TCP Congestion Control over IEEE 802.11 Wireless Lans based on

K-Means Clustering Focusing on Congestion Window Size and

Round-trip Time

Tomokazu Moriyama, Ryo Yamamoto, Satoshi Ohzahata and Toshihiko Kato

Graduate School of Informatics and Engineering, University of Electro-Communications,

1-5-1, Chofugaoka, Chofu, Tokyo 182-8585, Japan

Keywords: TCP, Congestion Control, Packet Loss Classification, IEEE 802.11 WLAN, K-Means Clustering.

Abstract: Recent IEEE 802.11 wireless LANs provide high speed data transfer using the newly introduced physical and

MAC technologies. Although packet losses over a wireless link are also decreased by the help of new MAC

technologies, some packet losses still occur randomly. Those packet losses invoke TCP congestion control,

which reduces the TCP level throughput, even if congestion does not occur at al. In order to resolve this

problem, some machine learning based approaches have been proposed, which use K-means clustering in

order to discriminate congestion triggered packet losses and wireless error triggered packet losses. However,

those proposals use only delay related parameters, but delay may increase due to non-congestion reasons, in

which case the conventional proposals fail discrimination. This paper proposes a method to classify packet

losses by the K-means clustering focusing on congestion window size and round-trip delay, and to stop

decreasing congestion window when losses are triggered by wireless errors. We develop the proposed method

as a Linux kernel module and show the performance evaluation results that the throughput increases by 40%

without increasing unnecessary packet losses.

1 INTRODUCTION

Recently, the data transfer throughput over IEEE

802.11 WLANs (Wireless LANs) has increased

significantly. The recent IEEE standards, such as

IEEE 802.11n and 11ac, introduced new PHY

(physical) and MAC (medium access control)

mechanisms (IEEE, 2016). The PHY mechanisms

including new modulation methods, MIMO

(multiple-input and multiple-output), and channel

bonding realize high data rate, and the MAC

mechanisms such as frame aggregation and block

acknowledgment (Block Ack) provide not only low

protocol overheads but also powerful data

retransmission capability.

Most of communications over IEEE 802.11

WLAN, such as web access and e-mail, use TCP

(transmission control protocol) as their transport

prptocol (IETF, 1981). One of noteworth functions in

TCP is the congestion control. When congestion

occurs at some nodes within a network, the TCP

module in a data sending node decreases its data

sending rate. However, since TCP works only at an

end node, it cannot detect a precise condition of the

node suffering from congestion. So far, dozens of

congestion control methods have been proposed

(Afanasyev et al., 2010), and most of them consider

that, if there are any packet losses, congestion occurs

somewhere in a network.

When a WLAN link exists within a path between

communicating nodes, the possibility of packet losses

will be larger than a path consisting of wired links

only, even if the recent IEEE 802.11 standards are

used. In such a case, TCP in a sending node considers

that congestion occurs and decreases the congestion

window size unnecessarily. This is a traditional issue

on TCP over wireless links and has been studied

actively (Sardar and Sara, 2006). There are many

proposals, such as modifying TCP, dividing TCP

connections, and support by lower layer protocols.

Recently, there are new trends; a machine learning

approach, i.e., the discrimination of TCP packet

losses by use of machine learning technologies.

Machine learning is a useful method which can be

applied to various fields. TCP communication is one

of the targets and several studies are proposed.

(Nunes et al., 2011) applied the Experts Framework

Moriyama, T., Yamamoto, R., Ohzahata, S. and Kato, T.

TCP Congestion Control over IEEE 802.11 Wireless Lans based on K-Means Clustering Focusing on Congestion Window Size and Round-trip Time.

DOI: 10.5220/0006836800250032

In Proceedings of the 15th International Joint Conference on e-Business and Telecommunications (ICETE 2018) - Volume 1: DCNET, ICE-B, OPTICS, SIGMAP and WINSYS, pages 25-32

ISBN: 978-989-758-319-3

technique to RTT (round-trip time) estimation.

(Mirza et al., 2010) proposed how to predict TCP

throughput using the SVR (support vector regression)

technique. (Chung et al., 2017) applied the random

decision forests to an MPTCP scheduler that selects a

subflow to send data segments by considering

performance metrics such as the MAC data rate,

signal strength, and network congestion.

As for the packet loss discrimination,

(Sooriyabandara et al., 2010) and (Morifuji and

Hiraki, 2013) proposed approaches that a data sender

infers the cause of packet losses by use of the K-

means clustering method (Hand et al., 2001). (Deng

and Cai, 2009) focused on MANET (mobile ad hoc

network) and adopted SVM (support vector machine)

to allow a data receiver to differentiate packet losses.

This paper improves the work done by

(Sooriyabandara et al., 2010) and (Morifuji and

Hiraki, 2013). These two papers focused only delay

related parameters such as one way delay and RTT.

In this paper, we use congestion window size (cwnd)

as well as RTT, and classify two types of packet

losses, i.e. congestion losses and wireless losses, by

use of the K-means clustering. Moreover, we propose

a congestion control method that skips the cwnd

decreasing when a packet loss is classified as a

wireless loss. The rest of this paper is organized as

follows. Section 2 gives some background

information including the overview of IEEE 802.11

WLAN and the TCP congestion control, some

previous comments on packet losses over WLAN, the

overview of K-means clustering, and some related

work. Section 3 describes the proposed method and

Section 4 gives the performance evaluation results. In

the end, Section 5 concludes this paper.

2 BACKGROUNDS

2.1 IEEE 802.11 WLAN and TCP

Congestion Control

As described above, the recent IEEE 802.11

standards, 11n and 11ac, introduce new MAC

mechanisms for high speed and efficient data frame

transmission; the frame aggregation and Block Ack.

The frame aggregation allows multiple data frames

(called MAC protocol data units: MPDUs) to be

aggregated and sent together. The whole transmitted

frame is called A-MPDU (Aggregation MPDU), and

is a collection of A-MPDU subframes, each of which

includes an MPDU delimiter, an MPDU body, and a

padding, as shown in Figure 1. An MPDU delimiter

contains the MPDU length, a cyclic redundancy

check (CRC) to detect bit errors within the delimiter

itself. A padding consists of 0 through 3 bytes, which

make the length of an A-MPDU subframe a multiple

of 4 bytes.

MPDU

delimiter

WLAN

data frame

（MPDU）

PAD

PLCP preamble,

PLCP header

MPDU

delimiter

WLAN

data frame

（MPDU）

PAD

MPDU

delimiter

WLAN

data frame

（MPDU）

PAD

A-MPDU

A-MPDU subframe

PLCP: physical layer convergence protocol,

MPDU: MAC protocol data unit,

A-MPDU: Aggregation MPDU, PAD: padding

Figure 1: Structure of A-MPDU.

The IEEE 802.11n and 11ac standards adopt an

acknowledgment scheme called high throughput

(HT)-immediate Block Ack. When a receiver

receives an A-MPDU, it replies a Block Ack frame

which contains a Block Ack Bitmap parameter

indicating whether it correctly receives a MPDU with

a specific sequence number. The Bitmap indicates

receipt or non-receipt of 64 MPDUs. The data sender

retransmits non-received MPDUs according to the

Bitmap. When a Block Ack frame itself is lost, the

whole A-MPDU is retransmitted by timeout.

TCP uses cwnd in addition with an advertised

window size (awnd) which a data receiver reports in

a TCP header. A data sender transmits data segments

according to the smaller of cwnd and awnd. TCP

Reno / NewReno (Henderson et al., 2012) is a

traditional congestion control method, which is still

used widely. cwnd is controlled in an AIMD (additive

increase and multiplicative decrease) mechanism.

When receiving an ACK segment reporting the

receipt of new data segments (a new ACK), cwnd is

increased by 1/cwnd (segments), and when any data

segments are retransmitted in response to three

duplicate ACKs (fast retransmit), cwnd is halved.

CUBIC (Ha et al, 2008) is a relatively new congestion

control method, which is a default in the Linux

operating system. cwnd increases in a cubic function

of time from the last fast retransmit. At the fast

retransmit, cwnd decreases to 70% (80% in the

original version) of the cwnd value just before the

retransmission.

In those methods, a data sender detects congestion

by a packet loss (retransmission). In TCP Vegas

(Brakmo and Peterson, 1995), on the other hand, a

data sender monitors RTT and estimates the queue

length at a bottleneck node. If the estimated queue

length is smaller than a threshold, a sender increases

cwnd by one segment during one RTT timeframe. If

the queue length is larger than another threshold, a

sender decreases cwnd by one segment during one

RTT timeframe. Otherwise, a sender keeps cwnd as it

is. That is, TCP Vegas is a mechanism based on the

delay.

DCNET 2018 - International Conference on Data Communication Networking

There are some congestion control methods based

on packet losses and delay. TCP Veno (Fu and Liew,

2003) and Compound TCP (Tan et al., 2006) are

examples. They decrease cwnd at a packet loss, and

control the cwnd increase and decrease depending on

the congestion status estimated by RTT. If network is

congested, they work like TCP Reno, and if not

congested, they increase cwnd more aggressively.

2.2 Packet Losses over 802.11 WLAN

As mentioned in the previous subsection, the recent

IEEE 802.11 WLANs provide highly reliable data

transfer by use of the HT-immediate Block Ack

mechanism, compared with the older IEEE standards

that used the one-to-one mapping between data and

Ack frames. However, a few packet losses occur

resulting from the retry-out in the MAC level

retransmission.

(a) Network configuration

(b) Packet loss rate

Figure 2: UDP data transfer over 802.11ac WLAN. (Dianu

et al., 2014).

Figure 2 shows a result of the performance

evaluation of 802.11ac LAN in an indoor

environment given in (Dianu et al., 2014). As shown

in Figure 2 (a), a sender is located at position S and a

receiver is at one of positions R1 through R5. A

sender generates 30 second UDP data traffic using

iperf. Figure 2 (b) shows the result of packet error

ratio in response to the distance between the sender

and the receiver. In the case that the distance is 10.1

m (position R2), there are some packet losses

although the rate is under 1%, which means that

wireless losses happen in an IEEE 802.11ac WLAN.

Figure 3 shows another example of performance

evaluation of 802.11n WLAN, which was conducted

for evaluating Bufferbloat problem (Nomoto et al,

2014). As shown in Figure 3 (a), a station starts

sending data to the server at the far most position for

30 seconds, moves to the nearest position, and stays

there for 30 seconds. Then, it moves to the far position

again. At the far most position, the station uses 6.5 or

13 Mbps data rate, and uses rate close to 300 Mbps at

the nearest position. Figure 3 (b) shows the time

variation of cwnd measured at the station. From time

0 to 40 seconds, and 80 to 120 seconds, that is, while

the station is moving or in the far most position, cwnd

keeps increasing. This means there are no packet

losses. On the other hand, while the station stays at

the nearest position, there are ten drops in the cwnd

graph, each of which corresponds to a packet loss.

Since there are no bottlenecks in the data transfer

from the station to the server, these losses are

considered as wireless losses. This result shows

wireless losses happen in an 802.11n WLAN.

(a) Network configuration

(b) cwnd vs. time

Figure 3: TCP data transfer over 802.11n WLAN. (Nomoto

et al., 2014).

TCP Congestion Control over IEEE 802.11 Wireless Lans based on K-Means Clustering Focusing on Congestion Window Size and

Round-trip Time

2.3 K-Means Clustering

In this subsection, we explain the K-means clustering

that we use to classify TCP packet losses. The K-

means algorithm is a type of unsupervised learning.

The goal of this algorithm is to categorize unlabelled

data into K groups. Specifically, it minimizes an

objective function ∅;

∅

‖





‖



∈







(1)

Here,  is a set of data, 



is a cluster, i.e. a disjoint

subset such that

⋃









 the set of data, and ̅



the cluster center of 



in a Euclidean distance sense.

The algorithm is summarized in the following way.

1. Assign data into K clusters randomly.

2. Calculate the center of each cluster by













∈



(2)

3. Reassign all data  into a new cluster in a way that

a Euclidean distance

‖

̅



‖

is the minimum for

̅



4. Repeat steps 2 and 3 until there are no changes in

clusters or the maximum repeat count is exceeded.

Figure 4 shows an example of the K-means clustering.

A hundred points are selected randomly in the field of

(0, 0) through (100, 100). The above algorithm is

applied once, twice, three times and eight times. As

the repeat count increases, the total of distance

between individual points and the corresponding

cluster center becomes small.

2.4 Related Work

repeat count = 1 repeat count = 2

repeat count = 3 repeat count = 8

Figure 4: Example of clustering by K-means clustering.

In order to classify TCP packet losses as a congestion

loss and a wireless loss, (Sooriyabandara et al., 2010)

uses two delay based parameters, one way delay

(OWD) and inter-arrival time (IAT) of ACK

segments, as data for the K-means clustering. A TCP

data sender keeps a record of OWD and IAT for the

three most recent ACKs, and if there is a packet loss,

that is, a duplicate ACK is received, a sender

classifies this loss event into two groups. A sender

determines this loss as a congestion loss or a wireless

loss depending on the mean OWD of last three ACKs.

If it belongs to a congestion loss category, then a

sender follows the standard TCP back-off procedure,

and if not, cwnd is not decreased. (Sooriyabandara et

al., 2010) shows some performance results by use of

the network simulator ns-2 successfully, but OWD is

difficult to measure in an actual network.

(Morifuji and Hiraki, 2013), on the other hand,

uses RTT for discriminating packet losses. It records

RTT for packets and discriminates the type of losses

based on RTT records. Only if a packet is classified

into the congestion loss, a data sender decreases cwnd.

Through a simulation based performance evaluation,

they confirmed that this method improves TCP

throughput.

3 PROPOSAL

Basically, both of the related work discussed in the

previous section use a delay based parameter for the

K-means clustering. That is, if congestion occurs,

OWD or RTT will increase, and so a packet loss with

a large delay may be a congestion loss in a high

probability. On the other hand, a packet loss with a

small delay might be a wireless loss. However, over

an IEEE 802.11 WLAN link, delay may change for

other reasons. For example, IEEE 802.11 WLAN

uses multiple data rates and the dynamic rate

switching, and when a station is located far from an

access point and the data rate is low, the transmission

delay will increase. Besides, in our previous paper

(Moriyama et al., 2017), we showed that, if there is

an unbalanced traffic load when the multi-user

MIMO is used together with the frame aggregation in

802.11ac, transmission delay may increase. Therefore,

it is possible that the K-means clustering using only

delay based parameters may lead to wrong

categorization.

In this paper, we focus on cwnd itself together

with RTT to apply the K-means clustering, because it

is considered that the probability of congestion

increases when the value of cwnd is large. In the

proposed method, cwnd is decreased as in the original

DCNET 2018 - International Conference on Data Communication Networking

Figure 5: Flow chart of proposed method.

TCP if the data retransmission is classified as a

congestion loss, but if classified as a wireless loss, a

data sender does not decrease cwnd.

Figure 5 shows detailed algorithm flow of the

proposed method. For new ACKs, a sender records

(cwnd, RTT) pair, and the average of three of these

pairs is maintained as y[j]. If a sender receives a

duplicate ACK segment, then it applies the newest

y[j] to the K-means clustering. If y[j] is categorized as

a wireless loss, then the cwnd decreasing is stopped

in the original ACK processing. Otherwise, the

original TCP congestion control is performed.

4 PERFORMANCE EVALUATION

4.1 Experiment Conditions

In order to evaluate the performance of the proposed

method, we implemented it over the Linux operating

system (Ubuntu 16.04LTS). We also implemented

the method described in (Morifuji and Hiraki, 2013),

which uses only RTT for the K-means clustering (we

call this method conventional method). The

maximum number of data used in the K-means

clustering is 10,000 and the maximum repeat count is

set to 50,000.

Figure 6 shows the experimental configuration. A

server, a data sender, is connected to Gigabit Ethernet,

which is connected with an IEEE 802.11ac access

point at the other end. There is one station, a data

receiver, in this WLAN.

We used two scenarios in the experiment. In

scenario 1, 30 msec delay and random packet errors

are inserted at the output port in the server. The

distance between the access point and the station is

about 1 m. The inserted packet loss rate is 0.03%,

0.3%, or 3%. In scenario 2, on the other hand, only 30

msec delay is inserted at the server, and the distance

between the access point and the station is about 7 m.

Server

(sender)

Access point

802.11ac

Station (receiver)

Ethernet

1Gbps

insert 30 msec delay and

random packet errors

〜 1m

(a) Scenario 1

Server

(sender)

Access point

802.11ac

Station (receiver)

Ethernet

1Gbps

insert 30 msec delay

〜 7m

(

)

Scenario 2

Figure 6: Experimental configuration.

In the experiment, the communication duration is

90 sec. We compare the performance of TCP Reno,

the conventional method (Morifuji and Hiraki, 2013),

and the proposed method. For each method, we

executed twenty-five experiment runs for measuring

throughput and number of duplicate ACKs.

4.2 Evaluation Results in Scenario 1

Figure 7 shows the average throughput measured in

scenario 1. The graph shows the results of Reno, the

conventional method (Conv. Method in the figure),

and the proposed method (Prop. Method). Along with

the increase of packet error rate inserted artificially,

Receiving ACK

segment

original

ACK

processing

duplicate ACK ?

i++;

x[i] = (cwnd, RTT)

j>3

j++;

y[j]=(x[i]+x[i-1]+x[i-2])/3

j>10

apply

-means to

y[j] is congestion

no action

stop decreasing cwnd

x[i]: recorded data,

j: number of data

y[j]: average of recent three data

TCP Congestion Control over IEEE 802.11 Wireless Lans based on K-Means Clustering Focusing on Congestion Window Size and

Round-trip Time

the throughput decreases in all the cases. The

throughput is the lowest in the original TCP Reno.

The two K-means clustering methods improve the

throughput. The conventional method provides better

throughput than the proposed method. Figure 8 shows

the number of duplicate ACKs during one

experimental run. This has a similar trend with the

average throughput. TCP Reno is the smallest and the

conventional method is the largest.

Those results mean that the conventional method

is the most aggressive, that is, the conventional

method handles large number of packet losses as

wireless losses and does not decrease cwnd. As a

result, the average throughput becomes high, and

consequently the number of packets sent increases,

which increases the packet losses again.

Figure 7: Throughput in scenario 1.

Figure 8: Number of duplicate ACKs in scenario 1.

4.3 Evaluation Results in Scenario 2

Figure 9 shows the average throughput in scenario 2.

In this scenario, both the conventional method and the

proposed method realize 40% improvement

compared with the original TCP Reno. On the other

hand, as shown in Figure 10, the number of duplicate

ACKs is 3.4 times in the conventional method and 1.5

times in the proposed method, compared with Reno.

This means that, although the throughput

improvement is similar for the conventional method

and the proposed method, the conventional method

induces the increase of packet losses.

The number of transmitted packets increases as

the throughput increases. Along that, the number of

duplicate ACKs also increases. In order to evaluate

the number of duplicate ACKs independently of the

amount of transmitted packets, we define the

congestion index α by the following equation;

α

  

  

(3)

Figure 11 shows the average of congestion index.

From this graph, the proposed method does not

increase the congestion index from the original Reno,

while the congestion index of the conventional

method increases, as twice as the original Reno. This

result means that the proposed method realizes high

performance without increasing unnecessary packet

losses, that is, without deteriorating congestion.

Figure 9: Throughput in scenario 2.

Figure 10: Number of duplicate ACKs in scenario 2.

DCNET 2018 - International Conference on Data Communication Networking

Figure 11: Ratio of duplicate ACKs for total segments in

scenario 2.

In the end of this subsection, we show a detailed

result for individual experimental run for 90 seconds.

Figure 12 shows the time variation of RTT and cwnd

for TCP Reno, the conventional method, and the

proposed method. Red dots in the graph of RTT show

three duplicate ACKs triggering fast retransmit. In the

case of TCP Reno, every three duplicate ACK

decreases cwnd as indicated in Figure 12 (a). Some of

these decreases are wireless loss driven events. In the

conventional method given in Figure 12 (b), cwnd

does not decrease when RTT is small, and so, there

are many chances that cwnd take larger value than

TCP Reno. But, sometimes cwnd does not decrease

even if the value is large (see around 50 seconds and

80 seconds). It is considered that these situations

deteriorate congestion. On the contrary, in the

proposed method given in Figure 12 (c), cwnd does

not decrease while RTT is small and cwnd decreases

when the cwnd value itself is large.

5 CONCLUSIONS

In this paper, we proposed a method to classify TCP

packet losses over IEEE 802.11 WLAN into conges-

(a) TCP Reno

(b) Conventional method

Figure 12: Time variation of RTT and cwnd in scenario 2.

TCP Congestion Control over IEEE 802.11 Wireless Lans based on K-Means Clustering Focusing on Congestion Window Size and

Round-trip Time

congestion losses and wireless losses by the K-means

clustering focusing on both congestion window size

and round-trip time. The proposed method modifies

the TCP congestion control such that if packet losses

are categorized as wireless losses, the congestion

window size does not decrease. We implemented the

proposed method within the Linux operating system

and conducted the performance evaluation using real

WLAN network. The results showed that the

proposed method provides 40% higher throughput

than TCP Reno and that it does not increase the ratio

of duplicate ACKs to the total packets, which the

conventional method focusing only on RTT suffered

from.

REFERENCES

IEEE Std 802.11-2016, 2016. IEEE Standard for

Information technology – Part11: Wireless LAN

medium Access Control (MAC) and Physical Layer

(PHY) Specifications.

IETF, 1981. Transmission Control Protocol, DARPA

Internet Protocol Specification. RFC 793.

Afanasyev, A., Tilley, N., Reiher, P., Kleinrock, L., 2010.

Host-to-Host Congestion Control for TCP. IEEE

Commun. Surv. Tut., Vol. 12, No. 3, pp. 304-342.

Sardar, B., Saha, D., 2006. A Survey of TCP Hnhancements

for Last-hop Wireless Networks. IEEE Commun. Surv.

Tut., Vol. 8, No. 3, pp. 20-34.

Nunes, B, et al., 2011. A Machine Learning Approach to

End-to-End RTT Estimation and its Application to TCP.

In Proc. 20th Int. Conf. on Computer Communications

and Networks (ICCCN), pp. 1-6.

Mirza, M., Sommers, J., Barford, P., Zhu, X., 2010. A

Machine Learning Approach to TCP Throughput

Prediction. IEEE/ACM Trans. Networking, Vol. 18, No.

4, pp. 1026-1039.

Chung, J., Han, D., Kim, J., Kim, C., 2017. Machine

Learning based Path Management for Mobile Devices

over MPTCP. In Proc. 2017 IEEE Int. Conf. Big Data

and Smart Computing (BigComp), pp. 206-209.

Sooriyabandara, M., et al., 2010. Experience with

Discriminating TCP loss using K-Means Clustering. In

Proc. 2010 Int. Conf. Information and Communication

Technology Convergence (ICTC), pp. 352-357.

Morifuji, F., Hiraki, K., 2013. Loss Classification

Algorithm for Enhancing TCP Data Transmission (in

Japanese). In IEICE Technical Report CPSY2013-19.

Hand, D., Mannila, H., Smyth, P., 2001. Principles of Data

Mining. MIT Press.

Deng, Q., Cai, A., 2009. SVM-based loss differentiation

mechanism in Mobile Ad hoc Networks. In Proc. 2009

Global Mobile Congress, pp. 1-4.

Henderson, T., Floyd, S., Gurtov., A., Nishda, Y., 2012.

The NewReno Modification to TCP’s Fast Recovery

Algorithm. IETF RFC 6582.

Ha, S, Rhee, I., Xu. L., 2008. CUBIC: a new TCP-friendly

high-speed TCP variant. ACM SIGOPS Op. Syst. Rev.,

Vol. 42, Issue 5, pp. 64-74.

Brakmo, L., Peterson, L., 1995. TCP Vegas: End to End

Congestion Avoidance on a Global Internet. IEEE J. Sel.

Areas Commun., Vol. 13, No. 8. pp. 1465-1480.

Fu, C., Liew, S., 2003. TCP Veno: TCP Enhancement for

Transmission Over Wireless Access Networks. IEEE J.

Sel. Areas Commun., Vol. 21, No. 2, pp. 216-228.

Tan, K., Song, J., Zhang, Q., Sridharan, M., 2006. A

Compound TCP Approach for High-speed and Long

Distance Networks. In Proc. IEEE INFOCOM 2006, pp.

1-12.

Dianu, M., Riihijarvi, J., Petrova, M., 2014. Measurement-

Based Study of the Performance of IEEE 802.11ac in

an Indoor Environment. In Proc. IEEE ICC 2014 –

Wireless Communications Symposium, pp. 5771-5776.

Nomoto, M., Kato, T., Wu, C, Ohzahata, S., 2014.

Resolving Bufferbloat Problem in 802.11n WLAN by

Weakening M\ac Loss Recovery for TCP Stream. In

Proc. IASTED Int. Conf. Parallel and Distributed

Computing and Networks (PDCN 2014), pp. 293-300.

Moriyama, T., Yamamoto, R., Ohzahata, S., Kato, T., 2017.

Frame Aggregation Size Determination for IEEE

802.11ac WLAN Considering Channel Utilization and

Transfer Delay. In Proc. WINSYS 2017, pp. 89-94.

DCNET 2018 - International Conference on Data Communication Networking