In contemporary JaaS configurations, resources are distributed to users primarily through the assignment of virtual machines (VMs) to physical nodes (PMs). This resource allocation is typically done in a way that does not consider user preferences and is unaware of the underlying network layout. The latter is of key significance as cost of the cloud's internal network does not grow linearly to the size of the physical infrastructure. In this paper, we focus on JaaS clouds built on the highly fault-tolerant and scalable PortLand networks. We examine how the performance of the could can benefit from VM placement algorithms that exploit user-provided hints regarding the features of sought VM interconnections within a virtual infrastructure. We propose and evaluate two such VM placement algorithms: the first seeks to rapidly place the required VMs as closely as possible on the PortLand network starting with the most demanding virtual link and by following a greedy approach. The second approach identifies promising neighborhoods of PMs for deploying the virtual infrastructure sought Both methods try to reduce the network utilization of the physical layer while taking advantage of the PortLand layout. Moreover, we seek to minimize the time expended for the placement decision regardless of the size of the infrastructure. Our experimentation shows that our methods outperform the traditional methods (first-fit) in respect to network usage. Our greedy approach reduces the network traffic routed through the top-level core-switches in the PortLand topology by up to 75%. The second approach attains an additional 20% improvement.