Performance Analysis of Big Data Motifs on Large Core Machines

The last decade has seen improvements in processor architecture that have enabled CPU dies to pack 64 to 128 cores, compared to 4 to 8 cores only a few years ago. Applications are being developed to take advantage of multi-core, high-performance computing. As a result of the volume of applications being developed, particularly in the Big Data and Artificial Intelligence fields, it is not possible to benchmark every single one of them. To tackle this problem, data motifs may be used. Data motifs can be defined as core units of computation that can be distinguished from other computational units by their data access patterns. Essentially, the most time consuming part of a program can be considered a motif. A program can be modelled as a pipeline of many motifs. As development paradigms shift towards making use of large core computing, which is defined in this paper as 16 or more cores, data access patterns for programs tend to change. As with most multi-threaded programs that follow Amdahl’s and Gustafson’s laws, there is a certain limit to the speedup that can be achieved, before synchronization begins to overbear the time saved through parallelism. Synchronization may consist of access to shared data and locking, as well as lower level mechanisms such as cache contention. The aim of this research is to corroborate the existence of data motifs through analysis of run-time data, analyze the performance of these data motifs on large core machines, and provide an understanding of how data access patterns vary with the underlying software stack. This paper details the results of the subsequent analysis and comparison of the motifs based on the shared memory reads and writes between their implementations.