Scheme for alternative packet overflow routing (SAPOR)
ABSTRACT Shortest path routing schemes, like open shortest path first (OSPF), have shortcomings when networks are highly loaded. Traffic engineering of IP networks is required to avoid this problem. Current efforts suggest the optimisation of OSPF weights to balance the network load more evenly. Also, more advanced technologies, like multiprotocol label switching (MPLS), are proposed. One major problem of dynamic routing efforts that use OSPF is the fact that many traffic flows are influenced by single weight changes. We introduce SAPOR (scheme for alternative packet overflow routing), which realises a methodology that can remember the routing of packets for the duration of a micro flow. This allows the rerouting of overflow traffic. In this case, well known concepts and methodologies from conventional circuit switched teletraffic engineering can be adapted for IP networks.
Conference Proceeding: Granularity of QoS Routing in MPLS Networks.Quality of Service - IWQoS 2001, 9th International Workshop Karlsruhe, Germany, June 6-8, 2001, Proceedings; 01/2001
- [show abstract] [hide abstract]
ABSTRACT: A system of techniques is presented for optimizing open shortest path first (OSPF) or intermediate system-intermediate system (IS-IS) weights for intradomain routing in a changing world, the goal being to avoid overloaded links. We address predicted periodic changes in traffic as well as problems arising from link failures and emerging hot spotsIEEE Journal on Selected Areas in Communications 06/2002; · 3.12 Impact Factor
Conference Proceeding: Traffic Engineering Using OSPF Weights and Splitting Ratios.[show abstract] [hide abstract]
ABSTRACT: A method is proposed to perform traffic engineering for Autonomous Systems by setting OSPF weights and distributing new metrics called node splitting ratios. The quality of the traffic engineering is comparable to that achievable using MPLS. The weights and splitting ratios can be calculated using linear programming techniques. This potentially allows the solution of very large problems of a scale often found in an ISP backbone. Furthermore changes to network topology or addition of new flows can be easily done without performing the whole optimisation procedure. The method we have presented here offers the ability to perform Traffic flow optimisation without maintaining any per flow state in routers and without introducing additional protocols or packet overhead, unlike MPLS which is commonly used for Traffic Engineering.Converged Networking: Data and Real-time Communications over IP, IFIP TC6 / WG6.2 Sixth International Symposium on Communications Interworking (Interworking 2002), October 13-16, 2002, Perth, Western Australia; 01/2002
Scheme for Alternative Packet Overflow Routing
Alexander A. Kist and Richard J. Harris
RMIT University Melbourne
BOX 2476V, Victoria 3001, Australia
Telephone: (+) 61 (3) 9925-5218, Fax: (+) 61 (3) 9925-3748
Abstract—Shortest path routing schemes like Open Shortest
Path First (OSPF) have shortcomings when networks are highly
loaded. Traffic engineering of IP networks is required to avoid
this problem. Current efforts suggest the optimisation of OSPF
weights to balance the network load more evenly. Also more ad-
vanced technologies like Multiprotocol Label Switching (MPLS)
are proposed. One major problem of dynamic routing efforts
that are using OSPF is the fact that many traffic flows are
influenced by single weight changes. The Scheme for Alternative
Packet Overflow Routing (SAPOR) which is introduced in this
paper, realises a methodology that can remember the routing
of packets for the duration of a micro flow. This allows the
rerouting of overflow traffic. In this case, well known concepts
and methodologies from conventional circuit switched teletraffic
engineering can be adapted for IP networks.
IP networks are a major technology for the transport of data.
More than half of the transported traffic is already IP based.
The core backbone network delivers the IP packets to their
destination. Core routers and the interconnecting bearers form
these networks. Most topologies allow many alternative paths
in such backbone networks. Shortest path routing, in particular
OSPF , is widely used to select the appropriate paths
for origin destination pairs. Over a wide range of operating
conditions, OSPF provides optimal solutions, in particular, in
Internet Service Provider (ISP) grade Internet environments.
The principal problem of shortest path routing, which many
efforts try to solve, is that by definition all traffic is routed
on the shortest path. If the load of the network increases,
the shortest path links will be highly loaded, whereas other
alternative network resources are potentially unused.
On the other hand, many current initiatives work on the
migration of telephony and other carrier grade services to
all IP next generation transport networks (e.g. 3GPP ). In
such a context, the Internet philosophy of “best effort” service
and “over provisioning” are not satisfactory any more. Even
current QoS ensuring methodologies like DiffServ and IntServ
do not address all problems. In particular, traffic-engineering
problems are not solved by these QoS technologies. But the
major selling points of carrier grade services are guaranteed
service levels. Traffic engineering efforts are necessary to pro-
vide carrier grade services using IP network bearers. Mutipro-
tocol Label Switching (MPLS)  is one possible approach,
to engineer the current networks and allow arbitrary routes,
although, MPLS requires the migration of whole network
regions to MPLS. Recent research also targets QoS routing and
MPLS (e.g. Ying-Dar Lin ). Other research efforts directly
target the optimisation of OSPF protocol weights to adapt the
cost metric for given traffic demand matrices. This includes
the Equal Cost Multi Path (ECMP) effort which is now part
of OSPF Version 2. ECMP splits the traffic between paths with
equal cost. Fortz and Thorup  use Tabu search techniques
to optimise OSPF weights. Murphy et al.  use a linear
programming approach and the fast solver CPLEX to generate
optimised OSPF weights. Other approaches by Harmatos 
use a heuristic and by Ye et al.  use online simulation to
get to similar results.
One major disadvantage of all OSPF rerouting techniques is
that all have an impact on existing traffic flows. For example,
if weights of links are changed, all existing flows will be
rerouted. It can introduce major traffic shifts and instabilities
in the network. This is one of the major arguments in 
“why weight changes are bad”. A second argument stresses
the importance that the network operators are in charge. This
paper introduces a Scheme for Alternative Packet Overflow
Routing (SAPOR). In this case, only new flows are redirected
by this scheme, the routing of existing flows remains the same.
This is in particular interesting, since a few long living flows
carry a large fraction of the traffic and a growing number of
real-time services would profit from persistent packet routing.
In principle, this scheme can be combined with any of the
above-mentioned efforts. It extends existing methods with the
possibility of dynamic routing without the negative side effects
of weight changes. SAPOR is rather an enabling technology
than a new optimisation effort.
The principal concept of overflow routing is not new and
widely known by the teletraffic engineering community. Con-
ventional circuit switched network operators have used these
methods for years. Dynamic non-hierarchical routing (DNHR)
 which uses different path sets for different times of the
day, for voice carriers, was initially developed by AT&T.
Other examples of these efforts include works on Dynamically
Controlled Routing (DCR) , Dynamic Alternative Routing
(DAR)  and State- and Time-Dependent Routing STR .
Gerald Ash’s book  presents a comprehensive discussion of
The discussion in this paper will use the notion of traffic
flows. A flow is the aggregate of a large number of packets
that are directed from the same origin node to the same
destination node in the observed network. These flows can be
measured in bytes per second. A micro flow is the aggregate
of packets between the same IP packet endpoints, i.e. the
same source and destination IP addresses. To further diversify
flows, port and protocol numbers can additionally be used to
define micro flows. The remainder of this paper is organised
as follows: The next section introduces the concept of SAPOR
and Section III describes the different functional components
in more detail. Section V discusses issues like the performance
and the requirements of SAPOR. The paper concludes with a
discussion of further work.
A router on the network layer has to route all incoming
packets on the appropriate links of the paths to the destination
node. The SAPOR scheme is located in routers. Figure 1
depicts the concept of SAPOR, i.e. the way outgoing links
are chosen for incoming packets. Compared to a conventional
system, SAPOR is located in between the router function
that switches the packets on the links and the routing table
that is generated by the shortest path algorithm. The main
functional groups of SAPOR are the Hash Function, the Token
System and the Routing Tables. The hash function consists
of two parts: the actual function (2) and the hash space (3).
The token system consists of the token buffers (6), the token
scheduler (5) and the token list (4) which is equivalent to the
hash space. The routing tables (7) are a selection of different
tables. The incoming packets are accumulated (1) and routed
on the emanating arcs (8). The dotted lines indicate the token
flows, the dash dotted lines indicate requests to the routing
tables and the solid line shows the virtual packet flow in this
scheme. Every outgoing link has an associated token buffer.
To distinguish the different links and buffers in the discussion,
colours are used to indicate their affiliation.
This paragraph describes the operation of SAPOR. The
incoming buffer receives new packets (1). As a first step,
the hash space is calculated for a packet of the origin and
destination address by a hash function (2). In the second step,
a token buffer is selected (6). This selection is firstly based on
the primary routing table (7). If the table indicates an outgoing
link (e.g. red) and an available (red) token (6), the token is
assigned to the hash space (3) and the number of tokens in the
token buffer is reduced. In the case where the space already
has an assigned token, this step is skipped. The current packet
and all subsequent packets are routed on the link with the
same colour (red). If the token buffer is empty, the secondary
routing table is considered and the same procedure is used for
the second table. If new packets arrive that are mapped on the
same hash space that already holds a token, the packets are
sent on the outgoing link with the appropriate colour. Tokens
have a finite time to live. If no new packets arrive within the
specified time, these tokens are selected and returned to the
appropriate token buffer by the token scheduler (5). If a packet
arrives before the time has expired the timer is reset.
Fig. 1.SAPOR Scheme
III. BUILDING BLOCKS
SAPOR uses three major building blocks: The hash func-
tion, the token system and the routing tables. This section
discusses the functional blocks in more detail.
A. Hash Function
The purpose of the hash function is to separate single micro
flows on the basis of flow specific parameters like the source
IP address, the destination IP address and other parameters.
A wide range of possible hash functions exist. The work by
Cao et al.  discusses hashing based schemes for Internet
load balancing and investigates the performance of different
methods. Most of their results can be applied in this context
as well. The important attributes of a hashing scheme for
SAPOR can be summarised as follows: The results have to
be evenly distributed over the hash space and the hash space
should be large enough that no frequent overlapping occurs.
For example, a 16 bit hash function yields 65536 hash spaces.
If link specific hash functions are used (See the alternative
Section V-F) the overlapping is less critical and smaller hash
spaces can be used. For the purpose of SAPOR, the XOR
folding of the source and destination address is sufficient.
Equation (1) shows this calculation.
⊕D1⊕ D2⊕ D3⊕ D4
The ith bytes of the source and the destination IP address
are denoted by Si and Di respectively. N limits the size of
the hash space.  indicates a sufficient spreading and this
scheme is simple to implement. Alternative hash functions can
also be used.
If the mapping of the hash space changes, disruptions can
occur since the mapping has to be reorganised. The concept of
robust hashing was introduced in  to combat this setback.
S1⊕ S2⊕ S3⊕ S4
However, this problem does not occur in this context, since the
token mapping is persistent, even when new links are added
or they disappear. For further discussion of failure modes see
B. Token System
The token system in combination with the hash function
is a major part of SAPOR. The number of available tokens
defines the size of the aggregated flows that are allowed on
outgoing links. This behaviour is similar to the token buffers
that are used for the leaky bucket scheme (e.g. ). The
number of tokens depends on the (statistical) properties of
the flow aggregates and the interval in which the token list is
updated. A token buffer is associated with a link and it has
a limited number of tokens. Every time a flow is transported
on this link a token is assigned to the flow. If no tokens are
available no more flows can be routed on this link. Tokens
have a (token) time to live (TTTL). If a flow transmits no more
packets and the TTTL has expired the tokens are returned to
the appropriate buffer.
1) Token Buffers: The token buffers can be simply imple-
mented by a counter that indicates the number of tokens in a
buffer. If a token is removed, the number decreases; if a token
is returned, the number increases. Under normal operating
conditions, this number is positive, but if the bucket size
changes during operation, the bucket count can be negative.
No tokens are available if the token count is less than one.
2) Number of Tokens: The major parameter that specifies
a token buffer is the number of available tokens ν. It defines
the maximum number of flows on the associated link. The
calculation that is shown in this section uses the assumption
that many micro flows exist and the measure of average flow
size is a valid approximation. This requires, in particular, that
no micro flow dominates over others, i.e. no single flow peaks
the flow aggregates. The average micro flow size is denoted by
f and is measured in bytes per seconds. The average duration
of a flow is denoted by t and measured in seconds. The flow
arrival rate is denoted by ϑ and measured per second. The
number of active flows A can be calculated by Equation (2).
A = ϑ · t
Equation (3) shows the calculation of the number of required
tokens νl for link l. The first tokens are determined by the
number of flows on the link, the second sum calculates the
number of tokens that are required due to the TTTL of the
tokens and the delayed reset of τ.
νl= Al+ δ · τ ·
where lmax is the number of outgoing links and δν is the
average flow change rate measured in new flows per sec. The
second product is necessary to scale the change rate to the
single emanating links. The possible number of flows Amax
depends on the capacity Cl, the average flow size f and the
utilisation ul. Amaxis depicted in Equation (4).
For example, an average flow size of 102.4bytes/sec, an
average duration of a flow of 30 seconds, an average flow
arrival rate of 207.8Flows/sec, a token update every 10
seconds and a utilisation of 60% yields 8144 tokens for one
single outgoing link (Equation (5)).
νl= 204.8 · 30 + 200 1/sec · 10 sec = 8144
The required link capacity can be calculated using Equation
Cl=6144 · 204.8 bytes/sec
= 1 Mbyte/sec
3) Token Scheduler: The token scheduler is the function
that determines “expired tokens”, clears the token list and
returns them to the token buffers. The token scheduler is
the most expensive function of the SAPOR scheme, but it
is not frequently used. It requires two major sub functions:
The traversing of the token list and the implementation of the
TTTL. The discussion in this section gives possible implemen-
tations, although there are many different ways for how the
same functionality could be achieved.
The token list can be implemented by using a combined
array-linked-list data structure. This avoids the requirement
that all spaces have to be traversed for the price of additional
memory. Every time a storage space is assigned, the “next
pointer” of the last added space, is pointed to the current space.
The current “next pointer” is set to a dummy end node. A
global pointer remembers the end of the list space. During
the iteration-step, the array-list is traversed in the sequence
of the linked list. The iteration-step also remembers the last
space it visited. In this way it is possible to delete a space
by connecting the “next pointer” of the previous space to the
space that is indicated by the “next pointer” of the current
space. Other implementations are also possible.
The TTTL can be implemented with a simple counter
variable. Every token list item has one such variable associated
with it. Zero indicates that the space is empty, i.e. no flows
are using this space. Empty spaces are skipped by the iteration
operation. If the number is positive it is increased every
time it is traversed by the iteration function. If a packet
arrives during the update interval the number is reset by this
event. The iteration function is executed every Tc seconds.
If n reaches the maximum count nmax the counter is set
to zero and the token is returned to the appropriate token
buffer. nmax is a positive integer that is larger than one.
In this case, the minimum time before a token is deleted is
(nmax− 1) · Tc and the maximum time before a token is
returned is nmax· Tc, where Tcis the time interval between
two iterations. This difference is due to the resolution which is
defined by the number of count steps 0...nmaxfor the timer.
Larger nmaxyield smaller differences between the minimum
and the maximum values. The mean time τ before a buffer is
delated is therefore defined by Equation (7).
τ =2 · n − 1
It is obvious that for large n the fraction approximates n.
4) Token Update Interval: Since the token scheduler is the
most expensive function, the token update interval Tcshould
be as long as possible. On the other hand, long intervals require
more tokens, a larger token list and the traversing of more
active tokens in the token list. The number should be chosen
on the basis of the evaluation of the given constraints. For the
discussions in this paper an arbitrary number of 10 seconds
has been chosen.
IV. ROUTING TABLES
The actual routing function of SAPOR uses a number of
possible alternative routing tables. They are denoted by pri-
mary, secondary, tertiary, ... n-ary table etc. These tables can be
generated by any means that are appropriate and useful for the
network configuration. SAPOR uses one primary routing table
per node and subsequent routing tables can be link specific.
If all routing tables are the same and built by a standard
routing algorithm (e.g. OSPF), the SAPOR scheme behaves in
exactly the same way as the original OSPF implementation.
In general, the mechanisms of the legacy systems should be
used for the primary routing table, for example, the original
shortest path routing tables. The routing decisions for the n-ary
routing tables and therefore paths can be based on any factors,
e.g. administrative decisions by human network operators. In
the case of SAPOR, the administrator can stay in charge of
any (overflow) routing decision. A simple automatic way of
generating these secondary routing tables is to remove the
outgoing link in the shortest path calculation and recalculate
the shortest path tree without the first link for the relevant
secondary routing table. The same principle can be applied
for all subsequent emanating links.
Secondary routing tables could also define a complete
redundant link system that is only used if the primary system
is overloaded or fails. As mentioned before in this paper,
existing teletraffic engineering methodologies can be applied
and used. The engineering efforts are reflected in the selection
of primary, secondary, etc. paths. Note that the path choices
are arbitrary but it has to be ensured that the primary and all
n-ary paths do not build loops. This is particularly important
since this scheme is node-local.
This section discusses some further issues that concern the
SAPOR scheme, in particular, requirements and performance
A. Speed of Changes
If the routing tables in the SAPOR scheme are changed, the
traffic flows do not change instantly, they are smoothly shifted
onto new links. The change speed is defined by the number of
new arrivals per second and therefore departures per second.
Equation (8) shows the duration of a change δt.
δt = ϑ · δν
Where ϑ is the flow arrival rate in flows per second and δν
is the number of existing flows that have to be rerouted. For
example, if ϑ = 200Flows/sec and δν = 1000 existing flows
have to be rerouted this will take 5 seconds.
B. Network Topology Changes
In the case of network topology changes, e.g. connection
or nodes disappear or become operational, the scheme has to
adapt to these changes. When new connections come online,
SAPOR will simply shift flows onto new capacities as they
become available. This will occur according to the speed of
change philosophy outlined in the previous section.
If nodes are no longer reachable, the speed of change might
not be fast enough, since packets will be still routed on
obsolete links. The failure mode for adjacent nodes is fairly
simple. If an adjacent link is failing, the corresponding token
buffer is emptied and all tokens in the token list that belong
to this link are flushed. In this way all existing flows will be
assigned to new links.
If ”non adjacent” links or nodes disappear, there are several
possible actions. Firstly, no specific action is taken and the
flows rearrange with the speed of change. Flows that are
routed to nodes that have lost their connectivity are disrupted.
Secondly, the token list is completely flushed if major routing
table changes occur. In this case flows that are routed by the
primary routing table end up on the same links as before,
the overflow routing however may be disrupted. Thirdly, only
tokens are flushed that belong to links that are the root to
changed network parts. These can be identified by comparison
of the different routing tables.
Lastly, only tokens of links are flushed that have lost their
connectivity. To identify changes in connectivity a simple
method can be used. Routing tables in nodes are written as
vectors, where every dimension identifies one emanating link.
The elements consist of a collection of nodes nxyz that are
reachable via this link. Equation (9) depicts an example of a
vector PT1which reflects a primary routing table.
All other routing tables can be written in the same way as
PT2, PT3etc. The connectivity vector PT can be calculated
by using the logical “or” function on the vector elements.
Equation (10) shows this calculation.
RT = RT1∨ RT2∨ ··· ∨ RTn
The elements of RT in one dimension identify nodes that
can be reached via the link that represents this dimension.
To find links that have lost their connectivity, the RT values
before and after the table changes have to be compared. This
calculation is depicted in Equation (11).
L = RTt∧ PTt+δt
The logical “and” operator results in differences in connec-
tivity that existed in the first vector but not in the second
vector. All links in vector L that are not empty have changed
connectivity and have to be flushed.
The fourth option is the least disruptive, but it also requires
the most effort. Further research has to identify the right
approach for specific networks and operating conditions. For
instance, the topology in the core part of backbone networks
will change far less frequently than the topology in access
parts of IP networks and the problems will therefore require
C. Requirements and Performance
1) Complexity: The complexity of SAPOR is twofold. All
operations that are required during packet routing are of the
order of one: O(1) i.e. the hash function, the check if a token
is assigned, the lookup of the routing table, the assignment of
a new token and the changes in the token buffer, therefore,
it is possible that all packets are routed in O(1) time. If a
token is already assigned to the buffer less O(1) steps are
required. The second operation is more complex, and therefore
more expensive, but it is less frequently required. To clear
and update the token list, all spaces have to be traversed. The
complexity of this process is therefore O(l) where l indicates
the size of the token list. l is equivalent to the hash space.
Using more memory and intelligent data structures can reduce
the number of spaces that have to be traversed.
2) Memory: The memory requirements for the routing
tables are equivalent to the requirements for tables of legacy
systems. The token buffers are simple counters and require
only several bytes per buffer. The token list is of size l and
requires therefore l bytes. However, a more advanced token
list requires k · l bytes.
As previously mentioned, SAPOR is an enabling technology
and not a routing methodology in the principal sense, there-
fore, actual performance improvements or comparisons with
existing technologies are not possible. The efficiency in terms
of improved utilisation etc. depends solely on the method-
ologies that are used to generate the alternative routing tables,
although SAPOR enables the use of these new methods. Future
research will have to focus on adaptation and development of
efficient routing methodologies for the SAPOR scheme.
E. Statistical Effects
All calculations in this paper are undertaken with the
mean values of the appropriate parameters. In practice these
simplifications will be sufficient, since many hundreds of flows
are observed. If SAPOR is used in very specific environments,
e.g. where many long-term flows exist, some of the parameters
might require adaptation to these networks.
Fig. 2.Alternative SAPOR Scheme
F. Modified Scheme
Figure 2 depicts a simple modification of the original
SAPOR scheme. Most of the function blocks are the same
as previously described. The major difference is the location
of the primary routing table. It is situated before the hash
function. The modified scheme also uses separate hash spaces
for all emanating links. The principal functionality of both
schemes is the same. The latter requires more separate building
blocks and can use shorter hash spaces, whereas the former
uses fewer building blocks and requires one larger hash space.
The modified scheme has the disadvantage that it requires a
lookup of the primary routing table every time a packet has
to be routed. The complexity is therefore slightly higher than
that of the original scheme.
One minor disadvantage of SAPOR compared to pure
management efforts is that SAPOR has to be implemented
in existing routers, although it uses methods that are already
implemented in existing routers. Simple changes to software
modules can implement SAPOR. It has to be noted that the
migration is a simple process, since SAPOR is a local router
function and can easily coexist with legacy systems. In fact, the
nodes can be updated one by one. Improvements are already
enabled with the first migrated node.
VI. FURTHER STUDY
Recent years have seen major efforts to develop QoS routing
schemes (e.g. Ying-Da Lin et al. ). The version of SAPOR
that was discussed in this paper does not directly target
different traffic classes, constraint based routing and QoS rout-
ing. Future schemes can support different QoS technologies.
Decisions within the scheme can be further separated and the
routing can be based on the type of service, for example.
Secondly, as indicated above, there are a wide range of
methodologies and schemes available from circuit switched
networks. Further research efforts will have to focus on the
investigation of the applicability of these schemes for packet
switched networks that use SAPOR.