A. and verified that the proper design of selection

A.       Contribution

        The state of art for
wormhole switched networks indicates many selection strategies with different objectives
like resource utilization, priorities based bandwidth reservation or sometimes
holding the use of resources meant for avoiding deadlock. This paper presents a
review of strategies proposed for selecting a single output channel in case
when routing algorithm is returning more than one output paths 3,4,5,6,7. The
path selection in these strategies is based on the congestion status of the
network. Therefore, we have classified the selection strategies according to
the method chosen for representing congestion. Section 2 summarizes some major
studies concerning the impact  of
selection function choice on network performance. Section 3 compares the
selection function in terms of strategies used for representing congestion
information and  discuss several open
issues in the choice of strategies. Finally, section 4 concludes the current
research on output selection function and suggests some future directions.

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!

order now



           Although a
significant performance can be achieved by using an adaptive routing algorithm,
many researches proved that the selection function choice is a significant
parameter for further improving the network performance. In 8 the study
considered various selection functions and verified  that the proper design of selection function
can greatly effect  the average message
latency and the saturation behavior of any adaptive routing algorithm that
relies on routing restrictions to avoid deadlock.


          A similar work is
presented in 9 in which authors performed a set of multi-factor experiments
to study the dependencies between routing policies and communication workloads
and showed that in addition to adaptivity, 
selection function greatly affect network performance under various traffic
patterns. Martinez et. al in 10 analyzed 
and compared several selection functions for irregular topologies in
order to evaluate their influence on network performance. The obtained results
show  the variations in latency  of about 20%, depending on the basic
selection function implemented. The authors also used time threshold on
two  best selection functions and showed
that the latency is reduced at the same time when throughput is increased about
10% for both selection functions. In 11
the author defined, compared and evaluated several selection functions in fat
tree topologies and proved that in order to effectively balance the utilization
links, a selection function plays a crucial role.


         Thus, all the above
reviewed studies show that choice of output channel selection schemes will
affect network performance  i.e
throughput and latency. Therefore, along with the routing function effect of a
selection function must also be considered.




             Although the Selection functions
chosen above make the design  easy to
implement, but they do not consider 
network conditions for path selection. In such cases selection strategy  may likely to select a path that is highly
contented and can  increase  packet latency with a severe drop in
throughput of the network. Therefore the selection function designed must
consider the network status on the downward routers. The choice of metric for
representing network status and closeness of metric to the actual value of
congestion will determine the effectiveness of 
a particular selection strategy. To locate an optimal path for a packet,
different metrics can be used. Based on the metric chosen selection strategies
can be classified as:


A.  Local Adaptive Selection

                 The most
direct method to evaluate an output path for adaptive routing algorithm is from
the local traffic monitoring in which each router analyzes the congestion
status of their immediate neighbors. The commonly used local information
metric  for output channel selection
is  buffer queue occupancy whereby the
length of data queues in FIFO buffer is represented by either multiple bit
signal or single bit signal for buffer status (free or busy). The current
router uses this signal to chose a best output path from the available
alternative directions at any instant as shown in figure 2. For source A to
reach destination K based on free available 
buffer slots, the path selected would be                       


             At each hop router with more
available  free slots in the input port
of  the neighboring router  is selected. In literature many proposals for
adaptive routing algorithm have considered this queue length oriented
selection, such as 12-14. Based on buffer occupancy, these strategies try to
distribute the packets from hot spots to a less congested area in the network.
In some of the other proposals count of free VCs (FVC) in the downstream nodes
is taken as the local selection metric 15. With every clock cycle FVC count
is passed to the neighboring nodes and upstream node considers this count for
channel selection. Although detecting and broadcasting such information to
neighboring routers requires extra logic implementation, these strategies  show lower latency and improved throughput
when compared to congestion oblivious algorithms 9-10.


In a different class of adaptive routing algorithm authors considered
cause of unavailability of free buffer space as the metric  of selection 16- 18. In 16 authors
considered link contention as selection metric to perform traffic prediction.
This selection uses hardware analysis probe for contention measurement that
accounts for the extra cost of the proposal. In CAS 17 concept of fluidity is
taken as congestion measure. A buffer is considered as fluid when a flit is
moving out of it in the current clock cycle. The routers make decisions based
on the downstream fluidity information that is measured and exchanged at
regular intervals between the nodes. The authors in BOFAR 18 proved
Cumulative buffer occupancy time as the more realistic selection than the
instantaneous count of free buffers and the fluidity information of downstream
nodes. In this proposal every router evaluates and exchanged the extra delay
encountered by each flit passing through that router. In port selection
decisions, neighbors with less buffer occupancy history are preferred. The
proposed model incurs slight overhead on