Most Infrastructure As A Service (IaaS) cloud platforms are implemented with a cluster of machines having multiple worker nodes and a single master node, which acts as the cloud front-end. In such an architecture the control regarding the Virtual Machine (VM) scheduling and policies compliance etc. lies largely with the front-end node of the cluster. Decision making in such IaaS cloud does not have any intelligent and real time participation from the worker nodes. This is undesirable for scenarios where the cloud is composed from nodes that are semi/fully autonomous and/or when their ownership is hierarchical in nature. It becomes even more significant issue in cases where the hyper visor is a Type-2 (i.e. hosted) kind. This is because a Type-2 hyper visor runs as a regular operating system (OS) process subjected to local OS policies regarding resource allocation and security etc. To address these issues, we propose here an architecture which makes use of real time system-state information from the cluster nodes and decentralizes the cluster node policy management etc., among other design decisions. We show that the scalability, security, availability and reliability of the IaaS cloud gets improved with our proposed architecture. We apply the proposed architecture to a real use case where it helps utilize the untapped computing capacity of large pool of powerful PCs without violating the unique Quality of Service (QoS) needs of their regular owners and users.