cassandra-stress utility performs warmup only when number of operations is specified.
In case of duration based execution it doesn't provide warmup. That leads to count startup/warmup induced latency that maybe considered as undesired statistics.
For example latencies difference between warmed and cold execution even for 15 minutes may reach to: 17% for 99th percentile, 35% for 99.9th and 55% for max latency.
Attached patch contains example how it may be fixed. It was chosen to use 1/5 of duration time as warmup time, but probably the proper way to do it is to add a new option.