Efficient tracing
In early tests, capturing the tracing data as tor plaintext log messages resulted in 20+ GB files for a 10m sim on a 1% network.
Using lttng tracing may help. It logs in an efficient binary format so should definitely be smaller initially. It's also designed to be easily compressible, so may compress smaller than the plaintext logs.
Ideally we'd like the trace data from all nodes in the network, but in case we need to prioritize, mikeperry says:
after overall performance characteristics that shadow already provides, the most important data is the queue lengths at exit relays (including edge connections), followed by queue lengths at intermediate relays, and CircEWMA (circuit scheduler) behavior
Edited by Jim Newsome