Modelling large-scale networks requires significant computational resources on a computer that produces a simulation. Moreover, the complexity of the calculations increases nonlinearly with increasing volume of the simulated network. On the other hand, cluster computing has gained considerable popularity recently. The idea of using cluster computing structures for modelling computer networks arises naturally. This paper describes the creation of software which combines an interactive mode of operation, including a graphical user interface for the OMNeT++ environment, with a batch mode of operation more natural to the high-performance cluster, "Sergey Korolev". The architecture of such a solution is developed. An example of using this approach is also given.
1. Introduction
Modeling large-scale networks requires significant computational resources on a computer that
produces a simulation. Many universities have at their disposal computational cluster architectures of
various sizes. The aim of this current work is to improve the quality of network simulation by using
cluster "Sergey Korolev" of Samara National Research University. OMNeT++ is used in this work to
model computer networks. OMNeT++ is a discrete event simulation environment [
The use of the high performance cluster "Sergey Korolev" often dictates a batch (and not
interactive) mode of operation [
For simulation of computer networks, it is more convenient to use an interactive mode of operation
alongside a graphical user interface. For these purposes, software should be created which is capable
of interacting with the user via graphical tools. In such a case, the OMNeT++ system [
In [
A number of papers are devoted to the use of clusters in scientific applications.
In [
The work [
There are works that offer novel algorithms for storing and processing data on clusters.
Article [
A number of papers are devoted to the use of the OMNeT++ simulator on cluster architectures.
The article [
The article [
This paper demonstrates the concept of cluster modeling of RTEthernet systems. The computing platform is based on a standard PC with an RT-Linux kernel, on which RTEthernet models work in the OMNeT++ simulation environment and the ARM9 microcontroller as a coprocessor.
The evaluation showed that the platform provides sufficient performance for the latency requirements of distributed real-time systems in the 230µs range, which is limited and has a linear dependence on frame size. The observed jitter is below 40µs.
The article [
So, to summarize, the above article describes an analysis of the parallelization problems inherent in the INET set of OMNeT++ models and demonstrates the possibility of INET parallelization, using a number of modifications to the models; this parallelization was primarily aimed at distributed initialization of the INET models.
The concept of distributed multistage initialization allows for the creation of a simulation model which supports distributed simulation execution. The tests showed that there is an acceleration of execution, due to parallelization, for the given example in most configurations, although there is still potential for further optimization. The implementation covers the most commonly used protocols, such as Ethernet, IPv4, TCP, and UDP, as well as the related models such as ICMPv4 or the application of UDP.
According to the official documentation [
To deploy a network simulator on a cluster, the steps presented in Table 1 are necessary. access the cluster; upload the necessary files to the cluster; create a task which build the OMNeT++ network simulator and wait for its completion (there is a version of OMNeT++ without an IDE, this is what was used here); create a task which build the INET Framework; create a test simulation model; create a task which runs the simulation and wait for its completion; download the result of the work.
However, interactive task execution is possible on the cluster - given access to a free node in the cluster. The simulation will be run as follows (in Figure 2).
It is worth noting that the main difficulties encountered when deploying a simulator on a cluster are to do with recreating the necessary environment and also there are the problems relating to the limited access available (all work is done in the console mode).
After initial testing, it was found that the INET Framework does not support parallel simulation
work. To solve this problem, the proposals in “Parallel INET for OMNeT++” [
To work with MPI, the Tictoc1 system was selected; this consists of two modules. These modules should be assigned to different MPI nodes, as shown in Figure3.
Then the project in OMNeT++ is built on the local machine.
To check the functionality from the project directory via mpirun, the following command is issued: $ mpirun -np 2 ./tictoc -m -u Cmdenv -c Tictoc1 omnetpp.ini. The result of such a command should look like this:
Next, the folder containing the project itself and the folder containing the source files must be copied over in order that the necessary libraries can be collected together into a working folder on the supercomputer, so that all can then be compiled (the libraries, and the project itself). The list of required libraries can be viewed using ldd %executable_file_name%, usually the necessary libraries will be only those that are in the OMNeT++ folder.
Next, it necessary to build the project on the cluster.
Then you need to create, in the working folder on the supercomputer, the task file *.pbs. Sample file content is as follows: #!/bin/bash #PBS -N helloOMNeT //name of task #PBS -A helloOMNeT #PBS -l walltime=00:01:00 //required time #PBS -l nodes=1:ppn=2 //number of nodes and processes #PBS -j oe //error stream to standard output cd $PBS_O_WORKDIR export PATH=$PBS_O_PATH # Loading Intel MPI v4 environment module load impi/4 export I_MPI_DEVICE=rdma export I_MPI_DEBUG=0 export I_MPI_FALLBACK_DEVICE=disable # command run task taking into account the features of the environment mpirun -r ssh -machinefile $PBS_NODEFILE -np $PBS_NP ./tictoc -m Cmdenv -c Tictoc1 omnetpp.ini. -u
We plan to test the work of OMNeT++ on the cluster "Sergey Korolev" using the project INET Framework. It is also planned to organize the possibility of full interactive work in the graphical interface of the modeling environment. To illustrate the effectiveness of the method used, it is planned to measure the performance of applications running on a cluster.
In this paper, we showed the development of software which allows us to combine an interactive mode of operation of the OMNeT++ modeling system and the batch mode of the cluster “Sergey Korolev”. in Via initial testing, it was discovered that the way by which networks are created in the course of simulation does not work in Parallel INET for OMNeT++. As a result, it was decided to create a network for parallel simulation using a Python script, and for ordinary simulation using a module created for OMNeT++. These results can be used to solve the other, related, problems [12].
This work was undertaken as a component of the project part of the state task of the Ministry of Science and Higher Education of the Russian Federation No. 2.974.2017/4.6.