KIND OF TRANSMISSION METHOD BASED ON THE NETWORK LEARNABLE POWER MODEL

Abstract

A kind of transmission method based on the learnable power model, which conducts periodic record for the historical change trend of the network. This method conducts weighting smooth processing on the round trip time and judges the changing trend of congestion control window. Then, it establishes model for the relationship between network power and the congestion control widow. When a new ACK is received, it immediately updates the window of power model. Finally, it forecasts the size of the congestion control window of the next time period by combining the congestion window and the network power changing trend. For the network packet loss or time-out events, the retransmission mechanism of traditional TCP is used, and when the packet loss ends, the power model process is used again. This invention reduces the influence of the network random events of the estimation error of traditional algorithm.

Claims

1. A kind of transmission method based on the learnable power efficiency model, it comprises the steps as follows: Step 1: Record the round trip time vector of all ACK packets received {right arrow over (D.sub.i)} and the sending window vector {right arrow over (W.sub.i)}, in which i means the current time period adopt the formula D.sub.g,i=α.Math.D.sub.g,i−1+(1−α).Math.{right arrow over (D.sub.i)} to conduct weighting smooth processing on the round trip time vector {right arrow over (D.sub.i)}; wherein, i−1 refers to the last time period, α is the weighting smooth factor and D.sub.g,i is the round trip time of maximum delay of ith time period after smoothing; Step 2: Calculate the normalized time delay change rate G.sub.i in accordance with the formula $G_i=\left(\frac{D_{g,i}-D_{g,i-1}}{D_{g,i}+D_{g,i-1}}\right).Math.\left(\frac{W_{g,i}+W_{g,i-1}}{W_{g,i}-W_{g,i-1}}\right).$ wherein, W.sub.g,i is the window size of the maximum time delay value of the ith time period; Step 3: Calculate the network power estimation value E.sub.e,i in accordance with the formula $E_{e,i}=\frac{T_i^{\Delta }}{D_i},$ Δ refers to the sensitive factor of delay-throughput and T.sub.i refers to the actual throughput of the ith time period, which are gotten through accumulation of the confirmed send window vector {right arrow over (W.sub.i)} in each time period, D.sub.i refers to all elements in the round trip time vector {right arrow over (D.sub.i)} of the i time period; Step 4: Establish the corresponding relationship model between the corresponding network power capacity E of every data packet and the send window size W, which is define as W(x+1)=f(E(x)+γ(x)); wherein, x is the time change parameter, i.e. time period I, W(x+1) is the congestion control window size of the next time period, f(x) is the window control function, E(x) is the estimation function of network power and γ(x) is the increase and decrease size of network power; Step 5: Judge the changing trend of network status; when G.sub.i>0 or $\frac{D_{\max ,i}}{D_{\min }}>R,$ the network becomes congestion and the network power should correspondingly decrease in the next time period; when G.sub.i≦0 and $\frac{D_{\max ,i}}{D_{\min }}>R,$ network becomes free and the network power should correspondingly increase in the next time period, which is $E_{e,i+1}=\{\begin{array}{cc}{E}_{e,i}-{\gamma }_2& \mathrm{When}.Math..Math.\frac{D_{\max ,i}}{D_{\min }}>R\\ E_{e,i}-{\gamma }_1& \mathrm{When}.Math..Math.G_i>0\\ E_{e,i}+{\gamma }_2& \mathrm{Otherwise}\end{array}$ in detail; E.sub.e,i+1 is the network power estimation value of the i+1 time period, D.sub.max,i is the maximum element in the round trip time vector D.sub.i of i time period, D.sub.min is the minimum element in the round trip time vector {right arrow over (D.sub.i)} and R is the maximum value of the link tolerance; γ.sub.1 and γ.sub.2 refer to the increase and decrease value of the network power; through a large quantity of repeated experiments, the optimum values are selected; Step 6: Calculate the W.sub.i+1 window value according to Step 4 and Step 5; Step 7: Divide every time period into n time slots; every time slot is λ ms; after the i time period, the data sent within n−1 time slots is taken as a part of data to W.sub.i+1. For instance, first n−1 time slots in every time period are used for the learnable model and the last time slot is used to send the remaining data; so the size of the data which shall be sent of the last time slot of the W.sub.i+1 window is W.sub.fin,i+1 and $W_{\mathrm{fin},i+1}=\max \left[0,\left(W_{i+1}-\frac{n-2}{n-1}.Math.W_i\right)\right].$

2. The transmission method based on the learnable power model as mentioned in claim 1 wherein the value of α is 0.875 in Step 1.

3. The transmission method based on the learnable power model as mentioned in claim 1 wherein the model W(x+1) is learned and gotten by adoption of the linear regression algorithm in the machine learning algorithm; the input parameters are the corresponding network carrying capacity E of every data packet and the send window size W of the ith time period.

4. The transmission method based on the learnable power model as mentioned in claim 1 wherein the value of γ.sub.1 is 2 ms and the value of γ.sub.2 is 3 ms.

5. The transmission method based on the learnable power model as mentioned in claim 1 wherein every time slot length λ is 5 ms, $n=.Math.\frac{{\stackrel{.Math.}{D}}_{\mathrm{ave}}}{\lambda }.Math.,$ in which {right arrow over (D)}.sub.ave refers to the mean value of all elements in the round trip time vector {right arrow over (D.sub.i)}.

6. The transmission method based on the learnable power model as mentioned in claim 1 wherein if the network packet loss or time-out issue occurs, the multiplicative reduction mechanism with β as the multiplicative factor shall be conducted, i.e. W.sub.i+1=β.Math.W.sub.i, in which β is the multiplicative reduction factor; and conduct the retransmission mechanism of the data packets to the traditional TCP; at this time, the window increases in the way of plusing 1.

7. The transmission method based on the learnable power model as mentioned in claim 6 wherein the value of β is 0.7.

8. The transmission method based on the learnable power model as mentioned in claim 2 wherein if the network packet loss or time-out issue occurs, the multiplicative reduction mechanism with β as the multiplicative factor shall be conducted, i.e. W.sub.i+1=β.Math.W.sub.i, in which β is the multiplicative reduction factor, and conduct the retransmission mechanism of the data packets to the traditional TCP; at this time, the window increases in the way of plusing 1.

9. The transmission method based on the learnable power model as mentioned in claim 3 wherein if the network packet loss or time-out issue occurs, the multiplicative reduction mechanism with β as the multiplicative factor shall be conducted, i.e. W.sub.i+1=β.Math.W.sub.i, in which β is the multiplicative reduction factor, and conduct the retransmission mechanism of the data packets to the traditional TCP; at this time, the window increases in the way of plusing 1.

10. The transmission method based on the learnable power model as mentioned in claim 4 wherein if the network packet loss or time-out issue occurs, the multiplicative reduction mechanism with β as the multiplicative factor shall be conducted, i.e. W.sub.i+1=β.Math.W.sub.i, in which β is the multiplicative reduction factor, and conduct the retransmission mechanism of the data packets to the traditional TCP; at this time, the window increases in the way of plusing 1.

11. The transmission method based on the learnable power model as mentioned in claim 5 wherein if the network packet loss or time-out issue occurs, the multiplicative reduction mechanism with β as the multiplicative factor shall be conducted, i.e. W.sub.i+1=β.Math.W.sub.i, in which β is the multiplicative reduction factor, and conduct the retransmission mechanism of the data packets to the traditional TCP; at this time, the window increases in the way of plusing 1.

Description

FIGURE EXPLANATION

[0025] FIG. 1 shows the flow Figure for the transmission method based on learnable power efficiency model.

[0026] FIG. 2 shows the function relationship Figure between network throughput capacity and the load.

[0027] FIG. 3 shows the function relationship Figure between data packet round trip time and the load.

[0028] FIG. 4 shows the unction relationship Figure between the network energy efficiency and the load.

[0029] FIG. 5 shows the relationship Figure between the time and the window in Hita potocol.

[0030] FIG. 6 shows the comparison Figure of average throughput capacity of Hita protocol and that of the reliable transmission protocol based on UDP.

[0031] FIG. 7 shows the comparison Figure of average throughput capacity of Hita protocol and that of the TCP protocol with better performance in HBDP network.

[0032] FIG. 8 shows the comparison Figure of actual throughput capacity of Hita protocol and that of the reliable transmission protocol based on UDP.

[0033] FIG. 9 shows the comparison Figure of real time end-to-end time delay of Hita protocol and that of the reliable transmission protocol based on UDP.

SPECIFIC EXECUTION METHOD

[0034] Further specific specification is given as follows by combining the attached Figures and the implementation case in detail. This invention offers a kind of transmission method based on learnable power efficiency model, which includes the steps as follows: [0035] Step 1: Record the round trip time vector of all ACK packets received {right arrow over (D.sub.i)} and the sending window vector {right arrow over (W.sub.i)}, in which i means the current time period. Adopt the formula D.sub.g,i=α.Math.D.sub.g,i−1+(1−α).Math.{right arrow over (D.sub.i)} to conduct weighting smooth processing on the round trip time vector {right arrow over (D.sub.i)}. Therein, i−1 refers to the last time period, a is the weighting smooth factor and D.sub.g,i is the round trip time of maximum delay of ith time period after smoothing; [0036] Step 2: Calculate the normalized time delay change rate G.sub.i in accordance with the formula

[00009]$G_i=\left(\frac{D_{g,i}-D_{g,i-1}}{D_{g,i}+D_{g,i-1}}\right).Math.\left(\frac{W_{g,i}+W_{g,i-1}}{W_{g,i}-W_{g,i-1}}\right),$

Therein, W.sub.g,i is the window size of the maximum time delay value of the ith time period; [0037] Step 3: Calculate the network power estimation value E.sub.e,i in accordance with the formula

[00010]$E_{e,i}=\frac{T_i^{\Delta }}{D_i},$

Δ refers to the sensitive factor of delay-throughput and T.sub.i refers to the actual throughput of the ith time period, which are gotten through accumulation of the confirmed send window vector {right arrow over (W.sub.i)} in each time period. D.sub.i refers to all elements in the round trip time vector {right arrow over (D.sub.i)} of the i time period; [0038] Step 4: Establish the corresponding relationship model between the corresponding network power capacity E of every data packet and the send window size W, which is defined as W(x+1)=f(E(x)+γ(x)). Therein, x is the time change parameter, i.e. time period i. W(x+1) is the congestion control window size of the next time period, f(x) is the window control function, E(x) is the estimation function of network power and γ(x) is the increase and decrease size of network power; [0039] Step 5: Judge the changing trend of network status. When G.sub.i>0 or

[00011]$\frac{D_{\max ,i}}{D_{\min }}>R,$

the network becomes congestion and the network power should correspondingly decrease in the next time period. When G.sub.i≦0 and

[00012]$\frac{D_{\max ,i}}{D_{\min }}>R,$

network becomes free and the network power should correspondingly increase in the next time period, which is

[00013]$E_{e,i+1}=\{\begin{array}{cc}{E}_{e,i}-{\gamma }_2& \mathrm{When}.Math..Math.\frac{D_{\max ,i}}{D_{\min }}>R\\ E_{e,i}-{\gamma }_1& \mathrm{When}.Math..Math.G_i>0\\ E_{e,i}+{\gamma }_2& \mathrm{Otherwise}\end{array}$

in detail; E.sub.e,i+1 is the network power estimation value of the i+1 time period, D.sub.max,i is the maximum element in the round trip time vector D.sub.i of i time period, D.sub.min is the minimum element in the round trip time vector {right arrow over (D.sub.i)} and R is the maximum value of the link tolerance; γ.sub.1 and γ.sub.2 refer to the increase and decrease value of the network power. Through a large quantity of repeated experiments, the optimum values are selected; [0040] Step 6: Calculate the W.sub.i+1 window value according to Step 4 and Step 5; [0041] Step 7: Divide every time period into n time slots. Every time slot is λ ms. After the i time period, the data sent within n−1 time slots is taken as a part of data to W.sub.i+1. For instance, first n−1 time slots in every time period are used for the learnable model and the last time slot is used to send the remaining data. So the size of the data which shall be sent of the last time slot of the W.sub.i+1 window is W.sub.fin,i+1 and

[00014]$W_{\mathrm{fin},i+1}=\max \left[0,\left(W_{i+1}-\frac{n-2}{n-1}.Math.W_i\right)\right].$

[0042] In Step 1, the value of α is 0.875. In addition, the model W(x+1) is learned and gotten by adoption of the linear regression algorithm in the machine learning algorithm. The input parameters are the corresponding network carrying capacity E of every data packet and the send window size W of the i time period.

[0043] In this invention, the value of γ.sub.1 is 2 ms and the value of γ.sub.2 is 3 ms. Every time slot length Δ is 5 ms.

[00015]$n=.Math.\frac{{\stackrel{.Math.}{D}}_{\mathrm{ave}}}{\lambda }.Math.,$

in which {right arrow over (D)}.sub.ave refers to the mean value of all elements in the round trip time vector {right arrow over (D.sub.1)}.

[0044] Further, if the network packet loss or time-out issue occurs, the multiplicative reduction mechanism with β as the multiplicative factor shall be conducted, i.e. W.sub.i+1=β.Math.W.sub.i, in which β is the multiplicative reduction factor. And conduct the retransmission mechanism of the data packets to the traditional TCP. At this time, the window increases in the way of plusing 1. For specific, the value of δ is 0.7.

[0045] Further specification is given for this invention through FIG. 2 to FIG. 9.

[0046] As shown in FIG. 2 and FIG. 3, when the load is relatively small, the increase of throughput capacity has linear relationship when being compared with the load and the delay increases slowly. When the load exceeds Knee point, the throughput capacity increases slowly and the delay increases quickly. When the load exceeds Cliff point, the throughput capacity decreases sharply and the delay increases sharply. From FIG. 3, we can see that when the load is near the Knee point, the use efficiency of the network is the highest. Congestion control means that the network nodes take measures to prevent the happening of congestion or response to the happening of congestion. In FIG. 2, the load is made to be near the Knee point. Congestion control mainly considers about the network environment between the end nodes, making the load do not exceed the transmission ability of the network. The flow control is implemented at the receiver side, making the sending rate of the sending end do not exceed the receiving ability of the receiving end.

[0047] As shown in FIG. 5, to reflect the change features of the channel quickly according to the information obtained through learning, Hita protocol subdivides the sending window W.sub.i into the smaller time slot of fixed length. After the i time period, the data sent within n−1 time slots serves as partial data of W.sub.i+1 as well. Size of the data which shall be sent in the last one time slot of W.sub.i+1 is W.sub.fin,i+1. The sending window value W.sub.i+1 is evaluated by using average RTT. Table 1 shows all parameters of the experimental test scenarios of SIIN satellite simulation platform of the Chinese Academy of Sciences, in which 10 kinds of satellite network topology structures are measured and the nodes include LEO, GEO satellite and the ground station.

TABLE-US-00001 TABLE 1 The Link Parameters for the Communication Scenarios of Different Satellites End-to-end Distance RTT average Measurement between value/Maximum Bit error 2BDP set in node node No. nodes(Km) value (ms) rate buffer (bytes) Node LEO-LEO Topology 1 7819.14 52 8.26E−07 1300000 types of Topology 2 6732.45 44 6.25E−07 1100000 satellites Topology 3 5984.51 38 4.71E−07 950000 that the Topology 4 Uplink: 5984.51 76 4.71E−07 1900000 transmission Downlink: 5984.51 4.71E−07 passing by Topology 5 5609.56 60 4.71E−07 1500000 GEO Topology 6 15165.73  100 3.87E−06 2500000 Topology 7 25914.00  170 9.74E−06 4250000 GND-LEO-GND Topology 8 Uplink: 3628.95 35 1.41E−07 875000 Downlink: 2005.36 2.68E−08 GND-LEO Topology 9 2319.14 25 2.68E−08 625000 LEO-GEO Topology 10 35100.82  250 1.85E−05 6250000

[0048] Table 2 shows the link models of different nodes tested through the simulation platform. The models in Table 2 are used to calculate the parameter information of the inter satellite links.

TABLE-US-00002 TABLE 2 Link Models of Different Types of Nodes Transmission node type LEO-LEO GEO-GEO LEO-GEO LEO-GND GEO-GND Working frequency 10 10 10 10 10 Transmitting power 20 20 20 20 20 Transmission gain 26 27 29 32 33 of transmitting antenna Transmission gain 54 54 54 54 54 of receiving antenna Receiver GT 18 18 18 18 18 Signal bandwidth 24 24 24 24 24 Assignment of bit −1 −1 — — — error rate Assignment of — — −1 −1 −1 uplink bit error rate Assignment of — — −1 −1 −1 downlink bit error rate Channel number “Convolutional K = 7 r = ½” Interlacing method “Matrix” “Matrix” “Matrix” “Matrix” “Random” Modulation method “DBPSK” “BPSK” “BPSK” “BPSK” “BPSK”

[0049] FIG. 6 shows the comparison Figure of average throughput capacity of Hita protocol and that of the reliable transmission protocol based on UDP, from which we can see that in most situations, the average throughput capacity of Hita protocol is higher than that of other protocols, whose highest increasing range reaches 5.94 times as that of QUIC+ protocol and reaches 2.03 times as that of UDT protocol. Combining the fairness index in Table 3 and the stability index in Table 4, although the throughput capacity of Hita protocol is lower than that of UDT protocol in some scenarios, it has good fairness and stability. The index value for fairness and stability of UDT protocol is far lower than that of Hita protocol. FIG. 7 shows the comparison Figure of average throughput capacity of Hita protocol with TCP protocol with better performance, from which we can see that the average throughput capacity of Hita protocol is far higher than that all kinds of variable protocols of TCP protocol, whose lowest increasing range is 1.17 times as that of other protocols and the highest one reaches 12.55 times.

TABLE-US-00003 TABLE 3 Comparison of Fairness Index of Hita Protocol and the Reliable Transmission Protocol Based on UDP Index value of Protocol name fairness (FI) Hita 0.99 UDT 0.71 Verus 0.98 QUIC+ 0.99 QUIC 0.99

TABLE-US-00004 TABLE 4 Comparison of Stability Index of Hita Protocol and the Reliable Transmission Protocol Based on UDP Index value of stability (SI) Topology Protocol name Tp-5 Tp-10 Hita 0.02 0.14 UDT 0.49 0.51 Verus 0.37 0.17 QUIC+ 0.36 0.16 QUIC 0.39 0.17

[0050] From FIG. 8, we can see that the actual throughput capacity of Hita protocol is far higher than that of other protocols. The occupies of Hita is about 80% of the link bandwidth while that of Verus occupies about 30% and that of UDP protocol occupies lower than 10%. From FIG. 9, it can be seen that Hita protocol can maintain the relatively low end-to-end time delay of about 10 ms.