Ad Hoc Networks : Technologies and Protocols - University of ...

TCP-aware Cross-layered Solutions 143
currently being used. This would prevent the retransmission timers at
the TCP sender from firing out during route computations and leading to
a degradation in throughput or in the worst case resulting in connection
stalls due to repeated timeouts. The prediction mechanism is merely a
heuristic and can fail either by predicting link failure wrongly or failing
to predict an actual link failure. In the first scenario, the throughput of the
corresponding connection is left unaffected since the source will continue
to use the current path until a new alternate path is computed or the current
path fails. Since the current path will not fail, the source will switch its
connection only upon the recomputation of an alternate path. Further, if
the current path is the best path, it will again be recomputed as the alternate
path thus preventing any sub-optimality because of wrong predictions.
The drawback of wrong predictions is the route recomputation overhead
that is incurred. On the other hand, if a route failure is not predicted
successfully, the performance of the connection will be only as bad as
the scenario wherein no prediction mechanism is employed.
Finally, the prediction mechanism will successfully predict only mobility
related route failures. Other possibilities such as congestion based route
failures will not be captured by the prediction mechanism and will trigger
normal route errors as usual.
Proactive Route Errors (PRE):
If a link failure occurs either due to congestion, or due to mobility but
has not been successfully predicted by the prediction mechanism, the
third mechanism in Atra tries to minimize the latency involved in the
route failure information being carried to the source(s) using the link of
failure. In the default set-up, DSR will issue a route error to only the
source of the packet that triggered the link failure detection at the MAC
layer. If multiple sources are using the same link, packets will have to
arrive from them at that node before route errors will be sent back to
them increasing the latency between the link failure detection and the
time at which the sources are informed. This latency is further inflated
because subsequently arriving packets from other sources will have to
go through the MAC failure detection time cycle before the link failure
is inferred. However, in Atra each node maintains a cache of the source
identifiers of TCP connections that have used a particular link in the past
T seconds. When a link failure is detected, all sources that have used
the link in the past T seconds are informed about the link failure through
normal route errors. This reduces the latency involved in the route failure
information delivery which consequently reduces the number of losses
and also triggers earlier alternate route computations.