Throughput Performance Analysis of the Multipath [601668]

Throughput Performance Analysis of the Multipath
Communication Technologies for the Cloud

SZILÁGYI Szabolcs , BORDÁN Imre, HARANGI Lajos, KISS Benjámin

University of Debrecen , Hungary ,
Department of IT Systems and Networks , Faculty of Informatics ,
26 Kassai Way , 4028 Debrecen , Hungary , E-Mail: [anonimizat]

Abstract – Cloud based computing is among the top
present -day research areas. Datacenters serving as the
backend for cloud solutions have to satisfy the demands
set by the time -critical applications emerging in our
rushing world. In order to provide physical redundancy,
the datacenters are equipped with redundant inter –
server connections that are left idle du ring normal
operation via the traditional TCP protocol, e.g. with
regards to link -capacity aggregation. Our paper
provides a comparison between two technologies
(MPTCP and MPT -GRE) that support multipath
operation and could prove useful in datacenter
envir onments by increasing efficiency and enhancing
user experience in the cloud.

Keywords: cloud network ; data center; MPT -GRE ;
MPTCP; multipath communication ; performance
analysis ; system throughput ; tunneling .

I. INTRODUCTION

Cloud based computing is one of the most popular
areas in IT today, enabling business processes, software,
platforms and infrastructures through its flexible
capabilities ( BPaaS : Business Process as a Service,
SaaS : Software as a Service, PaaS : Platform as a Service,
IaaS : Infrastructure as a Service). Since numerous
applications are communicating through the cloud, cloud
networking has a direct impact on user experience.

Core switches
Aggregation
switches
Access switches10 Gbps
1 Gbps
Host servers
10 Gbps

Fig.1. Datacenter topology example

In this paper, we present measur ements comparing
two technologies enabling multipath communication that
can greatly help in providing a more efficient operation in
datacenter environments, while also enhancing the user
experience. The MPT -GRE software library [1], capable of
effective link -aggregation even in quad -path Gigabit
Ethernet environments, was developed at the University
of Debrecen. In the current scope, we examine it in a dual –
path 10 Gigabit Ethernet measurement environment,
providing a comparison with Multipath T CP (MPTC P)
[2] from a maximum throughput and resource usage point
of view.
In the second chapter, we give a short description
about the operating principles of these multipath
solutions, followed in the third chapter by a presentation
of our measurement en vironment. Chapter four
showcases the results of the experiments, before we get
to the final chapter to draw a conclusion and touch on
further development opportunities .

II. MPTCP AND MPT -GRE IN A NUTSHELL

Numerous multipath solutions operating in different
layers of the computer networking model exist [3];
however, in our view, MPTCP is definitely to be regarded
as the flagship of this area. MPTCP was standardized in
2013 [4]. Since then, renowned network device
manufacturers and corporations devel oping operating
systems have integrated it into their own products (e.g.
Cisco, Apple) [5].
Practically speaking, MPTCP is a multipath extension
of the traditional TCP, realized via the use of so -called
TCP-subflows. This approach enables each physical
interface to be assigned a subflow that is responsible for
data transfer through that given interface. Earlier
publications have shown MPTCP to be quite effective
with respect to link -capacity aggregation (see e.g. [6]-
[8]). The following figure shows the lay ered architecture
of MPTCP:

+––––––––––––––– +
| Application |
+––––––––––––––– +
| MPTCP |
+–––––––- +–––––––- +
| TCP subflow | TCP subflow |
+–––––––- +–––––––- +
| IPv4/IPv6 (Physical) | IPv4/IPv6 (Physical) |
+–––––––- +–––- –––– +
Fig.2. MPTCP layered architecture

The MPT -GRE software started development in 2012 at the
Faculty of Informatics, University of Debrecen, and has since

gone through numerous versions [9]-[10]. This solution also
showed effective performance in earlier publications on
multipath environments [11]-[15]. Its operation differs from that
of both the traditional TCP and the MPTCP. It enables multipath
interface mapping via the introduction of a logical tunnel
interface. The operation of the interface as discernible to the
applications above the tunnel l ayer is completely unchanged.
However, beneath this logical layer, the mapping to physical
interfaces happens via utilizing the multipath GRE tunneling
technology. The following figure shows the MPT -GRE
architecture:

+–––––––––––– ––– +
| Application (Tunnel) |
+––––––––––––––– +
| TCP/UDP (Tunnel) |
+––––––––––––––– +
| IPv4/IPv6 (Tunnel) |
+––––––––––––––– +
| GRE -in-UDP | + ––+
+–––––––- +–––––––- +<–| MPT |
| UDP (Physical) | UDP (Physical) | + ––+
+–––––––- +–––––––- +
| IPv4/IPv6 (Physical) | IPv4/IPv6 (Physical) |
+–––––––- +–––––––- +
| Network Access | Network Access |
+–––––––- +–––––––- +
Fig. 3. The layered architecture of MPT -GRE

III. MEASUREMENT ENVIRONMENT

We have performed our measurements using the dual –
path 10 Gigabit Ethernet environment detailed below.
The two servers had the following specifications :

• Gigabyte Z77 -D3H motherboard with Intel Z77
chipset .
• Intel Core i7 -3770K 3.50 GHz processor with 4
cores and 8 threads .
• 4 X 4 GB 1600 MHz DDR3 SDRAM .
• Intel dual 10 Gigabit Ethernet server adapter .
• Ubuntu 16.04 LTS (XenialXerus) 64 -bit operating
system with 4.4.0 -62-generic Linux kernel module .

Server 1 Server 2enp4s0f0 192.168.1.1/24
tun0 tun0
10.0.0.1 /30 10.0.0.2 /30Tunnelenp4s0f1 192.168.2.1/24enp4s0f0 192.168.1.2/24
enp4s0f1 192.168.2.2/24

Fig. 4. Measurement environment

To establish a direct connection between the servers,
we used two 10 Gigabit Ethernet cables (HP X240 10G
SFP+ SFP+ 3m DA cable) and two 10 Gigabit Ethernet
Intel Server dedicated network interface cards with two
ports each. The motherboard’s integrated NIC was used
for remote management purposes, and it was always
disabled for the entire duration of the measu rements. For
the MPTCP tests, we installed the latest, 0.95 version (see
[2]). The MPT -GRE measurements were carried out
using the version available from our development
website. All experiments were run on the Linux Ubuntu
16.04 LTS distribution. The effectiveness of both
multipath technologies was examined through iperf3 ,
CPU utilization and FTP -based measurements. We wrote Python -based bash scripts to automate the process, and
repeated each measurement run ten times .

IV. MEASUREMENT RESULTS

A. iperf3 -based Measurements

We began with the iperf3 measurements. No physical
file download took place during these runs, as the direct
data reads and writes were performed between the
memories of the two servers. This got rid of the
bottleneck of hard drive read/write speed constraints. The
results were logged using a program named tee.
We performed the different measurements utilizing
one interface, and then utilizing two interfaces as well. As
it can be seen on Figure 5, increasing the number of
interfaces in the 10G environment did not always go hand
in hand with an obvious increase in throughput. Using a
single interface with MPT -GRE, we managed to reach a
throughput of 4.14 Gbps, while using two interfaces
resulted in 3.83 Gbps. This is progress compared to
earlier measurements (see e.g. [ 11]-[15]), but does not
take advantage of the potential offered by the applied
technologies.

Fig. 5. Testing the MPT and MPTCP iperf3 throughput
performance

The MPTCP achieved noticeably better results with
maximum speeds of 9.22 Gbps and 18.4 Gbps
respectively. However, in contrast to earlier results, we
can notice that MPT did not produce its fastest results
during measurements ran in a purely IPv4 configuration,
but instead when operating with an IPv4 tunnel over IPv6
physical paths.

B. FTP Measurements

The next round of scenarios involved FTP -based
measurements. The server on the left of Figure 4. was set
up as an FTP server that was used to download a 10GB
file to the server on the right.
We have performed the ramdisk configuration using
the following script:

sudo moun -t tmpfs -o size=11G tmpfs
/var/ftp/pub/

cp /var/ftp/10GB.zip /var/ftp/pub/

The FTP download process itself was automated as
follows:

!/bin/bash
#HOST="[fec0::2]"
#HOST="10.0.0.2"
#HOST="192.168.1.2"
HOST="[fec1:300::2]"
wget ftp://$HOST/pub/10GB.zip -O /dev/null
– -report-speed=bits 2>&1

Figure 6. shows the performance ach ieved by utilizing
two aggregated paths. We can similarly note in this case
as well that the IPv4 tunnel operating over IPv6 paths is
the most effective configuration when it comes to MPT
results. These settings allowed us to reach 2.97 Gbps
during the dow nload of the 10GB file. The other
configurations achieved maximum speeds of 2.94 Gbps,
2.83 Gbps, and 2.84 Gbps. The MPTCP performed better
in these runs as well, but the difference was not as big
compared to MPT results as in the case of the iperf3
measur ements: the maximum speeds here were 5.93 Gbps
over IPv4, and 6.13 Gbps over IPv6.

Fig. 6. Wget download speed comparison of the MPT
and MPTCP using 2 x 10 GE interfaces

We continued with further comparisons of MPT and
the MPTCP taking a look at FTP download speeds. First,
we performed a baseline measurement of network
performance with both MPT and MPTCP disabled. The
result we got was 8.78 Gbps, which equated to a
download time of 9.11s for the 10GB test file (see Figure
7.).

Fig. 7. The MPT -GRE I Pv4-IPv4 FTP throughput
performance using two 10 Gigabit Ethernet interfaces However, increasing the number of paths did not
bring increased throughput figures. This is discernible
during the usage of MPT and MPTCP as well. Using
MPT we achieved 2.94 Gbps , resulting in a 27.2s
download time. This is close to a threefold increase
compared to the baseline measurement.
The MPTCP results took a similar shape (see Figure
8.). In this case, the throughput was 5.93 Gbps, taking
13.49s to transfer the test file us ing two paths. This means
a one and a half times increase in transfer time compared
to baseline network performance .

Fig. 8. MPTCP IPv4 FTP throughput performance
using two 10 Gigabit Ethernet interfaces

C. CPU Utilization

Finally, we examined the CPU resource demand of
the MPT and the MPTCP solutions. Figure 9. shows the
most critical corner case, namely the CPU performance
figures we experienced while operating over IPv6.
Naturally, in the other less taxing cases the results were a
bit better; how ever, to keep the paper tidy, those are not
presented here.
We can see that CPU utilization per the applied
technology roughly reflects the achievable throughput
speeds. Processing the data transfer using MPT resulted
in a higher average of 23.3% and 23.4% CPU loads while
utilizing one and then two paths respectively. The
MPTCP was less taxing on the CPU, consuming 8.7%
and 13.7% of CPU resources depending on the number of
paths .

Fig. 9. MPT -MPTCP CPU utilization comparison in a two
path 10 Gigabit Ethernet multipath network environment

V. CONCLUSION S

In this paper, we presented the 10 Gigabit Ethernet
dual-path environment efficiency analysis of two
technologies supporting multipath communication,
which have proved effective in our earlier measurements
dealing with Gigabit Ethernet systems (see e.g. [ 16]-
[22]). Both the MPT -GRE and the reference MPTCP
solution performed below our expectations. Even though
MPTCP managed to beat MPT -GRE in itself with respect
to throughput capacity, it still could not prov ide a
sufficient enough performance. With regards to CPU
utilization, the MPTCP came out a clear winner against
the MPT -GRE technology. In our opinion, these
multipath solutions are capable of improving datacenter
performance when it comes to Gigabit Ether net
environments; however, 10 GE still seems too tough a nut
to effectively crack for both technologies. We believe that
further developments of these solutions and the use of
more modern server configurations could greatly improve
the results that we were currently able to achieve .

ACKNOWLEDGMENTS

This work was supported by the construction EFOP –
3.6.3 -VEKOP -16-2017 -00002. The project was
supported by the European Union, co -financed by the
European Social Fund .

REFERENCES

[1] "The MPT -GRE Project official website", [ August 2019] –

Research

[2] "The MPTCP Project official website", – http://mult ipath –
tcp.org/ [August 2019]
[3] M. Li, A. Lukyanenko, Z. Ou, A. Yla -Jaaski, S. Tarkoma,
M. Coudron, S. Secci, “Multipath transmission for the
internet: A survey”, IEEE Commun. Surveys Tutor., vol. 18,
no. 4, pp. 2887 ‒2925, 2016.
https://doi.org/10.1109/COMST.2016.2586112
[4] A. Ford, C. Raiciu, M. Handley, O. Bonaventure, “TCP
extensions for multipath operation with multiple ad-
dresses", RFC6824, RFC Editor, 2013 .
http://tools.ietf.org/html/rfc/6824
[5] S. Szilágyi, I. Bordán, L. Harangi, B. Kiss, “ Throughput
Performance Comparison of MPT -GRE and MPTCP in the
Gigabit Ethernet IPv4/IPv6 Environment ”, Journal of
Electrical and Electronics Engineering, vol. 12, no. 1, pp.
57‒60, 2019.
[6] C. Paasch, G. Detal, S. Barré, F. Duchêne, O. Bonaventure
“The fastest TCP connection with Multipath TCP ”.
http://multipath -tcp.org/ pmwiki.php?n=Main.50Gbps
[7] S. Szilágyi, F. Fejes, R. Katona, “ Throughput Performance
Comparison of MPT -GRE and MPTCP in the Fast Ethernet
IPv4/IPv6 Environment ”, Journal of Telecommunications
and Information Technology, Vol. 3, No. 2, pp. 53 -59, 2018.
https://doi.org/10.26636/jtit.2018.122817
[8] Á. Kovács, "Comparing the aggregation capability of
the MPT communications library and Multipath TCP", in
Proc. 7th IEEE Int. Confere nce on Cognitive
Infocommunications (CogInfoCom), 2016, pp. 157 –161.
https://doi.org/10.1109/CogInfoCom.2016.7804542
[9] G. Lencse, S. Szilágyi, F. Fejes, M. Georgescu, "Internet
Draft: MPT Ne twork Layer Multipath Library" – https://tools.ietf.org/html/draft -lencse -tsvwg -mpt-04
[September 2019]
[10] F. Fejes, "MPT source code on GitHub" – [March 2019]
https://github.com/spyff/mpt/
[11] B. Almási , G. Lencse, S. Szilágyi , "Investigating the
Multipath Extension of the GRE in UDP Technology",
Computer Communica tions, vol. 103, issue. C, pp. 29‒38,
2017. https://doi.org/10.1016/j.comcom.2017.02.002
[12] B. Almási, S. Szilágyi, “Throughput Performance Analysis
of the Multipath Communication Library MPT”, TSP 2013
– The 36th International Conference on
Telecommunications and Signal Processing, pp. 86 -90,
2013. DOI: https://doi.org/10.1109/TSP.2013.6613897
[13] B. Almási, S . Szilágyi, “Multipath FTP and St ream
Transmission Analysis using the MPT Software
Environment ”, International Journal of Advanced Research
in Computer and Communication Engineering, Vol. 2, Issue
11, pp. 4267 -4272, 2013.
[14] B. Almási, S . Szilágyi, “Investigating the Throughput
Performance o f the MPT Multipath Communication Library
in IPv4 and IPv6 ”, International Journal of Advances in
Telecommunications, Electrotechnics, Signals and Systems,
Vol. 5, No. 1, pp. 53 -60, 2016. DOI:
https: //doi.org/10.11601/ijates.v5i1.148
[15] S. Szilágyi, I . Bordán, L . Harangi, B . Kiss, “MPT -GRE:
A Novel Multipath Communication Technology for the
Cloud ”, 9th IEEE International Conference on Cognitive
Infocommunications: CogInfoCom 2018 Proceedings, pp.
81–86, 2018.
https://doi.org/10.1109/CogInfoCom.2018.8639941
[16] G. Lencse, Á. Kovács, “Testing the Channel Aggregation
Capability of the MPT Multipath Communication Library ”,
World Symposium on Computer Networks and I nformation
Security 2014 (WSCNIS 2014), Paper ID: 1569946547 ,
2014.
[17] G. Lencse , Á. Kovács, “Advanced Measurements of the
Aggregation Capability of the MPT Multipath
Communication Library ”, International Journal of
Advances in Telecommunications, Electrotechnics, Signals
and Systems, Vol. 4. No. 2. , pp 41 -48, 2015. DOI:
https://doi.org/ 10.11601/ijates.v4i2.112
[18] B. Almási, M. Kósa, F. Fejes, R. Katona, L. Püsök, “MPT:
A solution for eliminating the effect of network breakdowns
in case of HD video stream transmission ”, In Proc.
CogInfoCom 2015 Conf., pp. 121 -126, 2015. DOI:
https://doi.org/ 10.1109/CogInfoCom.2015.7390576
[19] F. Fejes, R. Katona, L. Püsök, “Multipath strategies and
solutions i n multihomed mobile environments ”, In Proc.
CogInfoCom 2016 Conf., pp. 79 -84, 2016. DOI:
https://doi.org/ 10.1109/CogInfoCom.2016.7804529
[20] F. Fejes, S. Rácz, G. Szabó, “Application agnostic QoE
triggered multip ath switching for Android devices ”,
Proceedings of the 2017 IEEE International Conference on
Communications : Bridging People, Communities, and
Cultures, IEEE Computer Society, pp. 1585 -1591,
Washingon, 2017.
[21] Á. Kovács, “Evaluation of the Aggregation Capability of
the MP T Network Layer Multipath Communication Library
and Multipath TCP ”, Acta Polytechnica Hungarica, Vol. 16,
No. 6, pp. 129 -147, 2019. DOI:
https://doi.org/ 10.12700/APH.16.6.2019.6.9
[22] G. Lencse, Y. Kadobayashi, “Comprehensive Survey of
IPv6 Transition Technologies: A Subjective Classification
for Security Analysis ”, IEICE TRANSACTIONS on
Communications , Vol. E102 -B, No. 10, pp. 2021 -2035 ,
2019.