You can usually find this kind of issues when applying policying in hardware-based units as switches are. As a matter fact QoS documentation under policing and marking section states:
The accuracy of the policer configured is +/- 1.5 percent.
The best advise I can give is to play with the burst rate, decrease it, tune the configuration to accomplish better results... I have learnt this from experience. Not sure if there's any bug ID for something but I have never been able to find one that justifies this behaviour. I'm also interested in hearing a better response, if there's one.
If you are testing with jperf change to udp and test again. The TCP sessions are dropping the sessions speed dramatically when some packets are dropped or you may use multiple TCP sessions. You may need to review the way you are testing. I would suggest to use shaper on output if you will use TCP for testing.