@article{Rosberg2011Flow,
author = {Zvi Rosberg and Ji Li and Fan Li and Moshe Zukerman},
title = {Flow Scheduling in Optical Flow
Switched ({OFS}) Networks under Transient Conditions},
journal = {IEEE/OSA Lightwave
Technology},
volume = {29},
number = {21},
pages = {3250-3264},
month = {Nov.},
year = {2011},
abstract = {Optical
flow switching (OFS)
has been recently introduced as a potential ¡°green¡± architecture addressing the
power issue of store-and-forward packet switching
in future MAN-WAN Terabit
networks. One key architectural component of OFS
differentiating it from other ¡°green¡± WAN architectures such as
optical circuit switching
(OCS), optical packet
switching (OPS) and optical burst
switching (OBS), is its centralized
flow scheduling.
Comparing the theoretical network capacity regions
of OFS, OCS, OPS and OBS has revealed that the
dominating theoretical capacity depends on the hardware as well as on the port
configuration. The dominating actual capacity (throughput) that can be achieved
also depends on the flow schedulers supported by
each architecture. Since centralized scheduling
incorporated in OFS is
the least restricting between all scheduling
methods, OFS is a promising ¡°green¡± architecture
option for future MAN-WAN Terabit networks. For
better understanding the actual potential throughput of
OFS, we study its scheduling problem
in a realistic traffic model where lightpath
requests arrive as a time-dependent Poisson process with Pareto distributed
lightpath service times. Lightpath schedules are taken at fixed time intervals
(larger than 100 ms) in a central node and
flows that have already been scheduled cannot be
interrupted before their completion. The scheduling
problem is represented as a discrete-time Markov decision process where the
objective function is given by the flow blocking
probability over a finite time horizon. We derive three lower bounds to the
objective function and propose several schedulers, with and without fairness
requirements. The performance of our OFS schedulers
are evaluated under both static and limited dynamic
routing, by emulating the algorithms on random network
topologies for two hours. The main result is that our proposed max-min fair
scheduler with limited dynamic routing significant- - ly outperforms all other
schedulers with static routing. Furthermore, its blocking probability is close
to the lower bound for static routing.}
}