Hum, I do not think the Modbus implementation works like that here. In keeping with Modbus’ history (originating as a shared serial bus) it waits for each poll query/response to complete (or fail) before executing the next one.
If considering some multimaster setup, you should bear in mind that some Modbus-TCP slaves only support one TCP connection at a time. This may lead into TCP disconnect/reconnect overheads I’ve already hinted at.
I think truly, we’re trying to say that what you propose should work. It’d be worth doing some more diagnostic work before adding complexity to try and get around it.
At the very least, simplify to find out what works. Shorter lengths? Reduced poll rate?
There is a very long running thread by @ssalonen which includes how-to get extra debugging info