ALPS: A Methodology for Application-Level Communication Characterization of Parsec 2.1

AbstractThis paper presents a methodology for analyzing communication of multi-threaded applications. Previous work relies on more or less accurate architectural models. Our measurement methodology has been designed to be completely architecture independent, since we want architects to have an undistorted view of the communication behavior. One part of our methodology is the concept of communication attribution which allows communication from underlying libraries, for example, to be attributed to application functionality. In this paper, we have applied our methodology to the Parsec benchmark suite. Our characterization shows how communication in Parsec scales, as we increase the number of threads from 2 to 256. Based on these results we have modeled communication growth as a power function and determined the characteristic exponents for each benchmark. Last but not least, our methodology enables researchers to track communication down to the source code