In a shared virtual computing environment, dynamic load changes of applications give rise to dynamic capacity demands, which results in lower resource utilization and application quality using the static resource allocation. We argue that finding an optimized resource allocation method to solve the problem of on-demand resource provision for VMs is the key to improve efficiency of data centers. However, the existing dynamic resource allocation methods focus on either local optimization within a server or central global optimization, limiting efficiency of data centers. We propose a two-tiered on-demand resource allocation consisting of local and global resource allocation with feedback to provide on-demand capacities to concurrent applications. Based on the proposed dynamic resource allocation mechanism and model, we propose a set of on-demand resource allocation algorithms. Using Rainbow, a Xen-based prototype we implemented, we evaluate the VM-based shared platform and our on-demand resource allocation algorithms. The experimental results show that Rainbow without dynamic resource allocation (Rainbow-NDA) provides 26% to 324% improvements in the application performance and 26% higher average CPU utilization than traditional service computing framework. The two-tiered on-demand resource allocation improves performance by 9% to 16% for those critical applications, 75% of the maximum performance improvement in comparison with Rainbow-NDA.