The buzz-word big-data refers to the large-scale distributed data processing applications that operate on exceptionally large amounts of data. Google’s MapReduce and Apache’s Hadoop, its open-source implementation, are the defacto software systems for big-data applications. An observation of the MapReduce framework is that the framework generates a large amount of intermediate data. Such abundant information is thrown away after the tasks finish, because MapReduce is unable to utilize them. In this paper, we propose Dache, a data-aware cache framework for big-data applications. In Dache, tasks submit their intermediate results to the cache manager. A task queries the cache manager before executing the actual computing work. A novel cache description scheme and a cache request and reply protocol are designed. We implement Dache by extending Hadoop. Testbed experiment results demonstrate that Dache significantly improves the completion time of MapReduce jobs.

Data Center Networks (DCNs) are the fundamental infrastructure for cloud computing. Driven by the massive parallel computing tasks in cloud computing, one-to-many data dissemination becomes one of the most important traffic patterns in DCNs. Many architectures and protocols are proposed to meet this demand. However, these proposals either require complicated configurations on switches and servers, or cannot deliver an optimal performance. In this paper, we propose the peer-assisted data dissemination for DCNs. This approach utilizes the rich physical connections with high bandwidths and mutli-path connections, to facilitate efficient one-to-many data dissemination. We prove that an optimal P2P data dissemination schedule exists for FatTree, a specially-designed DCN architecture. We then present a theoretical analysis of this algorithm in the general multi-rooted tree topology, a widely-used DCN architecture. Additionally, we explore the performance of an intuitive line structure for data dissemination. Our analysis and experimental results prove that this simple structure is able to produce a comparable performance to the optimal algorithm. Since DCN applications heavily rely on virtualization to achieve optimal resource sharing, we present a general implementation method for the proposed algorithms, which aims to mitigate the impact of the potentially-high churn rate of the virtual machines.