A typical Internet of Things (IoT) system usually consists of three types of components: the client devices that serve as information providers (sensors), the client gateway that gathers raw streaming data from the devices and delivers the data to the cloud, and the cloud that offers all kinds of services to transform the raw streaming data into valuable information. A message broker service cluster, such as a RabbitMQ service cluster, is usually employed as the hub to connect the clients and other cloud services. In the case that the message broker service is multi-tenant, the cloud service provider must deliver good QoS to each client that consists of several devices and gateways. However, a client may enormously occupy the service time and resource of the message broker, and thus other clients may suffer bad QoS due to resource contention. To this end, we propose a dynamic throttling mechanism to address this problem. The proposed mechanism first monitors whether a client uses too much resource. As soon as a 'bad neighbor' client is identified, the proposed mechanism dynamically generates throttling rules for each connection of the bad neighbor, and then applies the rules on each message broker service in the cluster. As the resource usage of the 'bad neighbor' client comes back to a normal degree for a while, the proposed mechanism then nullifies the throttling rules. Our preliminary results show that the proposed mechanism is feasible with little overhead.