OPTICAL BURST SWITCHING OBS 249
a Immediate setup with timed release b Immediate setup with explicit release
Time Time Control
packet Offset
Burst Control
packet Offset
Burst Release
packet A
B B
A
Figure 10.8 Immediate setup with timed or explicit release.
Control packet
Offset Burst
Time
A B
Figure 10.9 Delayed setup with timed release.
10.2.3 Scheduling of Bursts at an OBS Node
Depending on the scheme used to reserve and release OXC resources, there may or may not be a need for a scheduling algorithm. Let us consider first the case of immediate setup
with explicit release. Assume that an arriving control packet requests a connection setup on a specific wavelength i of output fiber j . Assuming that this wavelength is free, the
control unit immediately configures the OXC, and then forwards the control packet to the next OBS node. Wavelength i on output fiber j remains reserved until the control unit
receives a release message. During that time, the control unit cannot schedule any other bursts on this wavelength. Only after it has received the release message it can accept a
new request for the same wavelength. This simple scheme is fairly easy to implement, which means that the control unit can decide very quickly whether it can schedule a burst
on a wavelength of a given output fiber. Fast processing time of the control packets is critical to the network’s throughput. Therefore, the time to process a control packet should
be such that the control unit is not the bottleneck in the OBS network. For instance, if it takes 1 msec to process a control packet, then it can process 1000 control packets per
250 OPTICAL BURST SWITCHING
Control packet
Burst t
1
t
2
t
3
t
4
Figure 10.10 The delayed setup scheme.
second, which might be a lot less than the number of bursts that the OXC can switch. That is, the control unit imposes an upper bound on the number of control packets it can
process which in turn limits the throughput of the OBS node. Obviously, the control unit has to be able to process more control packets than the number of bursts that the OXC
can switch.
Consider the case of immediate setup with timed release. In this case, the control unit knows when a wavelength of an output fiber will become available. Consequently, it can
accept a new burst as long as the offset of the burst is such that it will arrive after the wavelength becomes free. For instance, if the wavelength becomes free at time t
1
, and the burst is due to arrive at time t
2
, where t
2
t
1
, then it will accept the new burst as long as there is enough time to configure the OXC between t
1
and t
2
. Now consider the case of delayed setup. In this case, the OXC is not configured until
just before the burst arrives. Let us assume that the control packet arrives at time t
1
see Figure 10.10. At time t
2
, the control unit has finished processing the control packet and the burst is not due to arrive until time t
3
. Depending upon the length of the gap t
3
− t
2
, it might be possible to squeeze one or more bursts in-between t
3
and t
4
. This technique is known as void filling. As mentioned above, the offset is a function of the number of OBS
nodes along the path. If there are many nodes, then the time between t
3
and t
4
might be large, and consequently we might be able to do void filling when the node is close to the
end device. However, this time gets shorter as we move away from the transmitting end device, which implies that void filling might not be feasible see Figure 10.9. Finally,
because of the timed release scheme, a new burst can be always accepted as long as its offset is such that it will arrive after the wavelength becomes free.
So far, we have assumed that there are no converters. Therefore, an incoming burst on wavelength i has to switched out on the same wavelength. Assuming full or partial
conversion, then the scheduling of burst gets more complicated, since now it is possible to schedule the burst on another free wavelengths. Let us assume full conversion. That is, there
are as many converters as the number of wavelengths. In this case, a burst can be always switched out as long as there is a free converter. This assumes that a converter can convert
the signal on an incoming wavelength to any other wavelength. This might not always be the case. That is, a converter might only be able to convert an optical signal on a wavelength
to other wavelengths that are within certain nanometers. In the immediate setup with timed or explicit release, scheduling a burst whose wavelength is not available simply involves
choosing one of the free wavelengths. In the case of the delayed setup with timed or explicit release, the scheduling algorithm becomes complex and CPU intensive.
10.2.4 Lost Bursts
Burst loss is an important performance measure of an OBS network, and it occurs when the on-the-fly connection setup is used. When a control unit receives a control packet, it
THE JUMPSTART PROJECT 251
extracts the burst’s destination from the packet and it then makes a decision as to whether its OXC can switch the burst or not. If it cannot switch the burst, it will not forward the
control packet to the next OBS node, and it will send a negative acknowledgment back towards the transmitting end device in order to release any resources already allocated.
However, the end device might have already transmitted the burst by the time it receives the negative acknowledgement, which means that the burst will get lost upon arrival at
the node.
A small number of FDLs can be used for reducing the burst loss rate, but, as with optical packet switching, FDLs are not commercializable. An alternative solution is to use
converters. In this case, if a burst is coming into an OXC on a wavelength which is in use at the destination output fiber, the burst can be converted to another free wavelength.
Finally, deflection routing can be used to divert a burst that would otherwise be lost. The alternative path, however, might have more hops than the original one and therefore
the offset might not be long enough to prevent burst loss. A possible remedy is to delay the deflected burst in an FDL so that its offset from the control packet is sufficiently long.
Another solution is to use an offset which is the upper bound of all offsets within the OBS autonomous system. This simple solution avoids the need for FDLs, at the expense
of delaying the transmission of some of the bursts more than necessary. Finally, deflected bursts impose an extra load on the alternative links, and this has to be taken into account
when planning the capacity of the OBS network.
10.2.5 Burst Assembly
Each end device maintains one queue per destination end device. Packets arriving at the end device are placed accordingly into the various destination queues, from where they
are transmitted over the OBS network in bursts. The size of the burst is controlled by a timer. When the timer associated with a destination queue expires, all of the data in
the queue are grouped into a single burst, which is then transmitted out according to the OBS scheme. The burst size has to be less than a maximum since large size bursts tend
to occupy the network resources for long periods of time thus blocking the transmission of other bursts. Also, the size of the burst has to be greater than a minimum because of
the signaling overheads for setting up a connection. Therefore, when the timer expires, a burst is assembled if its size is greater than the required minimum. Also, a burst can be
assembled before the timer expires if the data in the queue reaches the maximum size. The duration of the timer, and the maximum and minimum burst sizes can be estimated
using modelling techniques, such as simulation.
The end device can also introduce priorities when transmitting bursts. Specifically, each destination queue can be further subdivided into multiple QoS queues. The arriving pack-
ets are grouped into these queues, which are served according to a scheduler. In addition, different timers and maximumminimum burst sizes can be used for different queues.
10.3 THE JUMPSTART PROJECT
The Jumpstart project was carried out by MCNC, a non-profit research organization in the Research Triangle Park, North Carolina, and North Carolina State University, and it was
funded by the Advanced Research and Development Activity ARDA. The objectives of the Jumpstart project were to a define a signaling architecture for an OBS network, and b
