Performance Optimization for Large-Scale Slot Simulations (10^7-10^9 slots)

Vienna 5G System Level Simulator
1

Dear Vienna 5G System Simulator Developers,

I’m currently working on simulating extremely large numbers of slots (on the order of 10^7-10^9) to observe packet drop behavior under existing scheduling strategies.

My current approach uses parallel simulation with 8 chunks (each containing 1000 slots), running 125 iterations to achieve 10^6 slots. However, this configuration already takes approximately 20 hours to complete. Scaling this up to ≥10^8 slots would be computationally prohibitive with the current performance.

I understand that the simulator is a comprehensive system-level tool that models many functional modules in the communication process, which naturally leads to significant computation time. Could you please advise:

  1. Which specific components or modules in the simulation workflow are the most computationally intensive?
  2. Are there particular parts of the codebase that could be optimized to improve simulation speed?
  3. Would you recommend any specific configuration changes or simplifications that could enable large-scale slot simulations while maintaining reasonable accuracy?

My goal is to achieve statistically significant results from these large-scale simulations within feasible computation time. Any guidance on optimization strategies would be greatly appreciated.

Thank you for your time and assistance.

Best regards,
Nick Jay

2

Dear Nick Jay,

I recommend using the matlab run and time tool to evaluate your simulation, since the simulation time depends strongly on the chosen parameters.

At the end of the user manual a section elaborates on the computation time of some features, but not all features. I recommend having a look at it to get a feeling for which parameters have the most influence on the simulation time.

Since you asked questions about the 3GPP TR 38.901 channel model, I assume that you use this channel model. It is one of the more computationally complex models. If the channel properties specific to this model are not relevant to your results, I recommend switching to the most suitable PDP trace model. There, the channel traces can be precalculated, making the simulation faster.

Another thing to remember is that statistical significance can come from simulating more slots, but also from simulating larger networks. So you can play around with the network size to reduce overhead with larger networks, or reduce overall simulation time with smaller networks that need to consider fewer links.

Best regards,
Agnes