Deep Data Locality on Apache Hadoop

The amount of data being collected in various areas such as social media, network, scientific instrument, mobile devices, and sensors is growing continuously, and the technology to process them is also advancing rapidly. One of the fundamental technologies to process big data is Apache Hadoop that has been adopted by many commercial products, such as InfoSphere by IBM, or Spark by Cloudera. MapReduce on Hadoop has been widely used in many data science applications. As a dominant big data processing platform, the performance of MapReduce on Hadoop system has a significant impact on the big data processing capability across multiple industries. Most of the research for improving the speed of big data analysis has been on Hadoop modules such as Hadoop common, Hadoop Distribute File System (HDFS), Hadoop Yet Another Resource Negotiator (YARN) and Hadoop MapReduce. In this research, we focused on data locality on HDFS to improve the performance of MapReduce. To reduce the amount of data transfer, MapReduce has been utilizing data locality. However, even though the majority of the processing cost occurs in the later stages, data locality has been utilized only in the early stages, which we call Shallow Data Locality (SDL). As a result, the benefit of data locality has not been fully realized. We have explored a new concept called Deep Data Locality (DDL) where the data is pre-arranged to maximize the locality in the later stages. Specifically, we introduce two implementation methods of the DDL, i.e., block-based DDL and key-based DDL. In block-based DDL, the data blocks are pre-arranged to reduce the block copying time in two ways. First the RLM blocks are eliminated. Under the conventional default block placement policy (DBPP), data blocks are randomly placed on any available slave nodes, requiring a copy of RLM (Rack-Local Map) blocks. In block-based DDL, blocks are placed to avoid RLMs to reduce the block copy time. Second, block-based DDL concentrates the blocks in a smaller number of nodes and reduces the data transfer time among them. We analyzed the block distribution status with the customer review data from TripAdvisor and measured the performances with Terasort Benchmark. Our test result shows that the execution times of Map and Shuffle have been improved by up to 25% and 31% respectively. In key-based DDL, the input data is divided into several blocks and stored in HDFS before going into the Map stage. In comparison with conventional blocks that have random keys, our blocks have a unique key. This requires a pre-sorting of the key-value pairs, which can be done during ETL process. This eliminates some data movements in map, shuffle, and reduce stages, and thereby improves the performance. In our experiments, MapReduce with key-based DDL performed 21.9% faster than default MapReduce and 13.3% faster than MapReduce with block-based DDL. Additionally, key-based DDL can be combined with other methods to further improve the performance. When key-based DDL and block-based DDL are combined, the Hadoop performance went up by 34.4%. In this research, we developed the MapReduce workflow models with a novel computational model. We developed a numerical simulator that integrates the computational models. The model faithfully predicts the Hadoop performance under various conditions