Performance Characterization: Simulation Speed vs. Simulation Scale

[horser1]

Question

What is the expected order of magnitude for simulation speed as the simulation scale increases?

Context

As users scale up their simulations (e.g., increasing cluster size, number of replicas, request throughput, or model parameters), it would be valuable to understand:

• Simulation throughput: How many simulated timesteps or events per second can Frontier process at different scale levels?
• Scaling behavior: Is the simulation time complexity linear, polynomial, or worse with respect to cluster size, replica count, or request volume?
• Practical limits: What are the typical maximum scales (e.g., number of clusters, replicas per cluster, requests per second) that can be simulated in reasonable wall-clock time?
• Performance characteristics: Are there identified bottlenecks in the DES event loop, scheduling, or communication backend (analytical vs. collective_sim) that degrade performance at scale?

Motivation

This information would help users:

• Plan simulation experiments with realistic runtime expectations
• Identify potential optimization opportunities
• Select appropriate simulation scales for their research questions
• Determine whether analytical or collective_sim backends are suitable for their scale requirements

Related

This question applies to both supported architectures:

• co-location simulations
• Sequential pd-disaggregation simulations

[fwyc0573]

good suggestions, will do in following releases.