Recommended BGP/OSPF configuration for a dense 1000-router INET network

[anhndd]

I am using OMNeT++ 6.4.0 with INET 4.7.0 and need to simulate network topologies containing up to 1000 backbone routers. Each backbone router is connected to a local router and an end host. I evaluate several physical topology types, including ring, star, scale-free, and complete graphs.

My current dynamic-routing configuration assigns each backbone/local-router pair to a separate AS. OSPF is used only within each pair, while eBGP sessions are established between physically connected backbone routers. This works reasonably for multiple topologies except the complete topology.

In Cmdenv, this case can spend several hours in initialization before meaningful simulation progress occurs. I also observe very large memory usage. For example, for a single simulation process, the complete topology has reached approximately VIRT: ~49 TB - RES: ~369 GB even if it just show Initializing... after 8 hours of Setting up network "MyNetwork" whereas the other 1000-router topologies easily complete with VIRT: ~1.5 TB - RES: ~22 GB.

I am trying to determine whether this behavior is primarily an expected consequence of the number of network interfaces/BGP sessions, or whether there are INET/OMNeT++ configuration options that I should change for simulations of this scale.

I already disable most scalar/vector recording. The relevant configuration is approximately:

output-scalar-file = "/dev/null"
**.scalar-recording = false

**.vector-recording = false

# A small subset of application-specific signals is explicitly enabled:
**.<selected-signals>.vector-recording = true

Cmdenv output is currently redirected to a log file.

I would appreciate advice on the following:

1. Is a dense network with around 1000 routers and hundreds of thousands of physical links/BGP sessions within the intended practical scale of INET's current BGP/Ethernet models?
2. Are there recording, logging, statistics, visualization, packet, or other modules that should explicitly be disabled to reduce initialization time or memory consumption?
3. Does the very large VIRT value indicate a problem by itself, or is RES the more relevant quantity when evaluating INET memory consumption?
4. Are there known examples or scalability recommendations for OSPF/BGP simulations with approximately 1000 routers?