Methodology to predict the performance behavior of shared-memory parallel applications on multicore systems

摘要

With the advent of multicore architectures, there arises a need for comparative evaluations of the performance of well-understood parallel programs. It is necessary to gain an insight into the potential advantages of the available computing node configurations in order to select the appropriate computing node for a particular shared-memory parallel application. This paper presents a methodology to resolve this issue, by constructing a database with behavior information for various representative shared-memory programming structures, and then estimating the application behavior as a combination of these data. These structures are represented by small specific chunks of code called microbenchmarks (μB) based on the study of memory access patterns of shared-memory parallel applications. μBs set is run on each candidate node, and all execution performance profiles are stored in a database for future comparisons. Then, applications are executed on a base node to identify different execution phases and their weights, and to collect performance and functional data for each phase. Information to compare behavior is always obtained on the same node (Base Node (BN)). The best matching performance profile (from performance profile database) for each phase, is then searched. Finally, the candidates nodes performance profiles identify in the match process are used to project performance behavior in order to select the appropriate node for running the application. Resource centers own different machine configurations. This methodology helps the users or systems administrator of data centers to schedule the jobs efficiently.