Peilong Liu Hongyu Chen, Songjie Wei, Limin Li＊, Zhencai Zhu2,

1 Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 200050, China

2 University of Chinese Academy of Sciences, Beijing 100049, China

3 Innovative Academy for Microsatellites, Chinese Academy of Sciences, Shanghai 201203, China

4 School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China

5 College of Physics and Electronic Information Engineering, Wenzhou University, Wenzhou 325035, China

I. INTRODUCTION

Low Earth Orbit (LEO) satellite networks can achieve ubiquitous wireless coverage, which benefit the information access in areas and spaces that are uneconomic for terrestrial networks [1]. Moreover, it requires lower power for terminal devices and brings lower latency compared with geostationary orbit communication systems. Because of its irreplaceable advantages, LEO satellite networks, such as Iridium Next, LeoSat and Oneweb, are considered to be the essential complements for Next Generation Internet and Internet of Things(IoT) [2], [3].

While the use of inter-satellite links (ISLs)provides more flexibility and higher security,it raises the issue of routing in the space segment [4]. Different from traditional terrestrial Internet routing protocols, wireless mesh networks (WMN) routing protocols [5], [6] and hierarchical routing protocols of large-scale wireless sensor networks (WSN) [7]–[10], a specialized routing protocol for LEO satellite networks has to adapt to the dynamic topologies and link instability. In addition, routing complexity becomes more substantial as LEO satellites change their coverage areas and accordingly serve different amounts of devices.This ultimately results in a dynamically unbalanced and generally predictable traffic load over the entire constellation.

A number of load balancing routing schemes [4], [11]-[20] have been specifically proposed, which can be divided into two categories: centralized and distributed. To achieve the system optimum in traffic distribution, the centralized schemes generally adopt global optimization algorithms [14] (typically linear programming) for routing table calculation.However, it is difficult to adapt to the highly dynamic traffic requirements and may hard to be implemented in satellites due to its high complexity. Meanwhile, the distributed routing methods normally employ greedy shortest path algorithms that can improve robustness and speed of response by implemented onboard[12], at the expense of global performance.

To the best of our knowledge, the main load balancing method among distributed schemes is Distributed Traffic Detour (DTD): satellites make the decisions independently to detour a part of traffic from a default route to an alternative route. As shown in figure 1, when the link S4-S7 is considered to be almost congested, a decision is made by S4 to detourx% (0≤x≤ 100) traffic to link S4-S5. This DTD method allows satellites to provide real-time response to congestion by only considering the link states of local and neighbouring satellites.Nonetheless, for a group of satellites carrying high volume of traffic, DTD may lead to local optimum and result in cascading congestion[18], [19] due to its short traffic detour range.

Assuming the flow from S4 to S11 is distributed forwarded in the shortest path, DTD is enabled when a link is almost overloaded.However, detoured traffic may over flow other downstream links, such as S7-S8, S10-S13 and so on. To relieve cascading congestion,researchers have put forward various countermeasures by modifying DTD [4], [12]–[14],[20], the main principle is to collect local or neighbouring satellite’s residual buffering capacity and quantify detouring ratiox%. But DTD is still limited to neighboring area and thus easy to trigger cascading congestion [18],especially for satellites covering developed areas such as North America or Asia Pacific.

In this article, we propose a Hybrid-Traffic-Detour based Load Balancing Routing(HLBR) scheme for packet-switched LEO satellite networks, which utilizes Long-Distance traffic Detour (LTD) together with DTD to provide self-adaptive load balancing. As illustrated in figure 1, after the area with high volume of traffic is confirmed, LTD packets sent by S4 can take circuitous route toward S11 to achieve efficient traffic distribution and cascading congestion mitigation.

The remainder of this paper is organized as follows: section II presents the models and key concepts. Section III introduces the proposed HLBR scheme. Section IV evaluates HLBR by analytic formula. Section V presents simulation results to verify the performance gain in HLBR. Finally, section VI concludes the article.

II. SYSTEM MODEL

2.1 Constellation model

The HLBR is based on an Iridium-like Polar(Walker Star) [21] constellation that consists ofNsat=Norb?Nsposatellites, whereNorbis the number of orbits andNspodenotes the number of satellites per orbit. Each satellite can be connected with 4 neighbours at most by bi-directional ISLs (two interplane ISLs and two intra-plane ISLs). The ISLs in cross-seam and Polar Regions are switched off due to high dynamic motion, so the topologies can be represented as cylindrical Manhattan street network [22]. As shown in figure 2, satellites can also establish ground-satellite links (GSLs)with the devices located in their coverage. The transmitter of each ISL is allocated with a buffer queue.

Fig. 1. Hybrid-traffic-detour based method.

To hide the dynamics in topologies, the Virtual Node (VN) strategy is adopted, which can change the routing issue from moving physical satellites to stable virtual satellites[17]. Virtual satellites are supposed to cover distinct zones and have one-to-one correspondence with physical satellites. The relationship do not change until another physical satellite moves into the coverage of a virtual satellite while the former satellite leaves.

The identity of a satelliteViis given by the satellite’s orbit numberSo(1≤So≤Norb) and orbit phase numberSop(1≤Sop≤Nspo):

Fig. 2. Linking relationship of LEO satellite networks.

Fig. 3. Earth zone division and static device density index.

Assuming that the entire Earth’s surface is covered by logical zones which are embodied at any given time by a certain physical satellite, a graphG(V, E, W) as an abstraction of the constellation withNsatsatellites is constructed,whereV={Vi|1 ≤i≤Nsat} represents the set of satellites,E={Ei,j|1 ≤i≤Nsat, 1 ≤j≤Nsat,i≠j}represents the set of directed link from satelliteVitoVj,W={Wi,j|1 ≤i≤Nsat, 1 ≤j≤Nsat,i≠j}represents the weight of directed links.

2.2 Traffic model

As shown in figure 3, the Earth is divided into 22.5?×22.5?geographical zones, each zone is mapped with a satelliteViand static device density indexSDi. The statistics [23] from Organization for Economic Cooperation and Development (OECD) are utilized to estimate the density of IoT devices of each zone.

LetTRi,jrepresents the inter-satellite traffic requirement fromVitoVj, which is assumed to be proportional to theSDiandSDj[4]:

wheredi,jdenotes the distance between geographical zones that under the coverage of satelliteViandVjrespectively. In this paper,we setθ= 0.5 andφ= 1.5 [14].

traffic demands between source satelliteViand destination satelliteVjcan be calculated by equation (3) [14].

whereahrepresents the traffic percentage of time zone as illustrated in figure 4, andAdenotes total traffic in a day for whole networks in bits. Assuming the number of generated packets per second obeys Poisson distribution and the average data rate inViis

2.3 High-traffic-volume satellite block prediction

A group of neighbouring satellites carrying high volume of traffic are called High traffic Volume Area (HTVA). Since the HTVA is prone to cascading congestion, the perception of HTVA is necessary for efficient traffic distribution with global view. Given the predictable characteristic of topologies and terminal distribution, the traffic of LEO satellite networks can be decomposed into a predictable baseline and unpredictable fluctuations [14],so the prediction of HTVA is feasible and can benefit from geographical distribution of terminal devices. Set activated devices densityADiunder the coverage ofViis:

whereandare corresponding maximum value ofahandSDi. Figure 5 illustrated the unbalanced distribution of zones with top 32%ADiin Greenwich Mean Time(GMT) 10:00 and 00:00.

Since the reason ability of HTVA estimation is related with historical network status (e.g.queueing delay) and predictable activated devices density, and the aggregation degree of busy satellites can also denote the risk of cascading congestion, so High-traffic-Volume Satellite Block Prediction (HSBP) indexhifor each satelliteViis designed:

wherelhi,ahidenote local part and adjacent part of HSBP index, respectively.ECidenotes the estimated condition ofViand will be introduced in section III,adj(i) represents the adjacent satellites ofVi, a satellite with higherahiis prone to trigger and suffers from traffic detouring.hidenotes the probability ofVifor being a part of high traffic volume area by considering the local state together with the state of adjacent satellites. Because of the time-varying user activities, changing topologies and uncertain traffic fluctuations,hineeds to be dynamically adjusted according to the real-time network status.

III. PROPOSED HLBR SCHEME

The scheme is performed periodically as shown in figure 6. The routing cycle is ?TH,which can be divided into three parts: Network Status Collection (NSC), Circuitous Multipath Calculation (CMC), Forwarding and Hybrid-Traffic-Detour (FH). The corresponding period of time areTNSC,TCMCandTFH. Here a concept of Generation is raised, packets of Generationiare the ones that generated inith routing cycle.

3.1 Network status collection

NSC is first performed in each routing cycle to achieve network awareness. HSBP indexhiis needed to be collected for efficient traffic detour. Furthermore, queuing delay and propa-gation delay are vital parameters that influence the quality of service, so they also need to be collected. Set the link weightWi,jas:

Fig. 4. Daily variation of traffic volume.

Fig. 5. Activated devices density distribution in GMT 10:00 (left) and 00:00(right).

Fig. 6. Flow chart of HLBR scheme.

wheredenotes the expected queuing delay of directed linkEi,j, anddenotes the propagation delay for ISL from satelliteVito satelliteVj:

whereclis the speed of light, andDi,jdenotes the physical distance between satellites,which is assumed to be constant within each routing cycle to reduce computational burden.

The orbit speaker strategy [12] is adopted to limit the flooding area of the link state advertisement. There is a speaker satellite in each orbit. Speakers of adjacent orbits are connected by direct ISL. The strategy consists of three steps: firstly, non-speaker satellites send their HSBP index and link weight to the neighbouring intra-plane satellites which have the minimum hop distance to the intra-plane speaker. Secondly, each speaker generates orbit information packet (OIP) and delivers it to intra-plane satellites and adjacent speakers.Finally, when an OIP arrived at a speakerVi,it can be broadcasted toVi’s intra-orbit satellites and another downstream neighboring speaker. As a result, all satellites can get the latest global state information and establish a consensus ofG(V,E,W) and HSBP index distribution.

The packets relaying link weight and HSBP index are free from queuing delay. Each orbit speaker needsNspo/2 or (Nspo? 1)/2 hops at most to collect orbit information or broadcast OIP in its local orbit. OIP transmission between speakers needsNorbhops at most for pairs of speakers with maximum hop distance.Therefore, the lower limit ofTNSCis:

wheredenotes maximum propagation delay of neighbouring satellites. For a constellation consists of 8× 16 satellites with 800km orbit altitude,is 264ms. Taking unexpected events such as satellite failure into account,TNSCshould not be less than.

3.2 Circuitous multipath calculation

After finished updatingG(V,E,W) and HSBP index, CMC is executed. For each pair of satellites, a best hop of shortest path is prepared for distributed forwarding. Meanwhile, a LTD path is calculated for taking circuitous route to relieve cascading congestion.

AssumingViandVjare connected with direct ISL,Wi,jis needed to be modified to avoid predicted cascading congestion according to HSBP index:

whereAW(Ei,j) denotes the additional weight function of ISLs.

Pseudocode of CMC is presented in algorithm 1. After CMC, LTD paths and routing table of the shortest path are synchronized updated for transient loops avoidance.

The complexity of LTD path calculation is determined by its forward method. Distributed forward method demands an independent routing table for each pair of satellites, and the corresponding time complexity is.When source routing method is adopted for LTD,the time complexity can be decreased to, and the average storage requirement of LTD routing table is also reduced fromwhereSsatrepresents the storage occupied by a satellite ID, andHaveragerepresents the average hop count of LTD path between a pair of satellites. Furthermore, time complexity of routing table searching isfor distributed LTD, meanwhileO(1) for source routing.For a constellation of 128 satellites equipped with 300MHz FPGA,TCMCis around 40ms.

3.3 Forwarding and hybrid-trafficdetour

1) Practical Distributed Routing protocol:To achieve simplified forwarding and fast reaction to isolated congestion, a Practical Distributed Routing (PDR) protocol including DTD is proposed as a default routing method in HLBR scheme. PDR is parallelizable and requires less onboard resources compared with classical distributed load balancing routing method [4], [12], [13].

PDR enables satellites to provide a best hop and a DTD hop for each PDR packet, which represents the shortest path and the LTD path precalculated in CMC. The selection of next hop depends on the Next Hop State (NHS),when a direction is suitable for being next hop, the corresponding NHS is “Suggested”,otherwise is “Not Suggested’. For congestion avoidance, the Local Queue State (LQS) and Neighbouring Satellite State (NSS) are combined to determine the NHS.

Generally the state of ISL buffer is characterized by the Queue Occupancy Rate (QOR).When QOR is lower than a preset threshold αq(0% ＜αq＜ 1 00%), LQS is set to “Normal State” (NS), otherwise is set to “Busy State”(BS). BS signifies the incoming congestion and overflow. The Checking of QOR is periodically repeated in cycle length ofTQOR.Determination ofTQORand αqinVishould comply equation (12):

whereSBQdenotes the size of a buffer queue,is the propagation delay between neighbouring satelliteViandVj, (In?On)maxrepresents maximum difference of the output and input speed in directionnforVi:

whereCis the link capacity of ISL andNISLdenotes ISL number ofVi. Local buffer monitoring can avoid buffer over flow by choosing another less congested ISL. In addition, it is important to make sure that whether the best hop or DTD hop of neighbouring satellite is in BS state at a same time. Therefore, “Satellite Normal State” (SNS) and “Satellite Busy State” (SBS) are defined. SetVjas the nexthop satellite ofVi,

Algorithm 1. Pseudocode of CMC

wherePba(Vj) denotes the probability that the best hop and DTD hop inVjare all in BS, forNISL≤4:

whereNBSandNNSrepresent the number of ISLs which LQS is in BS and NS. When the value ofNSS(Vj) changes, it is spread toVj’s neighbouring satellites to help making the forward decision.

Moreover, HLBR uses NSS to indicate the traffic conditions in HSBP index adjustments.The NSS ofViin past Generations and the latest Generation τ are combined to perform condition estimation in Generation τ+1:

where therepresents theECiin Generation τ.denotes the proportion of time forVibeing in SBS in Generation τ. αdis a decay factor ranges in [0,1]. It is an empirical value that can be interpreted as an indicator for the weight of long-term traffic pattern and short range traffic fluctuation in the condition estimation.

A combination of NSS and LQS is used to determine the state of optional hops, determination rules are listed in Table 1, and the hop selection method is shown in Table 2.

2) Long-Distance traffic Detour:Networks with traditional DTD method are prone to cascading congestion in HTVA. Due to HSBP scheme and circuitous multipath, LTD path can bypass potential HTVA to avoid trapping in local optimum.

Table I. Rules for the setting of hop state.

Table II. Hop selection rules for PDR.

Table III. Symbols for analysis.

However, due to the extra hops caused by circuity, LTD path is not the best choice for networks without severe congestion, especially for latency-sensitive applications. So a LTD-Shift-Trigger (LST) is proposed to monitor congestion for PDR packets and trigger LTD when necessary. The LST is composed of two parts: surveillant module and shift-trigger module.

For surveillant module in satelliteVj,end-to-end delayPe2eand hop countPhopof PDR packets that sent toVjare concerned. A congestion binary functionfcon(Vi) forVjis defined, iffcon(Vi) is 1, the congestion of the flow fromVitoVjis judged:

whereINe2edenotes end-to-end delay binary index andINhopis hop count binary index:

PKrepresents an incoming packet fromVitoVj, which is the argument of hop count threshold function(PK) and average hop delay threshold function(PK). ThePdis(PK) represents the minimum hop distance betweenViandVj.

where αdisis hop threshold and αhdrepresents the average hop delay threshold. Range ofαdisis:

the upper limitNsat/Pdis(PK) denotes a packet passing through all the satellites in the network, lower limit indicates that the packet arrives at its destination by means of a minimum hop path. The range of αhdis:

whereSrepresents queue size andrepresents the average one-hop propagation delay. Upper limit indicates that the packet has endured maximum delay in each hop.

After the congestion is observed, a notification packet (NFP) is created and send back by means of shortest hop count path without suffering queueing delay, which benefits from its small size.

The maximum cycle length of NFPis:

wheredenotes maximum hop count of shortest hop count paths:

For a constellation consisting 8× 16 satellites with 800km orbit altitude,is 165ms,which is unnegligible.

As long as satelliteVireceived NFP created byVjof ongoing Generation, LTD fromVitoVjis activated, corresponding LTD binary triggerLTD(Vj) inViis set to 1.γ%(0 ≤ γ ≤ 100) of flow fromVitoVjwill be send by LTD in the current Generation. The packet type and hop sequence of LTD path are written in the packet header, after arriving at a relaying satellite, LTD packet is forwarded according to its header. When a Generation ends,LTD(Vj) inVireturns to 0.

Considering the timeliness of NFP, a shift-trigger module is set:

LTDm,m∈[1,αltd] is a shift register that keeps latest αltdGeneration ofLTD(Vj), it is updated when a Generation ends. Thefmt(Vj)denotes shift-trigger function. As a result, the final LTD triggerTRltdfor the flow towardVjis:

whenTRltd=1, LTD of current Generation is triggered by default. Shift-trigger module can compensate the propagation delay of NFP and promote the speed of congestion response especially for prolonged congestion events.

IV. ANALYSIS

To evaluate the performance of LTD, a simplified network model similar with Ref. [12] is illustrated in figure 7. Assuming a traffic flow fromV1toV9is concerned, end-to-end delay,packet drop rate of PDR and HLBR, including PDR and LTD, are compared.

The analysis focuses on the routing performance within a short period, so we can assume that the network status, best hop and DTD hop are fixed. Locate satellitesV3,V5,V6in a HTVA with higher congestion probability.Symbols are defined in Table 3, corresponding superscriptsnandhdenote whether a satellite is located in normal area or HTVA.Since a symbol can be related to at most three satellites (local satellite, best hop satellite and DTD hop satellite), the superscripts can also be configured as “n, h”, “n, n, h”, “n, h, n” and so on.

The probability of LQS varies in different kinds of satellites:

where χ%(0 ＜ χ ＜ 100) determines the difference of BS probability in HTVA and normal area. The relationship of the symbols is presented in Table 4, and figure 8 shows all available paths.

Fig. 7. Simplified model for analysis.

Fig. 8. Available paths.

Table IV. Relationship of the symbols.

Path selection probabilityPkand Path selection probability without dropPkwdare first determined:

Hq,k(x) is the hop function that can return the value of parameterxin theqth hop of the pathk. There are 4 hops for each available path,soqranges from 1 to 4. The end-to-end delay and the packet drop rate ofkth path is:

As a result, average end-to-end delayand average packet drop ratefor PDR are:

Assuming the LTD is activated, path 6 is the only choice that bypass HTVA. As a complement for DTD, Setγ% of flow is detoured by LTD method, probability of path selectionPkwdis change to:

In this case, average end-to-end delayand packet drop rateafter enabling HLBR (PDR+LTD) are:

To perform comprehensive analysis, path 1-3 that passing through HTVA are selected as the shortest path. The network status for path 3 being the shortest path is illustrated in figure 7.Based on parameters listed in Table 5,,are calculated and compared in figure 9(a) and figure 9(b). HLBR outperforms PDR in both packet drop rate and end-to-end delay, the advantages can gradually increase with the satellites in HTVA becoming more congested.

V. SIMULATION

5.1 Simulation setup

This section evaluates the performances of the HLBR by using OPNET Modeler. A constellation similar with Iridium is constructed in the form of Walker Star 86.4°:128,8,4, the satellites are evenly and uniformly distributed over eight orbits with 800km altitude. The crossseam ISLs are switched off due to the high dynamic motion in opposite directions, capacity of ISLs are equal to 260Mbps and the size of Drop-Tail based buffer queue is 19.3Mbit. All links are assumed to be free of bit errors. Other parameters are listed in Table 6. The simulations are run for 600s under the total traffic from 3.13Tbit/10min to 4.60Tbit/10min, and the performances among HLBR, TLR [12] and ELB [13] are compared.

5.2 Simulation result

1) Packet Drop Rate and Total Throughput:The average packet drop rate and total throughput are given in figure 10(a) and figure 10(b). As shown in figure 10(a), HLBR achieves the lowest average packet drop rate compared with TLR and ELB. For sending rate of 4.60Tbit/10min, the packet drop rates of HLBR, TLR and ELB are 3.17%, 6.42%and 14.09%, respectively. The underlying reason lies in the abilities of HLBR that the LTD path can bypass predicted HTVA to alleviate cascading congestion. TLR performs better than ELB due to its local DTD and the consideration of next hop satellite’s status, which can further reduce the probability of packet drop in the next hop satellite. ELB performs upstream DTD which may reduce reaction speed and result in a higher packet drop rate.

The validity of HLBR can also be evaluated by total throughput, which is shown in figure 10(b). The advantages of HLBR owe to the hybrid detour method that can achieve more efficient traffic distribution.

2) Queue Occupancy Rate:The status of buffer queue is recorded during the simulation. Figure 10(c) illustrates average QOR for networks under 4.41Tbit/10min sending rate.It can be seen that the average QOR of satellites with HLBR is lower than TLR and ELB.Maximum average queue occupancy rate for HLBR, TLR and ELB is 29.27%, 35.01% and 37.50%, respectively, which is mainly benefits from the load balancing advantage of hybrid traffic detour method. Satellites with lower QOR may reduce packet drop rate and protect system from severe congestions.

3) Cascading Congestion:The cascading congestion can be verified through traffic distribution. Figure 10(d) shows the real-time geographical distribution of satellites for GMT 10:10 and GMT 18:10 under 4.24Tbit/10min.Grey block represents the satellites with instantaneous QOR exceeding 25%, and the aggregation of which can be regarded as a HTVA. HLBR can achieve a smaller HTVA,so it can be deduced that the cascading congestion of HLBR is relieved.

Fig. 9. Analysis results: (a) average end-to-end delay, (b) average packet drop rate.

Table V. Parameters for analysis.

Table VI. Parameters for simulation.

Since the cascading congestion may increase packet drop rate, the dropped traffic because of DTD is recorded during the simulation, and shown in figure 10(e). HLBR outperforms ELB and TLR, which shows the effect in cascading congestion mitigation.

4) Average End-to-End Delay:The total delay of packets that arrived at their destination is recorded. Average end-to-end delay is shown in figure 10(f), HLBR performs better than ELB until sending rate exceeds 4.24Tbit/10min. Along with longer transmission distance, the more hops packets may experience, the higher drop probability it will be, especially for packets that traverse HTVA.HLBR can reduce the packet drop rate by using efficient LTD with global view. However,it may prolong the average end-to-end delay by allowing more long-distance packets to arrive at their destinations. Simulation Result shows that under the traffic of 4.60Tbit/10min,HLBR achieves 51.5% and 77.46% reduction in packet drop rate compared with TLR and ELB, at the cost of 3.45ms and 7.10ms average end-to-end delay, respectively. The primary goal of HLBR using LTD method is to mitigate cascading congestion and achieve efficient traffic distribution, while keeping endto-end delay at a reasonable and acceptable level.

5) Traffic Distribution:Load balancing routing protocols can achieve better traffic distribution than pure Single Source Shortest Path (SSSP) protocols. Researchers adopt Distribution Index (DI) [12], [14] to quantify the traffic distribution capability. Since the goal of load balancing is to avoid congestion and thus increase the total throughput, the DI and corresponding total throughput should be comprehensively considered. Therefore, a Distribution Effectiveness Index (DEI) is devised:

whereTHRrepresents total throughput,xidenotes the number of packets that traversed throughith ISL. Higher DEI value represents a more efficient traffic distribution. Figure 10(g)shows the results of DEI, which indicates that the HLBR outperforms TLR and ELB under simulated sending rates. This result proves the fact that the traffic distribution of HLBR is more efficient and leads to greater total throughput gain, especially under high data sending rate.

6) Total Flow Cost:Minimization of Total Flow Cost (TFC) [14] can be the objective function for centralized routing optimization:

wheredenotes the flow fromVmtoVnon linkEi,j. Figure 10(h) shows the TFC which is attracted from the simulation. Considering the queueing delay and prorogation delay as the link cost, for simulated scenarios,HLBR can move network equilibrium closer to the system optimum without centralized optimization. The better global performance mainly owe to the global view of LTD method.

VI. CONCLUSION

In this article, we propose a HLBR scheme for packet-switched LEO satellite networks.A hybrid traffic detour method which combining LTD and DTD is introduced to provide self-adaptive load balancing. HSBP scheme is first raised based on prior geographical information and real-time network status to predict areas which are prone to cascading congestion. Then, LTD path for cascading congestion mitigation is acquired by CMC and activated through LST according to the HSBP index and real-time network status. Simulation results show that HLBR works much better in packet drop rate, total throughput, average queue occupancy rate, effectiveness of traffic distribution and total flow cost, compared with ELB and TLR.

ACKNOWLEDGEMENTS

This work is supported by the National Science Foundation of China (No. 61472189),Zhejiang Provincial Natural Science Foundation of China (No. LY18F030015), and Wenzhou Public Welfare Science and Technology Project of China (No. G20150015).

Fig. 10. Simulation Results: (a) packet drop rate, (b) total throughput, (c) average queue occupancy rate, (d) real-time distribution of high QOR satellites, (e) dropped traffic due to cascading congestion, (f) average end-to-end delay, (g) distribution effectiveness index, (h) total flow cost.

[1] B. G. Evans, P. T. Thompson, G. E. Corazza, A.Vanelli-Coralli, and E. A. Candreva, “1945-2010:65 years of satellite history from early visions to latest missions,”Proceedings of the IEEE, vol. 99,no. 11, 2011, pp. 1840–1857.

[2] Z. Qu, G. Zhang, H. Cao, and J. Xie, “LEO Satellite Constellation for Internet of Things,”IEEE ACCESS, vol. 5, 2017, pp. 18391–18401.

[3] Y. Kawamoto, H. Nishiyama, Z. M. Fadlullah,and N. Kato, “Effective data collection via satellite-routed sensor system (SRSS) to realize global-scaled Internet of Things,”IEEE Sensors Journal, vol. 13, no. 10, 2013, pp. 3645–3654.

[4] O. Korcak, F. Alagoz, and A. Jamalipour, “Prior-ity-based adaptive routing in NGEO satellite networks,”International Journal of Communication Systems, vol. 20, no. 3, 2007, pp. 313–333.

[5] I. W. Lai, N. Funabiki, S. Tajima, M. S. A. Mamun,and S. Fujita, “Cross-layer selective routing for cost and delay minimization in IEEE 802.11ac wireless mesh network,”Wireless Networks, vol.23, no. 9, 2017, pp. 1–13.

[6] Y. Chai, W. Shi, T. Shi, and X. Yang, “An efficient cooperative hybrid routing protocol for hybrid wireless mesh networks,”Wireless Networks, ,vol. 23, no. 5, 2017, pp. 1387-1399.

[7] S. K. Singh, P. Kumar, and J. P. Singh, “A survey on successors of LEACH protocol,”IEEE ACCESS,vol. 5, 2017, pp. 4298–4328.

[8] S. Sharma, D. Puthal, S. Tazeen, M. Prasad, and A. Y. Zomaya, “MSGR: A Mode-Switched Grid-Based Sustainable Routing Protocol for Wireless Sensor Networks,”IEEE ACCESS, vol. 5, 2017, pp.19864–19875.

[9] Y. Xu, S. Wu, R. Tan, Z. Chen, M. Zha, and T. Tsou,“Architecture and routing protocols for smart wireless home sensor networks,”International Journal of Distributed Sensor Networks, 2013.

[10] X. Ding, M. Ling, Z. Wang, and F. Song, “DKLEACH: An optimized cluster structure routing method based on LEACH in wireless sensor networks,”Wireless Personal Communications, vol.96, no. 4, 2017, pp. 1–11.

[11] WANG, Houtian, ZHANG, Xiangjun, Ying, and Naijin, “Cross-layer design and ant-colony optimization based routing algorithm for low earth orbit satellite networks,”China Communications, vol. 10, no. 10, 2013, pp. 37–46.

[12] G. Song, M. Chao, B. Yang, and Y. Zheng, “TLR:A traffic-lightbased intelligent routing strategy for NGEO satellite IP networks,”IEEE Transactions on Wireless Communications, vol. 13, no.6, 2014, pp. 3380– 3393.

[13] T. Taleb, D. Mashimo, A. Jamalipour, and N.Kato, “Explicit load balancing technique for NGEO satellite IP networks with on-board processing capabilities,”Networking IEEE/ACM Transactions on Networking, vol. 17, no. 1, 2009,pp. 281–293.

[14] Z. L. Liu, J. S. Li, Y. R. Wang, X. Li, and S. Z. Chen,“HGL: A hybrid global-local load balancing routing scheme for the Internet of Things through satellite networks,”International Journal of Distributed Sensor Networks, vol. 13, no. 3, 2017.

[15] J. Bai, X. Lu, Z. Lu, and W. Peng, “Compact explicit multi-path routing for LEO satellite networks,”2005 Workshop on High Performance Switching and Routing, 2005, pp. 386–390.

[16] H. Nishiyama, D. Kudoh, N. Kato, and N. Kadowaki, “Load balancing and QoS provisioning based on congestion prediction for GEO/LEO hybrid satellite networks,”Proceedings of the IEEE, vol. 99, no. 11, 2011, pp. 1998–2007.

[17] H. Yan, Q. Zhang, and Y. Sun, “A novel routing scheme for LEO satellite networks based on Link State Routing,”2014 IEEE 17th International Conference on Computational Science and Engineering (CSE), 2014, pp. 876–880.

[18] G. Hu, Z. Wu, F. Jin, B. Yang, Y. Song, and Y. Fu,“Tunneling-based multi-path routing mechanism in packet-switched non-geostationary satellite networks,”International Conference on Algorithms and Architectures for Parallel Processing, 2015, pp. 729–742.

[19] J. Zheng, Z. Gao, B. Fu, and F. Li, “Cascading failures in congested complex networks with feedback,”Chinese Physics B, vol. 18, no. 11, 2009,pp. 4754–4759.

[20] L. Liu, T. Zhang, and Y. Lu, “A novel adaptive routing algorithm for delay-sensitive service in multihop LEO satellite network,”KSII Transactions on Internet and Information Systems, vol.10, no. 8, 2016, pp. 3551–3567.

[21] J. G. Walker, “Some circular orbit patterns providing continuous whole earth coverage,”Journal of the British Interplanetary Society, vol. 24,1970, pp. 369–384.

[22] L. Wood, “Internet working with satellite constellations,” Ph.D. dissertation, University of Surrey,2001.

[23] Organisation for Economic Cooperation and Development, OECD Digital Economy Outlook 2015. OECD Publishing, Paris, 2015.