Welcome to the forum. This looks like TCP delayed ACK behavior.
One of the services that TCP provides is reliability; it does this by sequencing each byte of data and retransmitting if data is loss is detected. The detected part is important - it doesn't mean data loss occurred, it just means that it was detected. In your case, if you look at the retransmission time (this would be the server side RTO, round trip timeout) it is just about 100ms. When the server puts bytes on the wire (MAC address says VMWare device, but you claim a Rockwell PLC, not sure why, or if it matters) it starts a timer. If it does not get ACK before the timer expires (RTO), it resends the data.
Client side behavior has an optimization called delayed ACK. It does not want to send an ACK for EVERY packet - it waits a little while to see if it has some data to send back, so will then send data and the ACK at the same time. More efficient. In your case, you have no data to send back quickly, but there is a timer here too - the delayed ACK timer. I don't know what OS you are using, but WindowsXP says 200ms (
https://support.microsoft.com/en-us/kb/328890). So it will wait that long before sending a response, even a discrete ACK. Notice the timing mismatch - server waits 100ms before timing out, client is waiting 200ms (or whatever, but longer than 100ms). So you get a true TCP retransmission every Modbus request cycle. This is a generalization; it's actually more complicated. RFCs and google can keep you busy for quite some time on the topic, as could
Steven's book.
Some of these values are adjustable by the user - Windows has the registry entry, and Linux can adjust RTO values per route (
http://indcontrolproto.blogspot.com/2014/02/tcp-timing-and-rto.html).
Anyway, this is not unheard of. It's not great behavior but usually networks are over provisioned so this has little impact. In your case with radio communications, I think you are correct to be a little concerned as this could waste precious bandwidth. Since the Rockwell device is embedded, the timers are likely all fixed. Worth asking Rockwell if there is a way to adjust, but I would not expect a good response. So if you feel this will be a real issue, focus on the client side. Might need to move your Modbus poller to a different OS or system that has more hooks or at least different behavior to avoid this.