We have witnessed the fast-growing deployment of Hadoop, an open-source implementation of the MapReduce programming model, for purpose of data-intensive computing in the cloud. However, Hadoop was not originally designed to run transient jobs in which users need to move data back and forth between storage and computing facilities. As a result, Hadoop is inefficient and wastes resources when operating in the cloud. This paper discusses the inefficiency of MapReduce in the cloud. We study the causes of this inefficiency and propose a solution. Inefficiency mainly occurs during data movement. Transferring large data to computing nodes is very time-consuming and also violates the rationale of Hadoop, which is to move computation to the data. To address this issue, we developed a distributed cache system and virtual machine scheduler. We show that our prototype can improve performance significantly when running different applications.