In a cloud computing environment, virtual machine allocation is an important task for providing infrastructure services. Generally, the datacenters, on which a cloud computing platform runs, are distributed over a wide area network. Therefore, communication cost should be taken into consideration when allocating VMs across servers of multiple datacenters. A network-aware VM allocation algorithm for cloud is developed. It tries to minimize the communication cost and latency between servers, with the number of VMs, VM configurations and communication bandwidths are satisfied to users. Specifically, a two-dimensional knapsack algorithm is applied to solve this problem. The algorithm is evaluated and compared with other ones through experiments, which shows satisfying results.