experience similar channel conditions (from the AP) and use
the same physical rate of 36 Mbps (802.11g) for their trans-
missions. For simplicity, we assume i.i.d. transmission fail-
ures, for all devices, i.e., any transmission from either the AP
(with large buffer) or devices, fails with a constant known
transmission failure probability. Note that these transmission
failures reduce the total attainable throughput of the system.
Furthermore, when devices are contending for the right to
transmit, the achievable throughput is also affected by the
channel access method. In particular, in addition to wireless
channel errors causing transmission failures, the random
access in 802.11 networks leads to collisions and reduced
throughput when more than one device contends for access
to the shared channel. Our overall finding from the ns-3
experiment is summarised in Table 1.
When d
1
and d
2
start TCP connections at the same time and
use 802.11 WiFi under a given wireless channel error proba-
bility of 0.3, their data rates after 20 seconds are, respectively
2.46 and 17.9 Mbps for the standard 802.11 MAC settings
(recall that the physical rates are 36 Mbps for both). It can be
seen that d
2
enjoys better throughput than d
1
via the same
AP. The problem is summarized in Table 2 (see Section 4 for
details). Note that simply splitting the AP buffer (separate
Buffering) for each device equally, similar to that in [12], [16]
under the considered scenario would yield rates of 5.65 and
14.71 Mbps after 20 seconds.
Remark 1.1. For the motivating example (Table 1), we later
show that employing the AC algorithm (proposed in this
paper) at the AP, then packets from d
1
and d
2
are admitted
differently into the AP and equal data rates (10.19 Mbps)
can be achieved by both IoT devices after 20 seconds. This
is known as balanced rate allocation in this paper.
1.4 Approach and Concept
We discover (and propose) an adaptive admission policy for
packets from multiple IoT flows entering into the shared AP
buffer to ensure low loss and consistent low delay. The con-
trol policy concurrently measures the arrival and departure
rate of packets belonging to each flow so as to ensure desired
sharing (e.g., equal/fair sharing) of buffer space by the pack-
ets belonging to each flow. With the policy, aggressive tradi-
tional Internet applications with bulk TCP flows can
perceive equal throughputs as those of small IoT-based TCP
flows. The mathematical interpretation of our adaptive admis-
sion control (AAC) approach is discussed further in Section 3.
Our approach in transient modeling of the behaviour of
IoT scenario (recall Fig. 1) and its traffic dynamics provides
important insights in developing the aforementioned AAC
approach. For transient analysis, we adopt a pragmatic
approach to model the IoT network traffic as fluids that flow
from sources to destinations and then capture the underlying
TCP congestion control (Section 2.1.1) as feedback dynamical
systems, which is explained in the following.
While modeling our IoT communication system, the exis-
tence of deterministic (evolution of TCP congestion windows)
and other non-exponentially distributed parameters such as
MAC transmission failures, transmission of fixed length pack-
ets etc. gives rise to stochastic modules that are nonMarkovian
in nature. Such nonMarkovian models can be Markovized
using phased-type approximation. However, phased-type
expansion increases the already large state space of our system
model. Furthermore, it becomes really intractable (severe),
when mixing deterministic times with exponential ones, in
our modeling. Therefore, an alternative idea is developed to
analyze the internal process model of our IoT communication
system (nonMarkovian), which can be shown to be a Markov
regenerative (semi-regenerative) process (see Markov Regener-
ative Reward [17] and its applications in [18], [19], [20], [21]).
1.5 Contributions
Our main contribution in this paper is the following.
C. We develop a novel adaptive admission control framework
for accommodating IoT flows over home WiFi networks,
under realistic channel conditions.
To prove our idea, we develop a mathematical framework
that accurately captures the transient dynamics and the
throughput unfairness perceived by uplink and downlink IoT
devices over an infrastructure 802.11 WiFi, under an errone-
ous wireless channel and finite (droptail queuing) AP buffer.
Often, it is the last-mile WiFi access network that limits the
overall bandwidth achievable by a device using a TCP flow.
A TCP segment sent by the device may suffer loss either due
to buffering at the AP or due to collisions and impairments
on the shared WiFi medium. Such losses interact with TCP’s
feedback mechanism in a complicated way which is best
captured by a packet-level discrete stochastic model (see
Section 2.3.2). Furthermore, for short-lived TCP flows, one
must study the transient rather than the steady-state dynam-
ics which is best captured by a fluid-like continuous model
(see Section 2.1.1). These two sub-systems (which represent
the ‘loss process’ and the ‘arrival process’) interact through
the ‘queue occupancy process’ of different types of traffic
(see Section 2.2). Therefore, our three main sub-contributions
are as follows.
TABLE 1
Motivating Example: Achievable Data Rates
Physical Rates
Effective Rates
AP 802.11 only Equal Splitting
d
1
36 Mbps 2.46 Mbps 5.65 Mbps
d
2
36 Mbps 17.9 Mbps 14.71 Mbps
TABLE 2
Network Parameters and Protocols
Parameters Value
PHY rate (802.11n on 2.4 ghz) 72.2 Mbps
PLCP premble 144 ms
Slot Time 20 ms
DIFS time 50 ms
SIFS time 10 ms
EIFS time 308 ms
Min. Contention Window 31
Max. Contention Window 1023
Retransmission limit 7
TCP version NewReno
TCP Header 20 bytes
TCP ack size 40 bytes
data size 1460 bytes
POKHREL ET AL.: ADAPTIVE ADMISSION CONTROL FOR IOT APPLICATIONS IN HOME WIFI NETWORKS 2733
Authorized licensed use limited to: SUN YAT-SEN UNIVERSITY. Downloaded on May 08,2022 at 13:32:32 UTC from IEEE Xplore. Restrictions apply.