Hadoop distributed file system (HDFS) becomes a representative cloud storage platform, benefiting from its reliable, scalable and low-cost storage capability. HDFS has been utilized in BlueSky, one of the most prevalent e-Learning resource sharing systems in China, to store and share courseware majorly in the form of PowerPoint (PPT) files and video clips. Unfortunately, HDFS does not perform well for massive small files since huge numbers of small files imposed heavy burden on NameNode of HDFS, correlations between files were not considered for data placement, and no prefetching mechanism was provided to improve I/O performance. This paper introduces a novel approach to improve the efficiency of storing and accessing small files on HDFS. Characteristics of file correlations and access locality remaining among small files in the context of courseware are well considered for storing and accessing them. Firstly, all correlated small files of a PPT courseware are merged into a larger file to reduce the metadata burden on NameNode. Secondly, a two-level prefetching mechanism is introduced to improve the efficiency of accessing small files. The experimental results indicate that the proposed approach is able to effectively mitigate the load of NameNode and to improve the efficiency of storing and accessing massive small files on HDFS. © 2010 IEEE.