Scaling gradient descent with Hadoop

The length of time each iteration of batch gradient descent takes to run is determined by the size of your data and by how many processors your computer has. Although several chunks of data are processed in parallel, the dataset is large and the processors are finite. We've achieved a speed gain by performing calculations in parallel, but if we double the size of the dataset, the runtime will double as well.

Hadoop is one of several systems that has emerged in the last decade which aims to parallelize work that exceeds the capabilities of a single machine. Rather than running code across multiple processors, Hadoop takes care of running a calculation across many servers. In fact, Hadoop clusters can, and some do, consist of many thousands of servers.

Hadoop consists of two primary subsystems— the Hadoop Distributed File System (HDFS)—and the job processing system, MapReduce. HDFS stores files in chunks. A given file may be composed of many chunks and chunks...