TCP Response to Explicit Link-Error Notifications
Ilpo Järvinen and Aki Nyrhinen

Explicit Link-Error Notifications (ELEN) allow TCP to distinguish
between link-error related and congestion losses whose disambiguity
results in suboptimal performance in wireless environment because
TCP is forced to respond to a non-existing congestion. In this demo,
we show that ELEN perceivably improves TCP performance in a wireless
environment that exhibit error-prone characteristics as TCP is able
to sustain good throughput.

The origin of TCP dates back multiple decades, since that time, it
has needed multiple amendments to cope with the problems that were
unknown at that time. Introduction of wireless access like it is today
adds losses due to link errors that often occur on wireless link because
of corruption and blackouts to this ever increasing list of problems.
TCP congestion control, on the other hand, is rooted were deeply to the
assumption that every loss means congestion on the path because in wired
environments bit errors are rare to occur. If network gives false
congestion signals often enough, TCP backs off unnecessarily and its
window becomes a limiting factor in network utilization. To counter
such suboptimal throughput, we have developed a logical link layer
protocol that detects losses that occur on the wireless links and then
sends an Explicit Link-Error Notification (ELEN) about them to TCP
layer using a cross-layer mechanism. Our ELEN-aware TCP is able declare
a recovery non-congestion related and cancel out congestion control
measures that were taken when a loss was detected if ELENs inform
that all losses were due to link-errors. 

Our demo setup is as follows. Web browser located on a mobile terminal
has an access to the Internet over an emulated high delay bandwidth
product wireless link. The wireless link provides ELEN upwards in the
local network stack to the sending TCP when link-error related losses
occur. The browser uses a HTTP proxy that is located in the access
network very close to the wireless link. We believe this scenario
represent a good candidate for ELEN deployment because ELEN-aware
link-layer setup can then be made local to an operator. During the
demo, the mobile terminal occasionally moves out-of-coverage of
current connectivity causing a link outage to occur. The link outage
is modeled by three-state Markov error model. A good state represents
perfect connectivity. Before the full link-outage state, the link
quality decreases considerably causing a short bad quality period
during which many losses occur, yet some data is still getting
through. After the link outage, full quality of connectivity is again
restored. Such error model roughly represent what happens, for example,
during a handover.

We present the ELEN effect with a TCP data flow that is relatively
long because then the benefits are clearly visible to human eye. From
statistical point of view, the improvements are present in shorter
TCP flows too but the time scale is not suitable for unaided eye and
other factors can affect more in an individual transfer (e.g.,
slightly different loss pattern during initial slow start). Both
Baseline and ELEN TCP are run simultaneously on different computers,
yet, the states of the error model are synchronized between them
using pre-generated state length distributions.