Performance of IP over optical networks with dynamic bandwidth allocation

Abstract

IP over optical network performance can be improved with dynamic bandwidth allocation, depending on the reallocation paradigm and the network topology. Under high connectivity, dynamic bandwidth allocation provides a notable boost to the network's traffic-carrying capacity.

Full text

Page 1

Performance of IP over Optical Networks with Dynamic
Bandwidth Allocation
Joel W. Gannett, George Clapp, Ronald A. Skoog, and Ann Von Lehmen
Telcordia Technologies, Inc., 331 Newman Springs Road, Red Bank, New Jersey 07701-5699
jgannett@research.telcordia.com

Abstract: IP over optical network performance can be improved with dynamic bandwidth
allocation, depending on the reallocation paradigm and the network topology. Under high
connectivity, dynamic bandwidth allocation provides a notable boost to the network’s traffic-
carrying capacity.
© 2005 Optical Society of America
OCIS codes: (060.4250) Networks; (060.4510) Optical Communications

1. Introduction
Running IP over optical networks that are capable of assigning bandwidth resources dynamically can deliver
improved performance [1]. Such networks shift bandwidth resources from one part of the network to another to
follow either time-of-day traffic pattern variations or focused traffic loads caused by exceptional events. This paper
addresses the performance of dynamic bandwidth allocation with respect to both the chosen reallocation paradigm
and the network’s connectivity. Our results indicate that the benefits of dynamic bandwidth allocation are modest
except for highly connected networks. In the latter case, dynamic bandwidth allocation provides a notable boost to
the network’s traffic-carrying capacity.
2. Simulation Results
Our simulation model is based on a network of IP routers with OC-192 ports. The ports incorporate SONET
VCAT/LCAS (Virtual Concatenation/Link Capacity Adjustment Scheme); hence, the 192 physical STS-1 channels
of each port can be grouped into one or more virtual links that provide one-IP-hop connections to other routers.
Although we are modeling the flexible bandwidth afforded by SONET VCAT/LCAS, the general principles we
demonstrate apply equally well to a network interconnected by a flexibly-switched WDM optical layer. In the latter
case, the switching granularity would be wavelengths rather than STS-1 channels.
To enable the simulation of large-scale traffic flow effects in complex networks, we model the usage of the IP
network as a stochastic sequence of abstract calls between routers, each with a finite stochastic holding time.
During the time it is active, each abstract call represents an increase of one STS-1 in the traffic flow between the two
routers at its end points. While it is active, such a call in our model consumes one STS-1 of bandwidth on each link
over which it is routed. Our call generation model uses exponential arrivals and exponential holding times. As each
abstract call arrives, its two end points are assigned randomly such that the expected time-average number of
requests terminating at a router is proportional to its number of ports (i.e., if router X has twice as many ports as
router Y, then X terminates twice as many abstract calls as Y on average over time). Although our traffic termination
model is unconventional, we find it more intuitive and more plausible than the conventional uniform traffic
assumption. Note also that no concept of distance enters into our selection of call end points, so routers that are
widely separated in the network may have as much traffic between them as routers that are close together. This
conforms to modern data networks, such as the Internet, where the interaction paradigm seems to be “everyone talks
to everyone.” Consequently, we feel that our distance-insensitive paradigm for assigning end points is more
appropriate than a conventional “gravity” traffic model.
We assume a centralized or distributed network management and control (M&C) system is used that issues
bandwidth reallocation requests to the VCAT/LCAS layer. In a real network, the M&C system would issue these
requests in response to changes in monitored performance metrics that indicate inadequate bandwidth (congestion)
in some part of the network. These metrics may include the packet delay or packet drop ratio. In our simulations,
we use the call blocking ratio (the ratio of blocked calls to total calls) as a surrogate for assessing network
performance. When an abstract call is blocked, that means there is less bandwidth on a link than there ought to be to
accommodate the offered traffic. For best performance, the blocking ratio should be kept as low as possible.

Page 2

We simulate two static bandwidth allocation schemes and two dynamic bandwidth allocation schemes, denoted
as follows: Static Allocation, Static Routing (SA-SR); Dynamic Allocation, Wait-For-Trouble, Static Routing (DA-
WFT-SR); Dynamic Allocation, Busybody, Static Routing (DA-BB-SR); and Static Allocation, Dynamic Routing
(SA-DR). SA-SR is similar to today’s Internet routing, that is, it is based on static routes between routers with static
bandwidth on the links. The static routes constructed in these simulations are minimum -hop paths between the
routers. DA-WFT-SR uses static routes and performs a bandwidth reallocation when (and only when) a call is
blocked. The reallocation algorithm incorporates a heuristic that tries to shift any available (unused) capacity over
to 100% utilized (blocked) links, thereby unblocking them. DA-BB-SR also uses static routing, but it checks after
each successfully routed call for any link that may have become 100% utilized as a result of that new call; hence,
DA-BB-SR reallocates more often than DA-WFT-SR. Once reallocation is indicated, DA-BB-SR uses the same
heuristic as DA-WFT-SR to unblock links by reallocating bandwidth. The reallocation paradigm must keep at least
one STS-1 allocated on every link at all times to avoid triggering IP reconfiguration. Finally, SA-DR leaves the
allocation static but uses Dijkstra’s algorithm to find the shortest path that routes around blocked links. SA-DR is
the only one of the four schemes that uses dynamic routing.
Our example network consists of 10 routers, numbered 0 to 9, with between 3 and 5 backbone ports for each
router and a total of 40 ports. The number of virtual VCAT/LCAS links between the router backbone ports is varied
from 20 to 45 in increments of 5. The end points of each link must terminate on ports of distinct routers, and we
allow no more than one link connecting two routers. Therefore 45 links is the maximum number of links allowed in
our example, as this results in one link between each pair of routers. The links are placed randomly, except the
placement algorithm guarantees (if there are at least 20 links) that each of the 40 ports is connected to at least one
link. Routers with more ports are likely to get proportionately more links. The 25-link network has all the links of
the 20-link network, plus five more, while the 30-link network has all the links of the 25-link network, plus five
more, etc. See Fig. 1 for a graph of the 30-link network.

Fig. 1. Network with 10 routers, 40 backbone ports, and 30 virtual links.
Each simulation experiment consists of 3,000,000 abstract call attempts, with the blocking ratio tallied at the
end. For the first 20% of the calls in an experiment (first 600,000 calls), the end points are generated randomly as
described earlier. For the next 20%, the focus is on router 8, which gets its average terminating traffic doubled from
about 10% of the total to about 20%. The focus shifts to router 1 in the next 20% of the calls, then to router 5 in the
next 20%. For the last 20% of the calls, the random end point generation paradigm of the first 20% is used again.
The dynamic allocation paradigms must track these changes in traffic focus to reduce the blocking ratio. Figure 2
shows the simulation results for the 30-link network. The graph shows the increase in call blocking as the network-
wide load in Erlangs (average call holding time) is increased gradually.
Figure 3 summarizes the results for the six network topologies simulated. We define the capacity for each
network topology and each allocation paradigm as the offered load in Erlangs that causes a blocking ratio of 0.001.

3. Concluding Remarks
With the Static Allocation, Static Routing (SA-SR) paradigm, which is similar to ordinary Internet routing, varying
the number of virtual links between routers implies two phenomena that seem to cancel. On the one hand, more
links means fewer multi-hop routes, which implies more efficient bandwidth usage and less blocking. On the other

Page 3

hand, more links means less static bandwidth per link, which means more blocking. The net result, shown in Fig. 3,
is that the network capacity oscillates around 800 plus or minus 200 Erlangs as the number of links is varied.
Call Blocking Ratio for 10-Router Network with 30
Links
0.00001
0.0001
0.001
0.01
0.1
700800900 1000 1100 1200 1300 1400
offered load (Erlangs)
blocking ratio
SA, SR
DA-WFT, SR
DA-BB, SR
SA, DR

Fig. 2. Call Blocking Ratio for the 30-link case.
Capacity of 10-Router Network vs. No. of Links
(Erlang Offered Load @ 0.001 Blocking)
0
200
400
600
800
1000
1200
1400
1600
1800
2025 3035 4045
no. of links
network capacity
(Erlangs)
SA, DR
DA-BB, SR
DA-WFT, SR
SA, SR

Fig. 3. Network capacity vs. number of links.
Similarly, the Static Allocation, Dynamic Routing (SA-DR) paradigm gives the highest network capacity but also
seems fairly invariant to the number of virtual links between the routers. The resulting capacity is almost twice as
high as SA-SR, being around 1500 plus or minus 150 Erlangs.
The two dynamic allocation paradigms, DA-WFT-SR and DA-BB-SR, perform similarly, with the more intensive
DA-BB-SR having a slight edge. These two paradigms show performance intermediate between the two static
allocation paradigms, with modestly rising performance as the number of virtual links is increased. Interestingly,
the performance of the dynamic allocation paradigms improves sharply, rivaling the performance of SA-DR, when
full virtual connectivity is provided and each router is just one IP hop away from every other router. Such a
topology is viable for small networks, but may be impractical when the number of routers is large.
Although dynamic bandwidth allocation may not quite match the performance of dynamic routing, our results show
that it can come close under certain conditions. Moreover, dynamic allocation networks may be easier to manage
than dynamic routing networks, as the latter are subject to thrashing, congestion spreading, and other instabilities.
Further work should explore what types of load variations, topologies, and other network properties make dynamic
bandwidth allocation a superior strategy to dynamic routing in terms of both management and performance.
References
[1] R. Skoog and S. Yun, “The value proposition for bandwidth on demand,” Optical Fiber Communication Conference 2003 Technical Digest,
Atlanta, Georgia, March 23-28, paper ThH3, pp. 482-483.