Fig. 1. Net spectral efficiency (NSE) versus SNR for various PDM-QAM
formats. As the available SNR decreases, the net spectral efficiency also
decreases requiring a higher symbol rate for a given data rate.
the bit error rate (BER) as a function of SNR. Rather than
assume soft decoding we conservatively use the hard decision
decoding bound for the binary symmetric channel3 such that
if pb = BER, then the maximum code rate r is given by [10]
r = 1 + pb log2 (pb) + (1 − pb) log2 (1 − pb) (5)
from this the NSE as a function of SNR can be obtained
for a given cardinality of QAM to give Fig. 1. We note that
for a terrestrial core optical transmission network the SNR
will typically be in the region of 5 to 25 dB for the fiber
and amplifier parameters previously discussed corresponding
distances ranging from approximately 100 to 10000 km.
Over the region of interest shown in Fig. 1 the NSE that
can be realized with PDM-QAM and optimal hard FEC can
be approximately bounded by the following expression4
NSE ≈ 2 log2
_
1 + SNR
_ 210 + 9SNR
325 + 22SNR
(6)
If the optimum launch power spectral density is used then
the SNR is uniquely defined by the route though the network.
Hence knowing the SNR then using the approximate realizable
bound (6) this then defines the appropriate amount of spectrum
that should be assigned in an elastic network.
VI. ROUTING ALGORITHMS
We consider our benchmark as shortest path (SP) routing
with first fit allocation of the optical spectrum, and two
congestion aware (CA) variants of shortest path routing that
are:
CA1. Selects the shortest path that avoids the fiber link
that is most congested, implemented with Dijkstra’s
algorithm on the graph where the edge weight for the
most congested path has been replaced by infinity.
CA2. The shortest path through a weighted network, where
the weight of an edge joining nodes i and j is given
by Wi j = Li j /ηi j where Li j is the physical length
and ηi j is the proportion of the total spectrum which
is still available on that edge.
Since the system operates with a constant power spectral
density the spectral usage is proportional to the total optical
3We note that the latest state of the art transceivers employing soft FEC
are close to this bound, e.g. with a 15% FEC overhead, hard FEC limit is a
BER = 0.018, c.f. BER = 0.019 reported for an implemented soft FEC [12].
4For SNR from −30 to 50 dB this expression, being a minimax fit with a
Pade approximant, gives the NSE with an accuracy of better than 5%.