The network simulation method has been studied for years. Compared to expensive and limited-scale network emulation, network simulation achieves a larger scale of systems and has a lower cost. Therefore, network simulation is very important to scientific research. Because contemporary computer networks consist of heterogeneous operating systems (OS), protocols, and applications, the authors argue that the existing simulation only integrates part of the OS, protocols, and applications. Their major contribution, VMSimInt, is a simulation system that integrates arbitrary networking environments, OS, and applications.

The existing packet-based simulation tools (for example, ns-3 and Mininet) have the issues of host OS timing, the need for full implementation of the transmission control protocol (TCP), and so on. This is true in current network simulation tools. To solve these issues, the authors propose a method that can integrate unmodified OS and applications with a modified virtualization tool. The key component of VMSimInt is the modified QEMU software. The authors extend QEMU to support clock control and input/output (I/O) redirection. The main contribution here is how they align the time of virtual machines (VMs) to be the simulated time so the simulation won’t be affected by the host timing issue. The authors also implement a relay program. However, the overhead issue is not discussed much. There may be hundreds and thousands of TCP flows in one network simulation run. Each flow may use an individual socket in the relay program. Therefore, whenever there are many TCP flows, large bandwidth (10 gigabits per second (Gbps) or higher), and long delay (100 milliseconds (ms) or higher), the central processing unit (CPU) or memory overhead may be an issue. Besides, how many simulation hosts are needed to host all the VMs in order for VMSimInt to be good enough for a large-scale high-speed network simulation?

In the evaluation, the authors first compare their approach to ns-3 and IKR simulation, and then evaluate the performance of their approach. The results show that VMSimInt can produce the same simulation results as ns-3 and IKR, and therefore prove the correctness of VMSimInt. In terms of performance, either under greedy traffic or light traffic, VMSimInt shows a longer simulation runtime than other peer solutions.

Overall, in evaluation, the setup is too simple to form a conclusion. For example, the bottleneck bandwidth is too small (5 megabits per second (Mbps), 10 MB/s), which is far away from the contemporary network environment (usually 1 Gbps or higher). This may be insufficient to evaluate the overhead on VMs and their approach.