission suspension operation is n e Indicates the number of competing flows on the bottleneck link;

[0022] If the minimum queuing delay is greater than 0, the sending of new packets is suspended to eliminate the backlog of packets to be sent.

[0023] A second aspect of the present invention provides a transmission control system based on flow and delay perception in a data center network, the system comprising:

[0024] A delay measurement module is used to calculate the round-trip delay and minimum queuing delay of a data packet without queuing delay to obtain a delay measurement result;

[0025] The transmission suspension module is used for the sender of the data packet to execute the transmission suspension operation of the data flow according to the delay measurement result and different traffic types using corresponding strategies.

[0026] A third aspect of the present invention provides a computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] The present invention combines traffic type and network congestion level to design a traffic and delay-aware transmission control method and system in a data center network. The method adopts a transmission pause technology to effectively avoid data packets from experiencing large queuing delays, while giving priority to short-flow transmission, significantly improving the overall network performance. At the same time, the method can be flexibly integrated into existing transmission control protocols without timeouts, quickly eliminating queue backlogs and improving link utilization, thereby further improving network performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0030] Figure 1 This is a schematic diagram of the many-to-one communication mode that is widely present in the Leaf-spine network topology.

[0031] Figure 2 It is a schematic diagram of the overall flow of a transmission control method based on flow and delay perception in a data center network provided by an embodiment of the present invention.

[0032] Figure 3 It is a schematic diagram of the overall framework of a transmission control method based on flow and delay perception in a data center network provided by an embodiment of the present invention.

[0033] Figure 4 It is a schematic diagram of queue delay improvement percentage provided by an embodiment of the present invention.

[0034] Figure 5 It is a performance diagram under a web search workload provided by an embodiment of the present invention.

[0035] Figure 6 It is a performance diagram under Hadoop workload provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The enumerated embodiments are only used to explain the present invention and are not used to limit the scope of the present invention.

[0037] This embodiment provides a transmission control method based on traffic and delay perception in a data center network, the method comprising:

[0038] Delay measurement: calculate the round-trip delay and minimum queuing delay of the data packet without queuing delay to obtain the delay measurement result;

[0039] Pause transmission,The sender of the data packet adopts corresponding strategies to execute the suspension transmission operation of the data flow according to the delay measurement results and different traffic types.

[0040] In this embodiment, the sender of the data packet uses software timestamp to estimate the round trip delay and queuing delay. The main reasons are as follows:

[0041] First, although the round-trip delay can be measured more accurately through the timestamp of the network card hardware, this method inevitably increases the deployment cost. Secondly, software timestamps have good applicability. They have been widely used in public cloud virtual machines and bare metals, and are suitable for large-scale deployment in production data centers. Finally, software timestamps have high accuracy. The latest research results show that even under 80% network load, the median accuracy and 99th percentile accuracy are hundreds of nanoseconds and 2 to 3 microseconds respectively. In short, the sender can accurately obtain the round-trip delay and queuing delay through software timestamps.

[0042] On the other hand, due to kernel scheduling and offloads (such as GSO / TSO and GRO / LRO), the latency measured by the TCP stack can be highly variable and inaccurate, especially in high-speed data center networks. To address this issue, the send timestamp carried by the data packet is only written by the NIC driver of the terminal host at the moment it is about to be injected into the network, and the receive timestamp is also recorded when the ACK arrives at the sending network card. The main reason for this is that the network card driver is closest to the line, which can eliminate the additional latency overhead caused by kernel scheduling and offloads. In addition, considering that ACK is inevitably ranked behind other packets when transmitted on the reverse congested path, it affects the measurement accuracy. In order to mitigate the impact of congestion on the reverse path, ACK is assigned to the highest priority queue to ensure that it can be forwarded first, further improving the accuracy of delay measurement.

[0043] Therefore, this embodiment can accurately measure the delay of data packet queuing.

[0044] As an optional implementation, the round-trip time of a data packet without queuing delay is calculated, specifically: the sender of the data packet continuously measures the round-trip delay RTT experienced by each data packet in the transmission path, and takes the minimum value of all measured RTT values ​​as the round-trip delay without queuing delay.

[0045] Calculate the minimum queuing delay of the data packet, including:

[0046] S101. Calculate the queuing delay T experienced by each data packet q , the calculation formula is:

[0047] T q =RTT rt -RTT min

[0048] In the above formula, RTT rt Indicates the RTT value measured in real time, RTT min Indicates the minimum measured RTT value.

[0049] S102, using the weighted moving average method to tolerate noise and calculate the final packet queuing delay T q , the calculation formula is as follows:

[0050] T q =(1-α)×T q +α×T p

[0051] In the above formula, α represents the weight for calculating the queuing delay, which ranges from 0 to 1. p Indicates the queuing delay experienced by the previous packet.

[0052] S103: The queuing delay T experienced by the previous data packet p Update to T q , and according to T q For the minimum queuing delay T min to update.

[0053] In order to maintain a stable link utilization, this implementation uses a weighted moving average method to tolerate noise, so as to effectively cope with the network environment of noise interference and irregular data packet transmission, thereby reducing the impact of short-term delay fluctuations on the overall measurement trend and link performance evaluation.

[0054] As a possible implementation mode, the sender of the data packet adopts a corresponding strategy to perform a data flow transmission suspension operation according to the delay measurement result, including:

[0055] Determine whether the minimum queuing delay is 0, and analyze the traffic type, wherein the traffic type includes long flows and short flows;

[0056] If the traffic type is long flow and the minimum queuing delay is greater than 0, the packet transmission operation is immediately suspended for a period equal to the measured minimum queuing delay;

[0057] If the traffic type is short flow and the minimum queuing delay is not 0, then 1 / n e The probability of deciding whether to perform the transmission suspension operation is n e Indicates the number of competing flows on the bottleneck link.

[0058] Exemplarily, the sender can transmit data according to the initial window or rate of each data stream. After receiving the ACK of the first data packet, the sender starts a delay measurement timer and sets the timeout of the timer to the minimum measured RTT value for periodic measurement of the minimum queuing delay. Specifically, the sender calculates the real-time measured RTT value based on the reception time of the ACK and the transmission time of the corresponding data packet (extracted from the timestamp field of the ACK). Then, the queuing delay experienced by the corresponding data packet is obtained by subtracting the minimum measured RTT value from the real-time measured RTT value. If the minimum queuing delay is 0, the sender can infer that all data packets have no queuing delay, and there is no need to suspend the sending of data packets. On the contrary, if the minimum queuing delay is greater than 0, it means that the bottleneck link has been fully utilized and the queuing delay is at least equal to the minimum queuing delay. In this case, the sender should suspend sending new data packets to eliminate queue backlogs.

[0059] However, if multiple data flows on the bottleneck link perform transmission operations simultaneously, high link utilization may not be guaranteed, thereby reducing network performance.

[0060] In order to solve this problem, this implementation method performs a transmission suspension operation according to the measured minimum queuing delay and traffic type. On the one hand, since most short flows are delay-sensitive traffic, as long as the long flow detects that the minimum queuing delay is greater than 0, the transmission suspension operation is immediately performed and the suspension time is the measured minimum queuing delay. This can minimize the interference of the long flow with the transmission of the short flow. On the other hand, the short flow can use the existing traffic measurement technology to obtain the number of competing flows n on the bottleneck link. e If the minimum queuing delay measured by the short-flow is not 0, then 1 / n e The probability of suspending transmission is used to determine whether to perform the transmission suspension operation. This can effectively eliminate queue backlogs and effectively avoid the bottleneck link from performing suspension operations due to too many short flows, which would reduce link utilization and transmission performance. This allows the method provided in this embodiment to prioritize the transmission of short flows and achieve ultra-low transmission delay and high link utilization.

[0061] In some embodiments, reference Figure 2 , the method can be implemented by the following operations:

[0062] Step 1: Parameter initialization, including setting the initial congestion window size and packet size of the data flow;

[0063] Step 2: Send data packets according to the window size;

[0064] Step 3: The sender receives the ACK and calculates the instantaneous round-trip delay based on its arrival time and header timestamp;

[0065] Step 4: Update the minimum round-trip delay RTT based on the instantaneous RTT statistics min ; And calculate the instantaneous queuing delay through the instantaneous round-trip delay and the minimum round-trip delay;

[0066] Step 5: Use the weighted average moving method to calculate the final packet queuing delay T q ;

[0067] Step 6: Substitute the queuing delay T experienced by the previous packet p Update to T q ;

[0068] Step 7: Update the minimum queuing delay T min ;

[0069] Step 8: Determine the current queue delay T min Is it greater than 0? If so, the long stream and the probability-selected short streams will suspend transmission; otherwise, the remaining data will continue to be sent;

[0070] Step 9: Enable the Pause Transfer Timer and set its Pause Transfer Time to T min ;

[0071] Step 10: Before the transmission suspension timer expires, the sender cannot send new data;

[0072] Step 11: After the transmission pause timer times out, the remaining data is sent;

[0073] Step 12: Determine whether the data has been sent. If so, end; otherwise, return to step 2 to continue sending new data.

[0074] Figure 3 Demonstrates the method provided by this embodiment (UL 2 Since the method is deployed between the network layer and the transport layer, it can be quickly integrated into the existing transmission control algorithm, effectively eliminating queue backlogs and thus reducing the queuing delay of data packets, greatly improving the overall performance of the network.

[0075] The following models are constructed for the existing solution and the method provided by this embodiment respectively to illustrate the technical effect that can be achieved by the method provided by this embodiment.

[0076] Model building of existing solutions:

[0077] Assume that n flows start transmitting simultaneously on a shared bottleneck link, and their initial congestion windows are all w0. The round-trip propagation time of the flow is RTT0, and the link capacity is C (packets / second). The transmission time of each data flow can be divided into the following three stages:

[0078] The first stage is that all flows continuously increase their sending rates until the bottleneck link is fully utilized. In this stage, all flows increase their congestion windows. Assuming that the number of rounds of flow transmission is R1, the link is fully utilized after R1 transmission rounds. Therefore, the number of rounds R1 and the number of transmission packets S1 can be expressed as:

[0079]

[0080] Considering that data center transport protocols widely use ECN as a congestion signal to control queue length and achieve high throughput, it is assumed that the timeout-free transport protocol supports ECN and the ECN marking threshold is set to K packets.

[0081] The second phase is defined as the queue length increases from 1 to the marking threshold K. In this phase, the total number of packets added by each RTT0 in all flows is n. Obviously, after After rounds, the queue length becomes K. At the same time, the average queue length is Therefore, the number of S2 packets sent by each flow can be calculated as:

[0082]

[0083] Ideally, the queue length should be kept at a fixed value K. However, due to the highly dynamic and bursty nature of network traffic in data centers, bottleneck links are likely to experience frequent transient congestion. Once burst traffic reaches a shallowly buffered switch, it will inevitably cause a violent jitter in the queue length. At this point, some packets are discarded due to switch buffer overflow, degrading network performance.

[0084] In the third stage, assuming that after the queue length is K, the background flow on the transmission path generates B every m RTTs f packets (i.e. instantaneous traffic), so that the average queue length of the m+1 round becomes Since the background flow periodically generates instantaneous traffic and the traffic intensity is the same each time, the average queue length in the third stage is At this stage, the number of packets sent by each flow is S3 = S-S1-S2, where S is the size of each flow. Therefore, the average queue delay QD experienced by each packet is a It can be calculated by the following formula:

[0085]

[0086] In the above formula, the variable t s Indicates the transmission time or forwarding time required for each data packet.

[0087] Model construction of the method provided in this embodiment:

[0088] For the timeout-free transmission protocol integrated with the method provided by this embodiment, all flows first increase their congestion windows until the bottleneck link is fully utilized. Assuming that the number of rounds of flow transmission is R1, the link is fully utilized after R1 transmission rounds, so the number of rounds R1 and the number of transmission packets S1 can be expressed as:

[0089]

[0090] During the data transmission phase, the sender monitors the status of the transmission path in real time. Once queue accumulation is detected in the switch cache queue, long flows and some short flows immediately execute a pause mechanism to eliminate queue backlogs. In other words, in the absence of background flows, the queuing delay experienced by each data packet is very low (close to zero).

[0091] Assume that the background flow generates B every m round trips f packets, the average queue length is Therefore, the average queue delay experienced by each packet is:

[0092]

[0093] Based on the average queue delay calculation formula obtained from the existing solution and the model constructed by the method provided in this embodiment, we plotted the queue delay improvement percentage under different numbers of flows. Among them, the queue delay improvement is defined as In addition, the variables RTT, m, and B are set to 20 microseconds, 200 microseconds, 100, and 65 packets, respectively. Figure 4 The comparison results are shown in .

[0094] Figure 4 In (a), the size of each flow is set to 1000 packets, and the queue delay improves significantly when the number of flows increases from 10 to 50. The transmission protocol that does not use the method provided in this embodiment obtains a larger queue delay in a higher concurrent flow scenario. With the help of the method provided in this embodiment, the transmission protocol effectively eliminates queue backlogs through a pause mechanism to obtain ultra-low latency. Compared with the original transmission protocol, this method reduces the queue delay by more than 93%. On the other hand, the larger the bandwidth, the smaller the forwarding delay of the data packet. Therefore, the percentage of queue delay improvement is reduced. Figure 4 In (b), the link capacity C is set to 10 Gbps and the calculation is performed by changing the flow size. The results show that as the flow size increases, the percentage of queue delay improvement also increases. The main reason is that with the help of this method, more packets will not experience queue delay, which reduces the average queue delay and thus obtains a higher percentage of queue delay improvement.

[0095] We use real workloads in data center networks to further evaluate the effectiveness and wide applicability of the present invention. In this test, we consider two typical workloads: web search and hadoop, whose traffic distributions are both heavy-tailed (i.e., a small number of long flows contribute most of the traffic). In addition, we integrate the proposed method into DCTCP, T-RACKs, CP, and pHost, and use the symbol and They represent DCTCP, T-RACKs, CP and pHost respectively, which enable the method provided in this embodiment. The performance indicators we tested are packet loss rate, AFCT of all flows, AFCT of short flows (≤100KB) and AFCT of long flows (>100KB). The experimental results are shown in Figure 5 and Figure 6 shown.

[0096] like Figure 5 and Figure 6The experimental results show that as the network load increases, the packet loss rate and average flow completion time of various protocols are constantly increasing. Among them, the packet loss rate of the CP protocol is larger than that of other protocols. The main reason is that CP can accurately identify packet loss and set the fast retransmission threshold to a larger value, so that when data packets are lost, the data flow will not reduce its sending rate. In addition, since the pHost flow transmits data at the line rate in the first RTT, this can easily lead to serious queue backlogs. Therefore, the pHost flow obtains a larger flow completion time than DCTCP, T-RACKs, and CP.

[0097] In addition, experimental results show that after the existing transmission control protocol integrates the method proposed in this invention, the queue backlog can be effectively eliminated by pausing the transmission mechanism, greatly reducing the large queue delay experienced by data packets. Finally, all transmission protocols integrated with this method achieve lower flow completion time and near-zero packet loss rate than the original transmission protocol, significantly improving network performance.

[0098] In another embodiment of the present invention, a transmission control system based on flow and delay perception in a data center network is provided, the system comprising:

[0099] A delay measurement module is used to calculate the round-trip delay and minimum queuing delay of a data packet without queuing delay to obtain a delay measurement result;

[0100] The transmission suspension module is used by the sender of the data packet to execute the transmission suspension operation of the data flow according to the delay measurement results and different traffic types using corresponding strategies.

[0101] The system is used to execute the method described in the aforementioned method embodiment. Its working principle and technical effects can all be referred to the aforementioned method embodiment and will not be repeated here.

[0102] In another embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the steps described in the above method embodiment are implemented.

[0103] The above-mentioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general or special-purpose computer.

[0104] An exemplary readable storage medium is coupled to a processor so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (Application Specific Integrated Circuits, referred to as: ASIC). Of course, the processor and the readable storage medium can also exist as discrete components in an electronic device or a main control device.

[0105] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A transmission control method based on flow and delay perception in a data center network, characterized in that: The method comprises: Delay measurement: calculate the round-trip delay and minimum queuing delay of the data packet without queuing delay to obtain the delay measurement result; Pause transmission,The sender of the data packet adopts corresponding strategies to execute the suspension transmission operation of the data flow according to the delay measurement results and different traffic types.

2. According to claim 1, a transmission control method based on flow and delay perception in a data center network is characterized in that: Calculate the round-trip time of the data packet without queuing delay. Specifically, the sender of the data packet continuously measures the round-trip delay RTT experienced by each data packet in the transmission path, and takes the minimum value of all measured RTTs as the round-trip delay without queuing delay.

3. According to claim 1, a transmission control method based on flow and delay perception in a data center network is characterized in that: Calculate the minimum queuing delay of the data packet, including: S101. Calculate the queuing delay T experienced by each data packet q , the calculation formula is: T q =RTT rt -RTT min In the above formula, RTT rt Indicates the RTT value measured in real time, RTT min Indicates the minimum measured RTT value; S102, using the weighted moving average method to tolerate noise and calculate the final packet queuing delay T q , the calculation formula is as follows: T q =(1-α)×T q +α×T p In the above formula, α represents the weight for calculating the queuing delay, which ranges from 0 to 1. p Indicates the queuing delay experienced by the previous data packet; S103: The queuing delay T experienced by the previous data packet p Update to T q , and according to T q For the minimum queuing delay T min to update.

4. The method for transmission control based on flow and delay perception in a data center network according to claim 1, characterized in that: The sender of the data packet uses a corresponding strategy to execute the data flow transmission suspension operation according to the delay measurement result, including: Determine whether the minimum queuing delay is 0, and analyze the traffic type, wherein the traffic type includes long flows and short flows; If the traffic type is long flow and the minimum queuing delay is greater than 0, the packet transmission operation is immediately suspended for a period equal to the measured minimum queuing delay; If the traffic type is short flow and the minimum queuing delay is not 0, then 1 / n e The probability of deciding whether to perform the transmission suspension operation is n e Indicates the number of competing flows on the bottleneck link; If the minimum queuing delay is greater than 0, the sending of new packets is suspended to eliminate the backlog of packets to be sent.

5. A transmission control system based on flow and delay perception in a data center network, characterized in that: The system comprises: A delay measurement module is used to calculate the round-trip delay and minimum queuing delay of a data packet without queuing delay to obtain a delay measurement result; The transmission suspension module is used for the sender of the data packet to execute the transmission suspension operation of the data flow according to the delay measurement result and different traffic types using corresponding strategies.

6. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

Images