Big Data Processing: Navigating in a Zoo of Yellow Elephants, Sharks, and Giraffes
The term “Big Data” was ranked 10 on the Global Language Monitor’s 2013 top business buzzwords. However, behind the marketing phrase hides the actual hard problem of how to cope with huge amounts of possibly unstructured data that are produced at very high rates. Storing and processing that data is highly valuable as it provides insights that can be used by businesses to enhance user experience (making recommendations, refining search results, etc.) or to support decision making. However, traditional approaches are not able to provide feasible analysis at the required scale.
In this talk I will give an overview of existing software systems that aim to achieve this goal. I will start with MapReduce and its open-source implementation Hadoop, the most prominent among big data processing systems, and look at its basic concepts that enable users to conveniently analyse data at scale. I will also discuss its drawbacks and then introduce more advanced systems such as Spark that enable richer analytics via general, flow-based, programming APIs. Finally, I will briefly talk about network transfers as one major bottleneck these systems face and introduce NetAgg, a system developed in the LSDS group. NetAgg can reduce the amount of transferred data by performing early aggregation inside the network and hence, reduce the time it takes to analyse data.
The aim of this talk is to give a high-level introduction to the Zoo of systems that are out there to process big data and present some research directions we are pursuing in the LSDS group.