Motifs used in Big Data workloads, with a specific focus on benchmarking has allowed more abstract and generic workloads to be defined as combinations of these motifs, where each motif is defined to be predictable and deterministic in nature. We theorize that if a workload is broken into a sequence of motifs, then a measure of a performance sample can be obtained at the start of each motif. In this paper we outline novel phase detection methods, using clustering algorithms and try to identify similar workloads across benchmarking suites to quantify similar behavior, and build a neural network-based approach to detect the phase transfer of workloads in their time of execution. We identify the phase determination changes for the Sort, TeraSort, WordCount and some graph benchmarks such as MD5, PageRank and Connected Components, identified as key repeating motifs in the benchmarks under consideration.
For more information, please read the Reference Report.