Given our experience, I would say stick with the continuous running. Running on a schedule will build up latency and the distribution agents can slow down as the distribution database gets bigger. We replicate in the region of 800 million transactions a day around our entire infrastructure which is roughly 200 million per day to each of 4 subscribers. This represents about 30-40GB per day per server. We run all our agents in continuous mode but that's not to say you shouldn't experiment. We also run the 3 server architecture I mentioned earlier.

Only getting 100 cmds/sec out of the distribution agent is not good really. I'd expect to see something more like 500-800 cmds/sec or even higher if insert in the only activity. Hopefully, everything is indexed properly. Nothing lying around which shouldn't be there! Apologies if that's an insult.

I really would recommend creating more publications to get better parallelism into your design. We only replicate 9 tables off our main publishers but have 3 publications each with a dedicated distribution agent.

Mike