Professional Documents
Culture Documents
A New Packet-Loss Minimization Routing Algorithm For ATM High-Speed Data Networks
A New Packet-Loss Minimization Routing Algorithm For ATM High-Speed Data Networks
A New Packet-Loss Minimization Routing Algorithm For ATM High-Speed Data Networks
N- M
(5)
where UU,,(p) is the set of up-stream links of link ( U , TI)
Figure 1: High-speed packet switch on path p .
A link cost function is defined to be the packet loss
Let Cu,udenote the total capacity of link ( U , w) and rate on a link. Let us denote average network packet
Fu,udenote input flow on an out-going link ( U , U). Packet loss probability by B ( F ) where F = {Fu,u},( u , u ) E L.
loss probability on link ( U , w), denoted by PRu,u(Fu,u), The average network packet loss probability is used as an
is given by overall cost function and is defined by the sum of link
packet loss rates divided by the total traffic demands
over the network:
1
B(F) - &,u(FU,u) (6)
where pu,w = Fu,w/Cu,u is the utilization factor oflink ( U , w) (U,U)€L
It is noted that PRu,w(Fu,u) is a function of Fu,u (also
pU,,,). Packet loss rate on link ( U , U ) , denoted by Bu,u(Fu,+,),where y = CwEw r,.
is given by The link-formulated ORP for high-speed networks is
to find a set of path { x , , ~ ) that minimize the cost func-
&,u(Fu,w) = F u , u . f'fL,u(Fu,u) tion (6) satisfying traffic balance equation ( 5 ) and flow
constraints which are given by (3) and (4). By eliminat-
ing constraint ( 5 ) , the above link-formulated QRP can
288
be transformed into the path-formulated ORP:
whereGisdefinedase= ~ l x G 2 x . . . G w . . . X G . M , w i t h
6, being a simplex defined by x , , ~ 2 0 , V p E P w , p #
pw,Vw E W . pw is the shortest path for commodity w. (b)
Since the ORP represented by (7) is not convex, gen- Figure 2: (a) Input link flows in ORP. (b) Input link
flows in R-ORP
eral gradient projection methods (GPM) can not be di-
rectly used to solve (7). In this paper, we propose a In R-ORP, for each link, it is assumed that no pack-
two-stage algorithm to find the optimal solut,ion of (7). ets are lost at upstream links. Therefore, the input link
flows of R-ORP will be different to the actual input link
flows of ORP. Figure 2 shows an example of differences of
5 Our Two-Stage Algorithm
Our two-stage algorithm is intended to solve the non-
a
input link flows on a single path oing through node j ,
k, 1, and so on, with node i as t e origin. We should
mention that since PRk,l(Fk,l) is usually very small for
convex ORP introduced in the Previous section. The general network conditions, ignoring PRk,l(Fk,l) in ( 5 )
idea of the algorithm is as follows: In the first stage will not incur significant of the solution
of our algorithm, we transfer the original non-convex of R-ORP. The of R-ORP should be satisfac-
ORP into the reduced convex ORP (R-oRP)i which is tory for most network routin problems. However, if
discussed in the next sub-sectionl and then, solve the R-
ORP. The solution of the R-ORP is usually very close to
3
more precise solutions are nee ed, we can refine the R-
ORP solutions by the second sta e solution process in
the true optimal solution. In the second sta e, we solve
the original ORP with the solution of R-8RP as the
d
our twestage algorithm as intra uced in the following
sub-section.
initial guess. The a1 orithm naturally avoids the local
minimum traps cause3 by the non-convexity of the origi-
nal ORP since the solution of R-ORP is around the true 5.2 Solving the ORP with Our Two-Stage
optimal. Test results show that our two-stage algorithm Algorithm
works well for most high-speed networks.
To get more recise o timal solutions, we propose a two-
sta e algoritim. At t i e first stage, we solve the R-ORP
5.1 Reduced Optimal Routing Problem (R- pro%lem using our GPM. Then the obtained solution is
ORP) used as the initial guess of the second stage problem,
which considers the detailed loss formulation. The in-
In the ORP mentioned in the previous section, path tuition behind the approach is that the dominant factor
flows are correlated with each other. Therefore, packet that decides the optimal routes should be the conven-
losses on a path affect other path flows, causing the ob- tional routing problem, the loss problem should be a
jective function (average network packet loss probabil- secondary issue. The difference between the routing so-
ity) to be non-convex with respect to its variables (path lutions in R-ORP and ORP depends on the differences of
flows). The idea of R-ORP in high-speed networks is to input link flows. If the differences are reasonably small
get rid of this correlation by ignoring the effects of any as shown in 3, ORP can be Obtained at the
packet losses at intermediate switching nodes. from the second stage with relatively minor efforts. Also, the al-
traffic balance equation (5) for a generic link (U, .), the gorithm avoids many local minimum traps caused by the
terms P R k , l ( F k J ) , ('7') E 'U,V(P) due to packet losses
non-convexity of the ORP problem since hopefully the
solution of the first stage is around the true optimal.
at up-stream links of link ( u , ~ are ) eliminated. Then
input flow Fu,,,on link (U,U ) is now expressed by The area in which the global
minimum can be found by GPM
+
to find the new values of 2 ( k l), which denotes the Fchp (kS.1) = z w , p (k+l).(
(i,WJ”,”(P)
l-PRi,i( F{,i (kS1)))
path flows of the startin6 links of all the paths. We need
to find the flows of the intermediate links of each path
since the flow is not preserved at each intermediate link.
In our algorithm, all the path flows (and thus, all the
link flows alon different paths) need to be adjusted in
’i
parallel. Para1 el flow adjustment is difficult since, as
mentioned at the beginnin of sub-section 5.1, there is +
In ( l o ) , F/,i(IC 1) is computed by
5
a correlation among all lin (path) flows, i.e., when a
link flow, F,,,, changes (due to some path flow adjust- F/,l(k + 1) = F$”’(k) + FGjp(IC+ 1) (12)
ment), all the global link flows may be influenced. This W’EW,(bI)€P’
is due to the fact that when a link flow changes, the link tW ’ , P ’ I # tw ,PI
flows along all the paths that share the link will also
need to be adjusted because of the change of the packet In the first step, (10) is calculated and in the second
loss probability on the link. Then, more link flows are step, (11) is calculated.
influenced and need to be adjusted, and so on. Fi . 4 We should mention here that even though, in ( l o ) ,
shows an example of a correlation among all the Enk
flows. When the flow on link (1,2) changes, the flow on we use some of the flow values, F$”’(k),of the (IC)-th it-
link (3,4) is influenced since packet loss probability on eration to compute the flows of the k+ 1-th iteration, we
link (1,2) is different. Then, due to the flow change of do not expect any significant flow errors since the differ-
link (3,4), all the flows of down stream links, such as +
ence between FS$lp‘( k ) and F$”’ (k 1)is very small af-
link (3,5) and (5,6), along path 2 need to be adjusted. ter certain number of iterations. For all the testfd cases
Because of the flow adjustments of these down stream
links, other link flow adjustments, such as,link (5,7) and shown in section 6, the differences between ( I C ) and
F$lP
(7,8) of path 3, will be needed. Since in our algorithm, Fz:’2p‘(k+ 1) are smaller than 2% of F$Jp’(k + 1) after
all the path flow adjustments (therefore, the link flows 50% of the number of the iterations needed for stage 2.
2
along different paths need to be performed in parallel,
we have to get rid o this correlation.
Please note that our two-step method described in (10)
and (11) is used for the computation of the second stage
of our two-stage algorithm. We can also apply the HAD
scheme [7], the textured decomposition scheme [8][9] and
parallel GPM algorithm [6], depending on the network
topologies, to our two-stage algorithm to speedup the
Pat computations.
292