Supporting QoS in the Next Generation Optical Internet 
using Optical Burst Switching (OBS)

Myungsik Yoo & Chunming Qiao, SUNY Buffalo, Buffalo, New York 14260

We consider the next generation Optical Internet where IP runs directly over
an all-optical (or almost all-optical) WDM layer. Like its current incarnation
where IP routers are interconnected with WDM links, such an Optical Internet
can greatly streamline both software (e.g. ATM protocols) and hardware (e.g.
SONET equipment).  Further, having an all-optical WDM layer, which consists of
WDM links and photonic switches/routers, will enable a huge amount of "through"
traffic to be switched, and hence, can not only reduce the required number of
expensive terabit routers and transceivers, but also increase the speed, the
bandwidth and the transparency of the communication pipes. However, for an
Optical Internet to be valuable, one must address, among other important
issues, how the WDM layer supports QoS, which is lacking in the current IP
(that provides only best-effort services). Such a WDM layer supporting basic
QoS (e.g. priority) will not only facilitate but also be complementary to a
future QoS-enhanced IP (e.g. one that provides differentiated services). In
addition, it is useful for certain WDM layer traffic for signaling and
restoration which may require a higher priority than ordinary traffic.          

Previously, we have proposed the concept of Optical Burst Switching (OBS) and
in particular, a high-speed, one-way reservation protocol called
Just-Enough-Time (JET) to efficiently support bursty traffic in optical
networks. Here, we propose a novel priority scheme that takes advantage of the
two features of JET-based OBS, namely, delayed reservation (DR) and the use of
offset time, to provide efficient support for QoS in the next generation
Optical Internet. Unlike existing priority schemes (e.g. Fair Queueing and its
variations) which buffer the packets (or cells) using different queues to
achieve different levels of services, our priority scheme does not mandate the
use of any buffer at the intermediate nodes, and hence is especially suitable
for the WDM layer as the technology for optical buffer (e.g. fiber delay lines
or FDLs) is not as mature as the electronic counterpart.

In this work, we compare two OBS protocols, namely, classless OBS and
prioritized OBS, respectively. In classless OBS, all bursts are treatly
equally, and a control (set-up) packet is sent, followed by the corresponding
burst after a base offset time to compensate for the processing delay
experienced by the control packet. In prioritized OBS, we first consider two
classes of services: 0 and 1, respectively. Class 0 corresponds to best-effort
services and can be used for non-real-time applications, while class 1
corresponds to priority services and can be used for real-time applications.
A burst receiving Class 1 services (called a Class 1 burst) is given a higher
priority than a Class 0 burst when reserving bandwidth to ensure that the
Class 1 burst incurs a lower blocking (dropping) probability.  This is
accomplished by simply using an extra offset time between a Class 1 burst and
its corresponding control (or set-up) packet (but no extra offset time will be
used when sending a Class 0 burst). The extra offset time allows a control
packet for a Class 1 burst to make more advanced bandwidth reservation, thus
having a greater chance of success, than the control packet for a Class 0
burst. As an example of how Class 1 bursts can be effectively separated from
Class 0 bursts by using a reasonably small extra offset time, assume that the
length of Class 0 bursts has an exponential distribution with an average of L
(e.g. 1 ms). Since at least 99% of the Class 1 bursts are shorter than 5L, at
most 1% of the Class 0 bursts may block a Class 1 burst which uses an extra
offset time of 5L. 

We derive the upper and lower bounds on the blocking probability of each class
in prioritized OBS. Our analysis and simulation show that the prioritized OBS
can efficiently support different levels of services. In particular, Class 1
bursts can achieve much lower blocking probability than Class 0 bursts, while
the average blocking probability over all bursts remains the same as that of
classless OBS. Note that prioritized OBS can also be extended to support more
than two classes of services by using different amount of extra offset time
for each class.