The implications from benchmarking three big data systems

This paper conducts comprehensive evaluations on three representative data center systems based on BigDataBench, which is a benchmark suite for benchmarking and ranking systems running big data applications, and explores the relative performance of the three implementation approaches with differentbig data applications.

Abstract

Along with today's data explosion and application diversification, a variety\nof hardware platforms for big data are emerging, attracting interests from both\nindustry and academia. The existing hardware platforms represent a wide range\nof implementation approaches, and different hardware platforms have different\nstrengths. In this paper, we conduct comprehensive evaluations on three\nrepresentative big data systems: Intel Xeon, Atom (low power processors), and\nmany-core Tilera using BigDataBench - a big data benchmark suite. Then we\nexplore the relative performance of the three implementation approaches by\nrunning BigDataBench, and provide strong guidance for the big data systems\nconstruction. Through our experiments, we have inferred that a big data system\nbased on specific hardware has different performance in the context of\ndifferent applications and data volumes. When we construct a system, we should\ntake into account not only the performance or energy consumption of the pure\nhardware, but also the characteristics of applications running on them. Data\nscale, application type and complexity should be considered comprehensively\nwhen researchers or architects plan to choose fundamental components for their\nbig data systems.\n