Lustre, Hadoop, Accumulo

TL;DR

This paper compares Lustre, Hadoop, and Accumulo, highlighting their foundational principles, capabilities, and performance differences on a hypothetical cluster, aiding in understanding their optimal use cases.

Contribution

It provides a foundational comparison and simple models for assessing Lustre, Hadoop, and Accumulo, including performance metrics and recent integration efforts.

Findings

Lustre offers 2x more storage capacity and higher bandwidth on general workloads.

Hadoop provides 4x greater read bandwidth on specific workloads.

Accumulo achieves 10,000x lower latency on random lookups.

Abstract

Data processing systems impose multiple views on data as it is processed by the system. These views include spreadsheets, databases, matrices, and graphs. There are a wide variety of technologies that can be used to store and process data through these different steps. The Lustre parallel file system, the Hadoop distributed file system, and the Accumulo database are all designed to address the largest and the most challenging data storage problems. There have been many ad-hoc comparisons of these technologies. This paper describes the foundational principles of each technology, provides simple models for assessing their capabilities, and compares the various technologies on a hypothetical common cluster. These comparisons indicate that Lustre provides 2x more storage capacity, is less likely to loose data during 3 simultaneous drive failures, and provides higher bandwidth on general purpose workloads. Hadoop can provide 4x greater read bandwidth on special purpose workloads. Accumulo provides 10,000x lower latency on random lookups than either Lustre or Hadoop but Accumulo's bulk bandwidth is 10x less. Significant recent work has been done to enable mix-and-match solutions that allow Lustre, Hadoop, and Accumulo to be combined in different ways.