Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

An Algorithm for Survivable Network Design Employing Multiple Self-healing Rings

J.B. Slevinskyl, W.D. Grove?, M.H.MacGregor


TRLabs, #800 Park Plaza,
10611 98 Avenue, Edmonton, Alberta, Canada, T5K 2W
phone: (403)441 3800 fax: (403)441 3600
e-mail: grover@trlabs.CA
1. On Research Assignment from AGT Limited 2.Also with the Dcpartment of E!lectrical Engineering,University of Alberta

Abstract: We have developed an algorithmic approach to as a "design proposer" adjunct to existing costing and analysis
synthesis of restorable networks that employ several self-healing tools such as in [1,2].
rings (SHRs). The method is near-optimal from a capacity effi- Related work on this problem is found in [3,4]where hub
ciency perspective. The multi-ring design configurations that detection and costing analysis are used to place rings in combina-
result may be used directly, or the algorithm may be used as an tion with hubbing with diverse 1:l protection systems. Another
automated "design propose?' to complement existing ring plan- approach, used in [5], involves manual development of a series of
ning analysis tools. A systematic search for an optimum multi- network designs using all length 3 rings, then all length 4 rings,
SHR design is structured as a greedy search for ring candidates and so on up to all length 9 rings and then selecting a set of rings
amongst the set of cycles of the network. Novel aspects of the approximately based on those ring sizes that were individually
approach are: most efficient in each region. The authors are also aware of
(i) determination and use of the cycle set of the network graph to unpublished efforts to apply simulated annealing to the multi-
structure a systematic search for collectively efficient ring sets SHR problem. Another general approach to ring placement is
that cover the network, visualization of communities of interest in the demand matrix by
(ii) use of several different metric functions to explore alternative variable thresholding [ 1,2], but this still relies heavily on the plan-
search paths in the design space. Each identifies a different locally ner's intuition and judgement.
optimal network construction from which the best global candi- There are two main concepts of efficiency in the overall
date is easily selected. problem of finding a cost-optimal multi-SHR design: (i) capacity
On small and medium test cases, results show that the algo- efficiency and (ii) traffic capture efficiency. Capacity efficiency
rithm produces good designs, strictly optimal in some cases, if has to do with how well balanced the working capacities on spans
assessed from the point of view of capacity efficiency. For larger of individual rings of the network are. The greatest capacity effi-
networks, where the strictly optimal result is not known, we still ciency arises in a perfectly balanced ring where demands are all
show improvements over previously published results. between adjacent neighbors only. Capacity efficiency is levered in
overall cost effectiveness by the costs for fiber, regenerators, and
I . INTRODUCTION ring terminal line interfaces and add-drop multiplexer (ADM)
An attractive feature of Selfhealing Rings (SHRS)is the rel- size. On the other hand, ADM interface costs are generated by
ative simplicity with which a single SHR operates and can be traffic demands that have to transit from one ring to another. Min-
designed. Somewhat ironically, however, design of optimal sw- imization of these costs implies maximum capture of traffic
vivable networks that require several SHRs is an extremely com- demands within each ring., i.e., the choice of a ring set which
plex problem. There are so many possible combinations of ring gives the greatest number of demands whose origin and destina-
number, sizing, spatial placement, and traffic mapping, that it has tion are in the same ring. The total optimization problem therefore
so far been unclear as to how one might structure a systematic embodies finding the best compromise of balance and capture
exploration of the design space to seek optimal solutions. It can efficiency effects. This overall problem is not yet solved. Here, we
be shown, for example, that in a network in which there are as few focus on one part of the problem: optimizing capacity efficiency,
as 8 distinct simple cycles (a graph-theoretical term to which we as a component strategy towards an overall optimization of both
return), there are 109,600different combinations of ring number effects in future work.
and placement (not all of these are survivable.) Moreover, as an Novel aspects of the following approach to optimizing
optimization problem, multiple-SHR network design does not capacity efficiency are:
lend itself to formulation by well-known methods such as gradi- determination and use of the cycle set of the network graph to
ent search or linear programming. structure a systematic means of determining ring choices, and
This is reflected in the fact that most ring-network planning coordinating the complete set of rings chosen for collective
tools today are primarily unulysis systems, as opposed to synthe- efficiency.
sis systems [1,2]. They analyze designs presented to them by a
concurrent use of several different cost (metric) functions to
planner who uses guidelines, intuition, and other heuristics to
develop a proposed design. Cost-effective designs are obtained synthesize ring sets that represent simultaneous alternative
this way by experienced planners, but it seems useful nonetheless search paths in the design space. Each identifies a different
to have an algorithm that could automatically synthesize near-
optimal multi-ring design configurations, either for use directly or
0-7803-0917-0/93$03.00 Cp 1993 IEEE
1568
locally optimal network construction from which the best glo- B. Balanced Shortest-pathDemand Mapping
bal candidate is easily selected according to efficiency or The input data for demand mapping is the network route
detailed cost assessment. topology and the point-to-point demand matrix. Demands are
Results show that the algorithm produces quite good mapped onto the spans of the network via shortest path routing,
designs, even strictly optimal, if assessed from the point of view with a load-levelling or balancing criterion to choose amongst
of capacity efficiency in test networks for which it is possible to multiple equal length route alternatives. This is a single route
exhaustively test all design alternatives to find the globally best selection that yields the best uniformity in span wi quantities
design for comparison. For larger networks which cannot be amongst equal length routing alternatives possible at each stage
exhaustively tested, we are able to show a case of improvement of demand mapping. This is analogous to Wu’s algorithm for
over a previously well-honed published result [SI. choosing between equal length routes in a bidirectional SHR [3],
Capacity efficiency is used as a working measure of opti- but applied to shortest routes across the basic span topology of
mality for research purposes. This does not translate directly into the network before any ring placement has been made.
strict cost optimality, but there is a correlation. Ultimately, the
combined simultaneous optimization of transport efficiency and C. Cycle-Finding
ring access efficiency, incorporating an arbitrarily detailed cost- A simple cycle is any closed path within the network that
ing model within the synthesis process, will result in a strictly has no repeated nodes. The cycle-finding step identifies all dis-
cost-optimal system design tinct simple cycles of the network b d only on the topology of
the network. This is critical because all possible rings that could
Outline: The next section describes the algorithm and its be placed are thereby enumerated. The multi-ring design problem
optimization search strategy. Section 3 explains three types of can then be formulated as a systematic saarch through the cycles
testing and the results obtained to judge performance of this for a set of rings that covers all network spans efficiently.
approach. The final section presents a summary of insights and Although cycle-finding follows demand mapping in
areas for further work based on this study. Figure 1, it could proceed in parallel with demand mapping
because cycle-finding and demand mapping are independent
phases executed prior to the ring selection search. Cycle-finding
Il DESCRIPTION OF THE ‘RINGBUILDER’ could run before,or in parallel with, demand mapping.
ALGORITHM A search over the sat of all cycles of the network can be
used to structure this problem, as opposed to separately consider-
A. Preliminaries ing the number, sizing and placement of rings to use. Finding a
The bidirectional self healing ring (BSHR) [6],also known selection of cycles by the criteria that follow implicitly specifies
as the shared-protection (SP-ring) architecture, is assumed. number, sizing and placement of all rings required for a near-opti-
Adaptation of the algorithm specifically for 2-fiber or 4-fiber mal survivable network. Because the number of distinct simple
variations follows easily. To accommodate modularity of the cycles in a network can be very large, computational efficiency
actual fiber systems used to implement SHRs, working and spare becomes an important issue. The number of simple cycles, c, in a
values can be quantized to nearest module numbers in evaluation network is bounded above by [71:
of the cycle metrics. Note that we equate links with demand units
(e.g. STS-1’s).
For present purposes, design optimality is defined in tenxis
of the aggregate redundancy, r, of a complete multi-SHR surviv-
able network design: where S is the number of spans and N is the number of nodes.
And there are:
s s
C
C s i / C wi
r=
i=l i=l
(1)

where siis the number of spare links on span i of the network, wi


z-
.

I= 1
I
C!
(3)

is the number of working links on span i of the network, and S is


the number of spans in the network. Capacity efficiency is possible ways of choosing designs based on these cycles. Not all
defined as l/r. of the permutations expressed in (3) will correspond to 100%
This is a valid basis for comparison when all network restorable coverings of the network, but it suffices to appreciate
designs are 100% restorable and when the same mapping of how large the design space is. For instance, in networks with only
working capacities (wi’s)is common to all designs. No changes 6,7and 8 distinct cycles there are, respectively, 1956,13,699and
to any working capacity are made after initial demand mapping. 109,600possible multi-ring designs. Mathematical permutation
RingBuilder has the overall structure shown in Figure 1. We (which is even stronger than combination) of cycles arises
will first outline the demand mapping, then discuss the cycle find- because the order of selecting cycles to become rings is a factor
ing, cycle metric evaluation and ring selection steps in more in the redundancy of the completed multi-ring design. This will
depth. become more apparent when ring set selection is explained.
Because the number of cycles in the complete cycle set can be so

1569
large we later describe a variant of RingBuilder where cycle-find- algorithm may find local rather than global minima. Usual opti-
ing executes progressively to expand the cycle set, and for each mization strategies in such circumstances are to try and explore
partial cycle set an optimal design is determined. In this way the the solution space from multiple starting points, or to find ways to
design evolves, improving with the amount of computer time diversify the search trajectory to improve the likelihood of dis-
invested. covering a global optimum.
The cycle finder uses recursive backtracking to enumerate RingBuilder uses a number of different metric functions to
all possible simple cycles of the network. The routine starts by achieve diverse trajectories of exploration in the design space.
choosing an arbitrary node in the network, and then initiates a Each metric guides the synthesis of a network design through a
depth-first search of the network looking for a cycle. A cycle is slightly different sequence of cycle selections. In addition, the
detected when a node already seen in the search path is encoun- aspect of searching from different starting points can be effected
tered for a second time. Once a cycle is detected, it is checked for by exploring the designs that result when the second-best cycle is
uniqueness (the same cycle may be found several ways), and chosen in the first design step. This variant is referred to in our
added to the list of cycles. A cycle is “named” uniquely by listing results as “second choice first” (SCF).
the nodes in the cycle in a particular ordering. This ordering starts
at the lowest-numbered node, and proceeds around the cycle in E. Cycle Metrics Employed
the direction of the lowest-numbered neighbor of the lowest- Seven metric functions which have been employed in Ring-
numbered node. Once the cycle has been tested for uniqueness, Builder are listed in equations (4) through (10) below. In all
the search backs up to the penultimate node in the search path cases, n is the length of the cycle and the wi’s are only those per-
and proceeds along the first remaining untried span at that node. taining to spans covered by the cycle. As outlined above, if any
If no untried spans are left, the search backs up to the previous span has been covered by an SHR selection in a previous itera-
node in the path, and looks for untried spans there. tion, then the corresponding wi for than span is set to zero in (4)
through (10).
D.Cycle Metric Evaluation and Ring-set selection
Once a complete (or partial) cycle set is available for the n- 1
network, the list of cycles is iteratively examined to find, in each
iteration, the one cycle that presents the highest value of a partic-
C wi
i=O
ular selection metric. The selection metrics are functions of the f o ( n , w $ = ( n x max ( w i ))
efficiency of each cycle as an SHR candidate. The selection met-
ric is calculated for each cycle in each iteration and inherently
takes into account: n- 1

(i) the exact path on the network that the cycle describes,
f l (n7wi) = n x max (wi> - C
i=O
wi
(ii) the wi values on the spans traversed,
(iii) the number of spans in the cycle, n, and,
(iv) any overlaps with other cycles that have already been
selected as SHRs.
i=O
The cycle that maximizes the metric in a particular iteration
is selected to become an SHR in the network design that is evolv-
ing under the guidance of that metric. When a cycle is selected to
become a ring, all of the wi covered by that ring are set to zero.
Zeroing out the wis implies that when assessing remaining cycles
as ring candidates, they may cover this span again but will gain
no “credit” for doing so. This is somewhat different than in [3]
where rings are not allowed to share spans, so that only ‘Olym-
pic’ (interlocking) ring topologies are formed. In contrast, Ring-
Builder allows results where rings may be interlocking or
partially overlapping (having spans in common).
Once a cycle is chosen as a ring, the SHR which it implies
is dimensioned so that s i = m (wi) in each span of the ring,
where max (wi) is over only those wi included in the ring. The
algorithm stops when all network spans are members of some
SHR.
The process of selecting the best cycle to become a ring is a
“greedy” approach. The best cycle in the q’th iteration depends
on all previous (4-1) ring selections. Obviously, therefore, the

1570
fn-1 1 work does provide a comparative basis for assessment of the
RingBuilder design results for these networks. Network E is the
previously-published“Bellcore”study network [SI,shown in Fig-
ure 6. Network F is based on an existing Canadian metropolitan
network and is shown in Figure 7. Network E has 307 distinct
These functions express a common concern with the central cycles, while network F has 438 distinct cycles.
aspect of SHR transport efficiency;that is, the mismatch between The last type of testing was on still larger networks for
the wi’s in a ring and the total sparing investment (n*max(wi)) which we do not know the strictly optimal result and the cycle set
within that ring, but they vary in their biases as follows: Metricfb is too large to be completely enumerated in a reasonable amount
is the simple efficiency ( Z h ) of an SP-ring. This is of obvious rel- of time. For this type of network, RingBuilder executes in a pro-
evance as a minimization objective but fo tends to select small gressive mode, building up a partial cycle set and improving the
rings (n=3) in which the ratio of working to spare is high, but the design for as long as we wished to let it run,or until the cycle set
absolute amount of working capacity covered may be too small to is completely enumerated. The algorithm looks for a new network
necessarily be the most effective ring choice.fi follows from this design after each new cycle is found. In these cases we have nei-
criticism of fo by calculating the value of the difference between ther an exact optimum nor published designs for comparison but
working and spare capacity in the ring. Note thatfo andfl are we are able to demonstrate systematically improving designs.
both maximized if the variation in working span capacity within
the candidate cycle is minimized, and thatfi reflects the absolute
amount of capacity involved, rather than a ratio. B. Results
The third metric, f’, maximizes the amount of working On networks A,B,C, and D where exhaustive search was
capacity protected by any candidate cycle. It favors large cycles. feasible, all possible designs for the networks were enumerated
Functionf3 is a hybrid that attempts to find cycles with low vari- and evaluated in terms of their redundancy, r (Eq. (1)). All
ance and high total coverage. Functionsf4 andfs are variations designs were then summarized in the form of the frequency ver-
that give different relative power law emphasis to the absolute sus redundancy plots shown in Figures 8-11. The abscissa is the
working capacity covered (to the credit of the cycle) versus the value of redundancy for a given design; the ordinate is the num-
total sparing required in the ring to do so. Finally,f6 is based on a ber of designs with that value of redundancy. To assess Ring-
slightly different model for SP-ring efficiency: a term (ZLt) is fac- Builder results, the redundancy of the networks designed by each
tored out of fo,and discarded. This results in an efficiency expres- metric are overlaid as shown in Figures 8-11. Bracketing the
sion which is less biased by the number of spans a candidate results indicates that the result was obtained with second-choice
cycle has. first (SCF).
It is important at this stage to stress that the question is not These results show that strictly optimal designs resulted for
which metric is best. Rather, each metric will drive the search networks A,B and C. SCF located the best solution in network C.
through a different trajectory to a different design that is “good” In network D, SCF located the third ranked design (out of 1956
by its particular biases. The final step of inspecting the seven possibilities). From Figures 8-11, we find that, in general, there is
designs generated to select the best from them requires an inde- no basis for preferring one cycle metric over another. The impor-
pendent re-evaluation of cost or efficiency optimality. tant aspect is whether at least one metric within the set of seven
identified an optimal design in each case.
III PERFORMANCE
a
In the second type of testing, the designs from RingBuilder
yielded an improvement compared to previously published results
A. Test Networks [5] for network E. In the RingBuilder design the network‘s 313
Three types of tests were arranged to assess RingBuilder’s working links are protected with 432 spares, versus 440 spares for
performance. In the first, tests, we used networks with a small the manually developed solution in [SI. For network F, Ring-
number of cycles, evaluated the redundancy for all possible Builder’s best design had a redundancy of 1.93. From experience,
designs, and compared the results to those from RingBuilder. We this is good given that the network has highly unbalanced work-
know how well RingBuilder performs relative to a strictly opti- ing capacity distribution (working span capacity ranges from 28
mal design in this case. Results are presented for 4 test networks to 481).
of this type, having (A) 6 nodes and 8 spans, (B) 7 nodes and 9 Lastly, RingBuilder was also tested on networks with very
spans, (C) 8 nodes and 10 spans, and (D) 9 nodes and 11 spans, large numbers of cycles. As described above, a ‘progressive exe-
respectively. Each network had 6 distinct cycles and thus 1956 cution’ form of the algorithm was developed for these purposes.
possible designs. Networks A, B, C, and D are shown (with wi Figure 12 presents a sample result for network G (the network is
span capacities indicated) in Figures 2 through 5. not shown) that has 39 nodes, 64 spans and over 35,000 distinct
The second type of test used two slightly larger networks cycles. The design evolution in Figure 12 shows that generally, as
(networks E and F) with enough cycles that the solution space the number of candidate cycles increases, RingBuilder finds bet-
was too large to locate the optimum by exhaustive search, but for ter designs. Sometimes many candidate cycles must be added
which the entire cycle set could be enumerated in a reasonable before a better design can be found, indicating that some cycles
amount of time. Although strictly optimal answers are not inde- were not able to contribute to the design. Also, the addition of a
pendently available for these two networks, previously published

1571
cycle occasionally results in generation of a poorer design. This
is due to the greedy nature of the algorithm; sometimes a new
cycle is utilized which degrades the efficiency of later choices. Shortest Path
Nonetheless, as the number of candidate cycles continues to
increase, the design once again improves. 13

IV. SUMMARY
Cycle Finder
A novel strategy for the synthesis of capacity-optimized

a3
survivable networks that are based on multiple SHRs has been Figure 2: Network A
developed. Novel aspects include systematization of the optimi-
zation problem in terms of a m y search in the cycle-set of the
network, and the use of multiple metric functions to explore sev-
eral mutes in the design space simultaneously; the best design
can then be chosen for implementation. The metric functions
shown in the paper am examples only. The basic technique is Cycle Metrics
extensible to cost functions of arbitrary complexity, possibly 38
including detailed real-world costing models. Tests on three
classes of test network showed that the designs obtained were Figure 3: Network B
either capacity-optimal or neu-optimal in all cases. Results also
showed the option for progressive design evolution for very large
networks. Ongoing work includes examination of the traffic-cap- 17 57
ture aspect of the multi-SHR design problem, and the mecha-
nized means for automatic determination of span eliminations (or
non-utilization of selected spans), in the generation of covering 51
cycle sets. Capacity 19
Update 59
Acknowledgment 54 27
The authors wish to thank William Misskey, TRLabs Visiting Figure 4: Network C
Professor for helpful comments on the manuscript.

References
[ 11 Northern Telecom, Tmsition Planner User’s Guide, 199 1.
[2] J. Hopkins, ‘“bols for planning survivable networks,” Phoe-
nix Comforum ‘90,Mar. 1990, pp. D3.SD3.19.
Post Processor
[3] T.-H.Wu, FiberNetworkService Survivability, Artech House
Inc., 1992, pp. 308-318.
Figure 1: Major Steps Figure 5: Network D
[4] 0. J. Wasem, “Optimal topologies for survivable fiber optic in RingBuilder
networks using SONET self-healingrings,” Pmc. IEEE
Globecom ‘91,Dec. 1991, pp. 2032-2038.
[5] W. D. Grover, “Case studies of survivable ring, mesh and
mesh-Arc hybrid networks,” Pmc. IEEE Globecom ‘92,
Dec. 1992, pp. 633-638.
[61 T.-H. Wu, Fiber Network Service Survivability, Artech House
Inc., 1992, pp. 131-138.
[7] P. Mateti, N. Deo, “On algorithms for enumerating all circuits
of a graph,” SIAM J. Comput.,Vol. 5 , No. 1, Mar. 1976.

Figure 6: Network E Figure 7: Network F

1572
I

!. :
01

3:
% 4

li
= 0 .............................

Rdundmncy
fff
PP2
f
3 --L+ZE&E
$!-* Rrnlo

3
-. 3
-&

3
Figure 8: Network A Solution Space Figure 9: Network B Solution Space

11 I

Figure 10: Network C Solution Space Figure 11: Network D Solution Space

3.0

2.5

2.0

I
1.0 10000 20000 30000 40000
c, Number of cycles available as ring candldates

Figure 12: Incremental Solution Trqjectory for 39 Node Network

1573

You might also like