Benchmarking performance

[VikasChidananda]

Some specification on performance benchmarks:

• run the benchmarks on CPU and GPU

• Profiling to find bottleneck (if any)

• Metrics for comparison: timesteps / unit time, updates / unit time

• vary NN architecture linearly with increasing nodes(N) [learnable parameters M = AN + b]

• [x] Open loop simulation (without control)

• [x] Closed loop simulation (with reinforcement learning / classical controls?)

Ideas for configuring num_sources and num_loads:

• [x] num_sources = num_loads
• [x] set of all (num_source_i, num_loads_j) such that num_source_i + num_loads_j = num_nodes : not required
• [x] stochastically choose num_source such that num_source > num_loads and num_source + num_loads = num_nodes: not required

[VikasChidananda]

Common pitfalls while benchmarking on Linux machines

[VikasChidananda]

• [x] compare performances with processor shielding: requires sudo rights to the system hardware - infeasible
• [x] GPU benchmarks

-RL benchmarks

• [x] decoupled/ independent agents