Upload
f
View
213
Download
1
Embed Size (px)
Citation preview
Analysis of Priority-based Adaptive Routing in
Satellite Networks
Omer Kor9ak and Fatih Alag6zComputer Engineering Department
Bogazi9i UniversityIstanbul, Turkey
korcakom(i)cmpe.boun.edu.tr, alagoz(&,boun.edu.tr
Abstract- In this paper we analyze a priority-based shortestpath routing approach for mesh-like modeled satelliteconstellations. In a mesh-like satellite network, there are manyshortest paths between a source-destination (s-d) pair in terms ofhop-count. Therefore, deciding on the best shortest path becomesa crucial task for utilizing the network resources. In ill, weintroduced a Priority-based Adaptive Routing (PAR) scheme thataccounts for utilization metric of Inter Satellite Links (ISLs).Moreover, to avoid needless splitting of a flow and achieve betterutilization of ISLs, enhanced PAR (ePAR) scheme is proposed.However, there are a number of ePAR parameters that should beadjusted depending on the network and traffic characteristics. Inthis paper, we provide a detailed analysis of ePAR scheme toform an opinion on the parameters setting.
Keywords- satellite networks, adaptive routing, prioritymechanism
I. INTRODUCTIONSatellites are expected to widely appear in future
telecommunication systems, due to their extensive geographicreach, inherent multicast capabilities, etc. Researches onmultiple fronts have been under investigation for improvingboth the performance and capability of satellite networks.These researches include but not limited to integration ofsatellite and terrestrial networks, integrated satellitearchitectures, stand alone satellite systems, beam scheduling,on board signal regeneration, adaptive modulation and coding,multiple access, flow control, resource allocation, and otherservice and system specific attributes. In this paper, weconsider developing an efficient routing algorithm forachieving increased throughput, decreased delay andmaintained Quality of Service (QoS) in satellite networks.
There are several routing algorithms proposed for LowEarth Orbit (LEO) satellite constellations. [2] deals withadaptive routing with a limited set of alternative routes.However, there may be many shortest paths in a mesh-likenetwork which can be fully utilized. [3] proposes a DatagramRouting Protocol, where Inter Satellite Link (ISL)'s areconsidered to have variable length and each satellite decides onthe neighboring satellite to find the shortest delay path. In thisapproach, a satellite may change its decision in case ofexcessive queue length; however it is desired to avoidcongestion before it happens. [4] proposes a Maximum-Flow
Minimum-Residual routing algorithm which depends on thetraffic load. Main drawback of this algorithm is that a priorknowledge of the flows in the network is needed and it doesnot consider dynamic changes in the traffic flow. [5] proposesan Adaptive Flow Deviation algorithm which is also notsuitable for the dynamic traffic which could be caused due toinherent nature of Internet traffic, movement of satellites, anddifferentiation of day and night usage, etc. [6] proposes aProbabilistic Routing Protocol (PRP), which is based on a costmetric as a function of time and traffic load. UnfortunatelyPRP works only for the case where traffic load is location-homogeneous. In [1] we propose a novel priority-basedadaptive shortest path routing technique (PAR) thatdistributedly sets the shortest path through a destination andwhich is more suitable for dynamic traffic. Further we make anenhacement on PAR for better utilization of ISL's and proposeePAR. Relying on simulation results we show that theproposed techniques not only increase throughput, but alsodecrease delay. In this paper, we make an analysis of ePAR.Since there are number of parameters that should be adjustedproperly, our analysis provides an opinion on the setting ofthese parameters. In Section II, we present the backgroundabout the proposed algorithms. In Section III, we provide theanalysis of ePAR. In Section IV, we conclude this work.
II. BACKGROUNDIn a mesh-like satellite network, there are many shortest
paths between a source-destination (s-a) pair in terms of hop-count. At each satellite node, more than one outgoing linkcould be on one of the shortest paths. Decision on sending thedata from which of those links has an important effect on thedistribution of the traffic and utilization of ISLs. In [1], wepropose a Priority-based Adaptive Routing Algorithm (PAR),in which direction decision is made at each hop by a prioritymechanism, depending on the past utilization and bufferinginformation about the links. Determination of the prioritymetric is a critical issue that effects the performance of PAR.One possibility is to use the following metric:
A = na (1)
where na is the number of packets that arrive to the link.However, the congestion in a link is not only related with the
629
This work has been supported by "Bogazici University Research ProjectsAffairs", Project No: BAP.04 S 104.
0-7803-9206-X/05/$20.00 ©2005 IEEE
number of arrivals to it. For example among the links withsame number of arrivals, the link with shorter queue lengthmay be favored. Therefore in [1], the following priority metricis proposed:
/ = UrO ,+ /3. q + *nd (2)
where ur is the utilization ratio of the link, lq is the averagequeue length, and dd is the dropped data per second. Using thismetric, traffic tends to distribute the links in a more balancedway. Similarly, (2) can be changed as:
= ao nt + /3.q +c5-fnd
( 2ta) (2ta(7)
Equation (7) is for PAR. For ePAR, it can be rewritten as:
= ,U°sd 1 _ to + to) (8)
This implies that, nsd and Iq can be calculated in samemanner:
(3)
where n, is the successfully transmitted data per second inthe corresponding link. Note that, a, / and a are designparameters that should be adjusted properly due to the trafficrequirements and network topology.
It is important to note that most of the contentions occurbetween packets with different source-destination (s-d) pairs.Therefore it would be better to switch packets with same s-dpairs to the same outgoing link. This suggests that theperformance of PAR algorithm may be enhanced by using thefollowing metric:
flsd = a (n, -nsd )+. lq +S nd (4)
where nsd is the amount of transmitted packetscorresponding to s-d route, and tUsd iS the priority metric fortraffic traversing on s-d route. At the expense of increasedcomplexity on satellite nodes, better ISL utilization may beachieved in this technique, which we called enhanced PAR(ePAR). In this work, for the sake of simplicity, we assume thatbuffers are sufficiently large, hence we ignore nd:
Asd =oa(n, -ntsd)+3 lq (5)
Defining a new variable 'td= n, - n, , (5) can be furtherreduced to:
Psd a nsd + B lq (6)
Considering that the latest utilization and bufferinginformation is more important than the older ones, an agingmechanism is needed to be utilized while computing thepriority metric. One possibility is to take the average of last tseconds. However this mechanism has some drawbacks.Firstly, the information belonging to earlier times has also animportance, and completely ignoring them is not reasonable.Moreover, storing the information about last t seconds involveshigh space complexity. In [1], an aging mechanism is proposedas follows: We define an aging period with length ta. At thebeginning of each period, we store the current ,u value in avariable called ,u. Then satellite starts to collect utilization andbuffering information in a new variable called ,u. At to'th timeunit of a given period, ,u is calculated as follows:
0 (1 _to nn ( to
( 2t ) ( 2ta)1~~~q,=" I q
(9)
(10)
In the rest of the paper we make an analysis of the ePARalgorithm, utilizing the mentioned aging mechanism.
III. ANALYSISConsider that a, /3 and a in (4) denotes the design
parameters that should be adjusted properly depending on thenetwork characteristics. By the optimal selection of theseparameters, not only better load distribution can be achieved,but also traffic flows can be made more stable. By stability, wemean avoiding the needless fluctuations due to redirection ofall the flows in a congested link, simultaneously. In thissection, we aim to form an opinion about how to set theseparameters to achieve more stable systems.
Suppose a satellite network in which arrival rate of an s-dflow is Poisson distributed with mean 1/ bps. We represent aflow with source x, and destination y with fk. We consider ascenario, where a new flow is participated to a link that isalready utilized just below its capacity, and the data arrival rateexceeds its capacity by the participation of the new flow.
We assume that the existing aggregate flow (Fe) in aparticular link has a rate ofX/2 bps. A new flow (fid) withrate 1;d / 2 is participated to the corresponding link. After theaggregation of the existing flow Fe, and the new flow fd, thetotal arriving flow becomes (X + rsd ) / 2. Supposing that thisvalue is greater than the link bandwidth, B /2, total amount oftraffic to serve is B / 2 and total flow to be blocked per secondis (X + rd - B) / 2. Assuming that the blocking probability foreach flow is same, blocked portion of Fe will be:
bF = (X+lrd-B).-.I -)2 X+r~d (1 1)
And the transmitted portion of Fe will be:
630
B (xtF X+r
Asd/
Since tF is the total transmitted data per second, except theportion corresponding to s-d traffic, it is the new value that nsdwill converge. We represent this value as YJsd .
an aging period. Calculation of nij value for any other flow f,(12) can be done in same way as nsd. The only difference is that we
replace yV1d with i,tf and y4° with yi,, in (18) in Table 1.
TABLE I. CHANGES IN UTILIZATION AND BUFFERING INFORMATIONWITH THE TIME
/
B(X+rYV d -:-. t (13)
Before the participation to the corresponding link, nsd wasequal to the total transmitted data in the link, since n,d waszero. We represent the old value of nsd with i/°d:
oxVlsd -= (14)
For any other flowfL that uses the same link (and hence thatis a part of Fe):
B X+rSd -~tu. =X.AJ t X + rsd
(15)
where Yj is the new value that nu will converge. The initialvalue for nu is determined by the difference between totalinitial flow and the flow with source i and destinationj:
X-= r0iJ 2 (16)
Amount of blocked data continuously increases as timepasses, and t seconds after the participation of the new flow,
length of queue becomes t sdB that the
buffer is initially empty. Since we assume increment in queuelength is linear, average queue length between time t1 and t2 is:
ti +t2 X+rSd-B
2 A
For clear illustration, we represent increment in the queuelength (per second) with a new variable f:
X+r,d -Bsd (17)
Table I shows how nsd value and the queue lengthinformation changes after the participation of new s-d flow, f5dto the existing flow Fe in a particular link. The effect of agingmechanism is also considered. Let tj < ta where t0 denotes thelength of aging period. Without loss of generality, we considerthat the participation of new flow occurred at the beginning of
1nd =0sd
* lq
d= V'1;0 nsd = ,'d hhenceI (
X~~~~(Iq-..
sd = ysd nn = do ( ~~~~2ta ) 2ta )
0)f =O; g =2 *$ Xhence lq = 2.K* i*(2l
nd Vlsd 2+ Vfsd22 2yt~
I = ta .taq 22
n,s - n1;*sd = .d, hence
2 2
'nsd ='sd 2 '2 sdd2 22a
¢ 1° ta-4In = 2ta 2t 2,hence
( 2ta +to2 )2t
l
q 'q 2ec
4Y1sd 2+2YJs4 2 ta 22)
'q
C)
a)
V)
I<
0 1 (1- In,d = V°d *-+ Y1 d *' -
2md22lq (a~ +1+3t,a 2m +5ta, * 2m1 + - .+(2m-1)-t *-12
(t m 2i-1
.4I=I 2 )
We can find the sum of series that is included in (19) inTable 1:
631
0 1 ti I I tinsd = Vsd - - I - + Vfvd . 1-- -4 (18)2m 2tc, 2m 2m 2t,I/
t,, m 2i- I t ' K 2mt,, + t t)
Ilq -.E . 1-- + ' '- -. (19)4 i=, 2m-1 2t,, 'I \1 2 2t, )
m 2i-1 6T -. =-+4m-6.j= 2m-1 2m
Therefore, (19) is reduced to:
I t 6 4 }(,3+)) (20)
Figure 1. Illustration of(18).
Figure 1 illustrates, how the nsd value changes, as timepasses. A is the initial value for nsd, namely ys , and B is thevalue that nsd converges, namely VdI nsd (tq) stands for nsd valueat time ta.
Figure 2. Illustration of (20).
Figure 2 illustrates the increment in the 4q value with time.Iq (t0) stands for the lq value at time ta. Even though theincrement in the queue length is linear, the increment of 4q isnot linear because ofthe effect of aging mechanism.
Figure 1 and Figure 2 illustrates the case that no other newflow is participated and none of the flows is redirected toanother link. At some point, the corresponding link willbecome unfavored for the flows over that link, and redirectionswill occur. However, if all of the flows are redirectedcontinuously, it can lead to congestion in altemative link too.In that case, all of them will again redirected to this link, andthis will lead to needless fluctuations and disruption ofstability, which will decrease the performance of the system.To avoid this scenario, we can redirect the flows with smaller
data rates first, by adjusting the ePAR parameters. If 4q is toodominant for determining priority metric, we cannot achievethis. Difference between priority metrics for two flowsfu andfkion the corresponding link is determined by:
-ij =lkl a In,, -nklI (21)
Difference between nu and nkI values is the following:
= ~ ~t'-v'~ {-4+*.-J (22)|nyl YN<S12m (2t )| 2- 2 2ta )
For {ij}.{s,d} and {k,l}.{s,d} case, (22) is reduced to:
i -2-(X2ta)2 X+rSd 2m r 2ta
and it is equal to:
kj| , 2m (I2taJX+d( 2+ 2m 2ta )For sufficiently long aging periods, we can assume that
redirections occur in first aging period, hence m is equal tozero. Then (23) is reduced to:
li.i lk I 2A 1(( f2ta iJ+X + l;d 2t ) (24)
For the case that {kl}=ts,d}, (22) reduces to following:
In -n.d|=-| 1+01- .i. 1- +j A| 22 2ta) X + rd 2 2 2ta)
For m = 0 case it is:
Inii-nsdl=lt(-1- )X+|X+r-.-|2t,La2A X 'd 2(25)
As we mentioned before, 4q should not be too dominant indetermining priority metric. This suggests that difference occurbetween priority metrics for two flows, with reasonably distinctdata rates, should be near to the change occur in .4lq value in agiven reasonable time td.
Again assuming that m = 0, and initial 4q value is zero,difference occured in the 4q value in time td is the following:
lq (td )-lq (0) XXr B
4ta A(26)
Now, we can set:
632
B
----- ----
A~~~~~//?
A 1 ta 2t 4tTime
IqPta) ,,,;,1 {3Nv --- -- -- -- -- -- -- - - --I-- -- -- -- -- -- -- - - --I - -- -- -- -- -- ---- - - -- -- -- -- - ----
. ~/,q(2) .--
0--- -I ----------------
0 t 2t 3t 4tTime
a *In,, (td) - nkl (td )l = / (lq (td ) 1q (O)) (31)where fu andfki has reasonably distinct data rates, and td is
chosen appropriately. If td is too small, then system suffersfrom concurrent redirections. In the other case, if it is too large,then the effect of queue length will be decreased indetermination of the priority metric, and this can decrease theperformance of the system. If the data rate of Fe is much higherthan the data rate offd, we can ignore the case that {k,l}={s,d},and thus (27) can be rewirtten as:
R= tdl d
((2t a d ) +B +Br
If the s-d flow does not partitioned to various links, theexpected value for rsd is 1, and hence we may reduce (31) to thefollowing:
(R= t ) B(32)
__ X+rd-Btd sd
a 4ta
_8 r, r,((1 td
A ]( 2ta
A.) B tdX+ r 2taX+sd2t
(28)
which reduces to:
X = B -rSd and X = B are two extreme points for X, and areasonable Xvalue should be chosen between these two values.For example, setting it to the mean value between B - rSd and
B, that is B -_ d (or B - 0.5 ifwe set rS = 1 ), makes sense.2d
a I td *(X + rsd -B)2
rk-rr| I ((2t -td)+ -
If we set the reasonable rate difference |rj, - r,, to 0.5, (29)reduces to:
a 2t (d)+ r-B)
((a d)X + rd)
R~~ ~~~~~~~ /
ad Bx
Figure 3. Illustration of (30).
Figure 3 illustrates the reasonable a /a values fordifferent X values. Note that the y axis is equal to zero in thecase that the total flow does not exceed the link capacity afterthe participation of new flow, and thus queue length does notincrease (X = B -rSd). In the case that X = B, the link wasalready fully utilized before the participation of new flow. Atthat point, corresponding a / , is equal to R. From (30), thevalue ofR is found to be:
IV. CONCLUSIONIt is very crucial to best utilize all the appropriate paths
between two nodes in a satellite network. Due to the dynamicnature of satellite traffic, traffic sensitive routing approachescould be said to be more appropriate. Priority-based adaptiverouting is a novel example of such routing techniques, and it isa promising scheme for use in satellite networks. A prioritymetric, including utilization and buffering information, isdefined for PAR and for enhanced version of PAR (ePAR).The priority metric includes some parameters that should beadjusted for best performance. In this work, we make a detailedanalysis of ePAR to make sense about how to adjust theseparameters, in order to achieve more stable and high-performance systems. Although our analysis may seems to berestricted with large buffer size and long aging duration, it canbe easily extended for a wider domain. Unfortunately, due tothe simple characteristics of mesh-like satellite constellations,tailoring this analysis into real satellite constellations maybesomewhat complicated.
REFERENCES[1] 0. Korgak and F. Alagoz, "Priority-based Adaptive Shortest Path
Routing in IP over LEO Satellite Networks", 23rd ALAA InternationalCommunications Satelllite Systems Conference (ICSSC'05), Rome,Italy, Sept. 2005.
[2] 0. Korgak, I. Kaya, H. G. 14alikli, and F. Alagoz, "PerformanceEvaluation of Adaptive and Static Routing Algorithms and ContentionResolution Techniques in LEO Satellite Constellations", IEEE RecentAdvances in Space Technologies RAST, Istanbul, Turkey, June 2005.
[3] E. Ekici, I. F. Akyildiz, and M. D. Bender, "A Distributed RoutingAlgorithm for Datagram Traffic in LEO Satellite Networks,"IEEE/ACM Trans. on Networking, vol. 9, no. 2, pp.137-147, April 2001.
[4] R. Kucukates and C. Ersoy, "High Performance Routing in a LEOSatellite Network", Proceedings of the IEEE ISCC, pp. 1403-1408,Antalya, Turkey, 2003.
[5] I. Gragopoulos, E. Papapetrou, and F.-N. Pavlidou, "Performance Studyof Adaptive Routing Algorithms for LEO Satellite Constellations UnderSelf-Similar and Poisson Traffic", Space Communications, Vol. 16, pp.15-22, 2000.
[6] H. Uzunalioglu, "Probabilistic Routing Protocol for Low Earth OrbitSatellite Networks", Proceedings of IEEE Int. Conf. Comm. (ICC),Atlanta, GA, USA, June 1998.
633
(27)