CN101056260A - ECN mechanism-based congestion control method in the mixed network - Google Patents

Abstract

The invention discloses a congestion control method based on an ECN mechanism in a hybrid network, comprising the following steps: intercepting the latest ECN feedback information of a router, and calculating the congestion probability according to the distribution of the congestion marks in it, and the sender uses the change trend of the congestion probability Carry out congestion prediction, and adjust the congestion control mechanism according to the prediction result. The present invention performs congestion inference through the change law of the congestion probability value, that is, it is inferred as congestion only when the congestion probability value shows an increasing trend, thereby avoiding the instantaneousness and blindness of network state judgment when independent ECN feedback is used. The result of congestion prediction is used to control the TCP packet loss response behavior, which can more accurately eliminate the impact of non-congested packet loss on TCP congestion control in a hybrid network.

Description

Congestion Control Method Based on ECN Mechanism in Hybrid Network

technical field

The invention relates to a congestion control method based on an ECN mechanism in a hybrid network.

Background technique

Different from wired networks, wireless networks have high bit errors, complex channel fading, and sudden equipment noise. These characteristics will lead to frequent packet loss, but traditional TCP simply attributes all packet loss to network congestion. And blindly adopt the control strategy, which greatly reduces the performance of TCP. Therefore, in a wired/wireless hybrid network, TCP needs to eliminate the impact of non-congested packet loss on the original congestion control mechanism.

In a wired/wireless hybrid network, TCP performance is mainly improved by distinguishing between congestion packet loss and wireless error packet loss. Packet loss discrimination methods include NCPLD ^[4] based on round-trip delay, TCP-Biaz ^[5] based on packet arrival interval, TCP-Spike ^[6] based on ROTT, TCP-Bayes ^[7] based on RTT change, and TCP-HACK ^[8] for header checksums, etc. These existing methods mainly have the following problems:

(1) Most discrimination methods adopt end-to-end measures, and they have strong dynamics in composing increasingly complex hybrid networks, so it is difficult to establish the correspondence between these measures and network states. This will directly affect the accuracy of packet loss discrimination and the corresponding TCP performance. The TCP-Bayes method pays full attention to this feature, and establishes an HMM model that shows the dynamic relationship between RTT and packet loss properties, and keeps the discrimination accuracy above 80%, but obtaining this model requires learning and training based on historical data , the price paid will greatly limit the application of this method in TCP control.

(2) The current evaluation of the differentiation method uses the accuracy index of the packet loss event, but in the hybrid network, due to the long delay characteristic, the response to a single packet loss event is not real-time, that is to say, the difference obtained by the sender The information is actually the state of the network at a certain point in the past. Then even if the correct rate of distinction reaches 100%, it may not help to improve the performance of TCP. How to effectively use the packet loss event and its distinguishing information in TCP is very critical.

As an important TCP protocol in the current hybrid network, TCP Westwood and Jersey both adopt the method of distinguishing packet loss.

TCP continuously measures the effective bandwidth by monitoring the return ACK rate, and uses the current effective bandwidth to estimate the congestion window and slow start threshold. This method adopts different control strategies based on the measurement results of the system. In fact, the rules of packet loss identification are implied into the global control. Its bottleneck lies in the delay and cumulative effect of ACK arrival, which affects the accuracy of bandwidth estimation.

TCP Jersey ^[13] : Based on Westwood, TCP Jersey enhances the ability to distinguish packet loss by means of the ECN mechanism. When packet loss occurs, the sender detects ACK feedback. If the congestion flag (Congestion Explicit) is set to 1, it will The packet loss is attributed to congestion, otherwise it is attributed to errors. This is the Wireless_ECN method proposed by Fei and Shiduan, hereinafter referred to as WECN. However, some studies in recent years have pointed out that the packet loss discrimination ability of the ECN mechanism is relatively weak, and it depends entirely on the accuracy of congestion perception. For this reason, a lot of work has improved the packet loss discrimination algorithm based on the ECN mechanism from two aspects.

The first line of work aims to improve the accuracy of the ECN mechanism for the perception/marking of congestion.

SpecTCP minimizes congestion and packet loss by optimizing various parameters of AQM (buffer size, marking threshold, packet loss probability, etc.), but this parameter optimization is difficult to guarantee in a network with multiple flows and bottlenecks because it is statically configured. Minimize packet loss due to congestion.

Multilevel ECN subdivides the marking probability interval into two levels, and designs corresponding marking probabilities respectively. This method improves the accuracy of ECN marking, but also increases the complexity of router work, and the setting of parameters becomes more sensitive.

The second aspect of work aims to reduce the delay of ECN feedback and ensure the real-time nature of TCP response. If the ECN mechanism can accurately mark congestion in various traffic occurrence models, how should it respond to ECN congestion marking?

The ACK Spoofing/Congestion Signal Cancellation method allows the router to update the network status in a timely manner during the process of feeding back the congestion mark. ACK Spoofing notifies the sender of congestion in advance by generating 3 repeated ACKs. Congestion Signal Cancellation caches the congestion mark in advance and waits for the next The router confirms the network status at all times and then decides whether to send the mark. These two technologies are to allow the sender to capture the latest network status in the shortest time, but the completion of such work by the router will undoubtedly increase the system overhead and implementation difficulty.

It can be seen that the congestion control method based on the ECN mechanism has received extensive attention, and the existing optimization schemes have shown great application potential. However, for the second problem described above, that is, how to use ECN congestion feedback to control TCP behavior, it is still necessary to find a more concise and effective solution, which is also a key point to improve the performance of the TCP protocol in the current hybrid network.

Contents of the invention

In order to solve the technical problems in the above-mentioned congestion control method based on the ECN mechanism, the present invention provides a congestion control method based on the ECN mechanism in a hybrid network. The method can combine the packet loss event and the congestion probability to regulate the rate, thereby ensuring the robustness of TCP under the condition of increasing link errors.

The technical scheme for solving the above-mentioned technical problems of the present invention includes the following steps: intercept the latest section of ECN feedback information of the router, and calculate the congestion probability according to the distribution of the congestion marks in it, the sender performs congestion prediction according to the change trend of the congestion probability, and according to the prediction result Adjust the congestion control mechanism.

In the above-mentioned congestion control method based on the ECN mechanism in the hybrid network, the congestion probability is calculated as follows:

$\mathrm{<span\; class="notranslate">\; CP</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; Ack</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; ecnecho</span>}$

In the above formula, Ack is the latest ACK sequence collected by the sender, Ack[i]->ecnecho is the value of the CE bit in the i-th ACK packet, w _i is the weight of ACK[i], $\underset{i=0}{\overset{m-1}{\mathrm{\&Sigma;}}}{w}_i=1.$

In the above-mentioned congestion control method based on the ECN mechanism in the hybrid network, when the change of the congestion probability value shows an increasing trend, the sender infers that the current network is in a congested state.

In the above-mentioned congestion control method based on the ECN mechanism in the hybrid network, the TCP sender starts the corresponding congestion control mechanism such as speed reduction only when the predicted result is congestion, otherwise, the lost data packet is just resent.

The technical effects of the present invention are: in the present invention, the sender refreshes the ACK array when receiving each ACK packet, and calculates a new congestion probability value, controls the normal ACK completely according to the inherent ECN mechanism, and for the repeated ACK, according to The congestion probability value infers the network state, and only when the inferred result is congestion and packet loss, the sending window size is reduced. The method of the present invention performs congestion inference through the change law of the congestion probability value, that is, it is inferred as congestion only when the congestion probability value shows an increasing trend, thereby avoiding the instantaneousness and blindness of network state judgment when using independent ECN feedback . Since only the sender code needs to be modified in the implementation of the present invention, the method of the present invention can be applied to various TCP protocols to adapt to the transmission characteristics of heterogeneous links in the wired/wireless hybrid network, so as to ensure that the network obtains higher throughput and link utilization, while maintaining the friendliness with the original TCP protocol and the fairness between flows.

The present invention will be further described below in conjunction with the accompanying drawings.

Description of drawings

Fig. 1 is a flowchart of the present invention.

Fig. 2 is the TCP throughput simulation environment of the present invention.

Fig. 3 is the effect diagram of the performance analysis of the present invention applied in Westwood, wherein the present invention is named CPECN. Figure 3(a) is the throughput comparison of different marking methods; Figure 3(b) is the end-to-end packet loss rate comparison; (c) the change of the sending window.

Fig. 4 is the simulation environment of the TCP protocol friendliness of the present invention.

Detailed ways

Referring to Fig. 1, Fig. 1 is a flowchart of the present invention. The process is as follows: After TCP starts, it first enters slow start, and when the congestion window exceeds the slow start threshold, it enters the congestion avoidance stage. In the congestion avoidance phase, if a timeout occurs, slow start is restarted; if a new ACK is received, congestion avoidance is performed according to normal TCP; but if repeated ACKs are received, congestion prediction is required. The congestion prediction method is as follows: intercept the latest ECN feedback information of the router, and calculate the congestion probability according to the distribution of the congestion marks in it, the sender performs congestion prediction according to the change trend of the congestion probability, and adjusts the congestion control mechanism according to the prediction result. If the predicted result is congestion, the sender will adjust the rate and retransmit to enter fast recovery; if the predicted result is non-congested, then the sender will only retransmit and enter fast recovery.

In the present invention, "congestion probability" is used to indicate the analysis result. Specifically, the TCP sender collects the feedback information of the latest ACK, and uses the EWMA (Exponentially Weighted Moving Average) method to predict the congestion probability, which is called the "congestion possibility factor". The calculation method is as follows:

$\mathrm{CP}=\underset{i=0}{\overset{m-1}{\mathrm{\&Sigma;}}}{w}_i*\mathrm{Ack}\left[i\right]->\mathrm{ecnecho},$ Formula 1)

where Ack[i]->ecnecho={0,1}, $\underset{i=0}{\overset{m-1}{\mathrm{\&Sigma;}}}{w}_i=1,w_i=\left\{\begin{array}{c}\mathrm{\&alpha;},i\&Element;[0,m/2-1]\\ \mathrm{\&beta;},i\&Element;[m/2,m-1]\end{array}\right.,$

The larger m is, the more ECN feedback is obtained, which is more conducive to the accuracy of prediction probability, but at the same time it will affect the real-time performance of congestion control and error control. Following Reno's practice of 3 consecutive repeated ACKs to enter slow start, and intercepting the next ACK packet to determine the state of the intermediate router, so in the experiment, the P congestion probability is calculated in units of 4 ACK packets, and the probability is calculated by combining the two units before and after. Analysis, ie m=8. According to the EWMA method, a larger weight α is taken for the latest 4 ACK packets, and a smaller weight β is taken for the previous 4 ACKs, so that the sum of the full probabilities of the first 4 ACKs is equal to the full probability of the last 1 ACK. probability, then: $\left\{\begin{array}{c}\mathrm{\&Sigma;\&beta;}=\mathrm{\&alpha;}\\ 4*\mathrm{\&alpha;}+4*\mathrm{\&beta;}=1\end{array}\right.$ Namely: α=0.2, β=0.05.

To sum up, adopting the method of the present invention can prevent the TCP connection from being affected by the short-term recoverable non-congestion packet loss event, thereby enhancing the ability of TCP to fight against error links.

Utilizing the NS2.28 network simulation platform, we realized the method of the present invention in TCP Westwood, and tested its performance. NS network simulator is a general multi-protocol network simulation software, which is publicly released on the Internet (URL: http://www.isi.edu/nsnam/ns), and has been widely used by network researchers. NS2.28 is one of its versions.

We use the topology shown in Figure 4 to test the performance of the method of the present invention. As shown in Figure 4, a single TCP flow is sent from R0 to R3, and the long-delay Internet passing through multiple routers is simulated between R1 and R2, and the last hop is a wireless link with error packet loss. Specifically, the link bandwidth of the user access terminal is 100M, the wireless link bandwidth is 10M, the delay is 3ms, the bottleneck bandwidth is 5M, the delay is 40ms, and the buffer size on the bottleneck is set as the pipeline capacity. In Westwood we also added the WECN method as a comparison. In order to test the TCP performance in the case of link packet loss, we introduced a simple error packet loss model in the simulation experiment, and used the link packet loss rate as a parameter of the model. By default the simulation lasts 100 seconds. The packet size of all protocol streams is set to 1Kbyte.

1. Analysis of throughput

Fig. 3 (a) has provided the comparison of the throughput when the WECN method and the method of the present invention are applied to TCPWestwood under different packet loss rates, assuming that a simple threshold marking method is adopted on the router, as can be seen from the figure, no matter how the threshold on the router is set , the application of the WECN method can obtain higher throughput than the application of the method of the present invention.

In order to further analyze the reasons for the increase in throughput, Figure 3(b) and Figure 3(c) show the change of link packet loss rate and TCP window when the marking threshold is 3/4 of the queue size (here the packet loss rate is equal to refers to the link packet loss rate specified in the error model, and the end-to-end loss rate refers to the ratio of received packets to sent packets calculated by the TCP sender), Figure 3(b) shows that Westwood’s end-to-end packet loss rate varies with With the increase of link packet loss rate, the increase range is the largest, especially when the packet loss rate exceeds 0.005. Application of the WECN method and the method of the present invention does not increase the end-to-end packet loss rate, indicating that the method of the present invention improves throughput by increasing link utilization.

Figure 3(c) compares the window changes during the period from 0 to 30 ms (take Loss Rate as 0.01), Westwood has entered slow start many times, especially around 30 ms, Westwood’s sending window has serious jitter; after applying the WECN method, The sender also enters a slow start after packet loss and receiving the congestion mark fed back by a single ECN; after applying the method of the present invention, since the prediction result based on the change trend of the congestion probability value is "False", the TCP does not stop when the packet is lost. Entering the slow start, the method of the present invention has a judgment different from that of the WECN method at around 22ms. Generally speaking, the sending window of the method of the present invention is more stable.

In the network shown in Figure 2, the ON/OFF flow that obeys the exponential distribution is loaded, and its rate is 5Mbps. In the RTT time of about 100ms, there are 25ms in the "on" state, and the rest of the time slots are in the "off" state. Table 1 describes the results of the UDP/TCP mixed flow test.

Table 1 Comparison of TCP/UDP throughput in the case of loading on/off flow Packet loss rate 0.001 0.005 0.01 protocol TCP UDP TCP UDP TCP UDP Westwood 1694 936 1457 971 1346 970 Westwood with WECN 1437 1114 1393 1041 1359 1039 Westwood with this invention 1679 994 1618 1039 1482 1055

It can be seen from Table 1 that Westwood has a better ability to share bandwidth with UDP streams in a mixed network, but as the packet loss rate increases from 0.001 to 0.01, the throughput of TCP streams decreases by 20%; Competing with the UDP flow for the channel does not show greater superiority; after applying the method of the invention, the overall throughput of TCP is optimal, and with the increase of the packet loss rate, the throughput only drops by 11%.

2. Fairness and friendliness analysis

Fairness and friendliness are important properties of the TCP protocol. Multiple connections of the same TCP must coexist harmoniously and quickly converge to a fair point. At the same time, the new TCP protocol should be able to coexist with other traditional TCP protocols without causing the use of traditional protocols. The TCP flow is starved. We use the topology shown in Figure 6 to test the fairness and friendliness of TCP when the method of the present invention is applied.

In the fairness test, 10 identical TCP streams are used to access the bottleneck link. The total bandwidth of the 10 streams is 1000 Mbps, and they share a 20 Mbps long-delay bottleneck link. See the annotations in Figure 4 for details. The test results in Table 2 show that the fairness of TCP Westwood under different packet loss rates is better after applying the present invention.

Table 2 Fairness comparison of 10 identical TCP flows Loss Rate Westwood Westwood with WECN Westwoodwith the invention 0 0.935 0.962 0.997 0.001 0.998 0.999 0.999 0.005 0.999 1.0 1.0 0.01 0.999 0.998 0.999

In the friendliness experiment, 10 TCP streams are assigned to two types of protocols respectively, one type adopts the Westwood protocol applying the method of the present invention, and the other type adopts the original Westwood protocol. Table 3 shows the variation of the average throughput when the two types of protocols occupy different connections, from which it can be seen that the method of the present invention can coexist friendly with Westwood in various situations.

Table 3 The inventive method and Westwood friendliness experiment Westwood Source Westwoodwith Invention Source no loss Loss Rate: 0.001 Westwood Westwoodwith the invention Westwood Westwoodwith the invention 3 7 1162.00 1174.85 1161.33 1174.14 5 5 1316.80 1331.40 1312.80 1329.40 7 3 1519.28 1536.66 1511 1539.00 Westwood Source Westwoodwith Invention Source Loss Rate: 0.005 Loss Rate: 0.01 Westwood Westwoodwith the invention Westwood Westwoodwith the invention 3 7 1289.66 1314.85 1099.00 1186.57 5 5 1294.80 1319.80 1269.80 1352.2 7 3 1485.28 1508.33 1388.14 1506.00

From the above experimental data, the performance of these two aspects of the Westwood protocol using the method of the present invention is relatively ideal. This is because this TCP protocol for wireless links in Westwood itself has overcome some defects of traditional TCP protocols such as Reno. First of all, the method of adjusting SSTHRESH and CWND adopted by Westwood can control the TCP rate more effectively and reduce the impact of link errors than the method of directly setting these two values to 1 or halving them in Reno; secondly, the application The WECN method plays a certain role in separating error control and congestion control at the sending end by means of the congestion marking of intermediate routers. On the basis of them, the method of the present invention makes more reasonable use of the feedback congestion marks, so it performs better in terms of its own fairness and friendliness with Westwood.

Claims (5)

1. A congestion control method based on an ECN mechanism in a hybrid network, comprising the steps of: intercepting the latest section of ECN feedback information of a router, and calculating the congestion probability according to the distribution of congestion marks wherein, the sender performs congestion by the variation trend of the congestion probability Prediction, and adjust the congestion control mechanism according to the prediction result. 2. the congestion control method based on ECN mechanism in the hybrid network according to claim 1, described congestion probability is calculated as follows: $\mathrm{<span\; class="notranslate">\; CP</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; Ack</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; ecnecho</span>}$ In the above formula, Ack is the latest ACK sequence collected by the sender, Ack[i]->ecnecho is the value of the CE bit in the i-th ACK packet, w _i is the weight of ACK[i], $\underset{i=0}{\overset{m-1}{\mathrm{\&Sigma;}}}{w}_i=1.$ 3. According to the congestion control method based on the ECN mechanism in the hybrid network according to claim 1 or 2, when the congestion probability value changes and presents an increasing trend, the sender infers that the current network is in a congested state. 4. The congestion control method based on the ECN mechanism in the hybrid network according to claim 1, only when the predicted result is congestion, the TCP sender starts the corresponding congestion control mechanism such as slowing down, otherwise it just resends the lost data packet. 5. The congestion control method based on the ECN mechanism in the hybrid network according to claim 3, when the sender receives an ACK, first refreshes the ACK array and calculates the latest congestion probability, if the ACK is a normal value, then No matter what the current congestion probability is, congestion avoidance or slow start is maintained like the inherent TCP congestion probability protocol; if the ACK is a duplicate ACK and the result predicted by the congestion probability is congestion, then the sender will perform rate adjustment and retransmit Enter fast recovery; if the ACK is a repeated ACK and the result predicted by the congestion probability is non-congested, then the sender only resends and enters fast recovery.

Images