The data is consistently replicated on all nodes so the data transfered between nodes would depend upon the number of mutations you are doing. Does your mutations have long string values ?
Can you try disabling logs and check the IO usage and also can you please file an issue on github regarding logging. We could work on improving it.
The high io might not be due to logs. You can get an estimate of the io done by dgraph by checking the size of p and w directly every 10 seconds.