Internet architecture enables only a single data path between two communication endpoints within a communication session. On the other hand, decent communication equipments (laptops, tablets, phones) are equipped at the factory with several network interfaces (Ethernet, Wi-Fi, 3G, 4G). It does not worth not using these hardware-given possibilities, which could increase the performance of the communication between two devices, using two or more communication paths. In this paper we presented a possible solution by implementing the MPT-GRE software library. This software was developed under Linux and is based on a totally new architecture, in comparison with the classical TCP/IP model, providing an easy-to-use extension of the current TCP protocol stack (see Figure 1). In our previous papers we investigated its performance in various laboratory measurement environments. In this paper we tried to do it in a much more realistic environment, using the Dummynet WAN emulation software. The measurement results confirmed that the MPT multipath solution could eficiently aggregate the performance of physical connections in the emulated WAN environment as well.
Multipath communication technologies are one of the hot research topics nowadays.
What better proof of this than Apple and Cisco integrating MPTCP1, considered
as the flagship of multipath technologies, into their operating systems [
In our earlier publications (see e.g. [
The testing environments we have used previously (see e.g. Figure 3) can be considered ideal in the sense that they did not contain any network environment parameters that could negatively afect network performance (e.g. delay, jitter, packet-loss). For this reason, we find it important to further examine the efectiveness of MPT-GRE in a more realistic environment (see e.g. Figure 2).
When wants to test a newly developed networking software in a realistic environment, we practically have three possibilities: • internet 1The MPTCP Project oficial website: 2The MPT-GRE Project oficial website: en/mpt/ 3GRE-in-UDP Encapsulation standard: https://tools.ietf.org/html/rfc8086 https://www.multipath-tcp.org/ https://irh.inf.unideb.hu/~szilagyi/index.php/ • network simulation
The first contains advanced technological solutions, but it is, in turn, quite a costly method. The latter is not always capable of providing a reliable and precise
Using the internet can be given in the case of single-path communication. However, being able to test multipath systems, the presence of dual-home technology is essential, i.e. we need to have multiple ISP connections available.
Network simulation aims to replicate the key parameters of desired network environments with the help of mathematical models, with greater or lesser success.
The essence of network emulation is to replicate real network behaviors. Two main types exist: • hardware realization (see e.g. Figure 4) • software implementation test environment, but it is cost-efective. Some examples of network emulator software: Dummy Cloud4, Dummynet5, NETEM6, NIST Net7, SoftPerfect Connection Emulator8, WANem9.
Given that a hardware implementation of a WAN emulator suficient for our goals would be around 6000 EUR + VAT10, after having reviewed the software solutions, our choice was Dummynet.
The Dummynet WAN emulator was developed in 2010 at the University of Pisa,
and later got integrated into the FreeBSD operating system11. It provides a suitable
framework for testing multipath solutions, enabling the setup of packet-delay, jitter
and packet-loss network parameters [
4Dummy Cloud oficial website: http://www.dummycloud.com/ 5Dummynet Project oficial website: http://info.iet.unipi.it/~luigi/dummynet/ 6NetEm’s manual page: https://man7.org/linux/man-pages/man8/tc-netem.8.html 7NIST Net home page: https://www-x.antd.nist.gov/nistnet/ 8SoftPerfect Connection Emulator: https://www.softperfect.com/products/connectionemulator/ 9WANem oficial web page: http://wanem.sourceforge.net/ 10Linktropy Mini-G’s price at November 2020: https://www.digital-hands.eu/products/ apposite/linktropy-mini-g/ 11FreeBSD Manual Pages: https://www.freebsd.org/cgi/man.cgi?dummynet 12Using Dummynet in FreeBSD: http://noahdavids.org/self_published/using_dummynet. html Server 1 tun0 10.0.0.1/30 172.16.2.0/24
Dummynet .1 eth1 eth3 .1 .1 eth2 eth4 .1
Tunnel
All three machines were running Linux Ubuntu operating system. We examined the efect of packet-delay, jitter and packet-loss on file download speed, download time, and CPU performance. Bash and Python scripts – available on our website 13 – were used to automate the measurement process. We repeated each series of measurements ten times. 3. Measurement Results First, we checked how packet-delay afected download speed (see Figure 6).
MPT, with delay
We gradually increased the delay values on a scale of 0-190 ms using Dummynet. Everything proved to be stable until 100 ms. Above 100 ms, we experienced a continuous decrease in file download speeds. Using the 190 ms delay value, the download speed decreased to 107 Mb/s, while download time increased from 47 seconds to 78 seconds.
13Our test scripts can be downloaded from: https://nas01.inf.unideb.hu/share.cgi?ssid= 03CsniS
A similar efect could be experienced in the case of increasing jitter values (see Figure 7). With a 160-190 ms delay fluctuation, the download speed practically decreased by half, while the download time doubled.
MPT, with jitter 200 150 100 50 0 40-70 70-100 100-130
Jitter [ms]
Throughput [Mb/s] Download time [s] 0 130-160 160-190
Applying even a minimal data-loss rate (1 ‰ ), we witnessed a drastic performance decline (see Figure 8). The download speed fall to a quarter, while the download time quadrupled. Therefore, we did not experiment with further dataloss rate values.
0
0.001
Packet loss rate [‰] speed [Mb/s] time [s]
Regarding the efect of packet-delay on CPU performance, we did not experience significant fluctuation (see Figure 9). CPU utilization hovered between 15-22% in every case.
Introducing jitter however, had noticeable efects on CPU utilization (see Figure 10). With higher jitter values, we experienced a drop in CPU utilization.
While examining CPU loads, the efect of packet-loss also proved to be drastic (see Figure 11). Using a data-loss rate of 1 ‰ , utilization dropped from 17 to 5.7 percent.
MPT, with delay 25 20 15 10 5 0 18 16 14 12 10 8 6 4 2 0 20 15 10 5 0 0 70 160
190 100
130
Delay [ms]
We also carried out further measurements, mixing the parameters of the diferent paths. E.g. using only delay on one path, while using only jitter on the other. These scenarios brought similar results as well.
In our current paper, we extended the performance-analysis of our own multipath solution, MPT-GRE, using an emulated WAN environment. We examined the effect of diferent network parameters, like e.g. packet-delay, jitter and data-loss rate on file download speed, download time, and CPU utilization. The worst performance we experienced was with the application of the 1 ‰ packet-loss rate. Among our future goals we are planning to extend our measurements to Gigabit Ethernet IPv4/IPv6 environments, and we would also like to get hands-on experience with the capabilities ofered by hardware WAN emulators.