Section: New Results

Modeling Affinity of Multithreaded Applications

With the increasing complexity and scale of multi-core processors, optimizing thread placement becomes more and more challenging. Our goal is to better understand which characteristics of a multi-threaded application can have an impact on a placement decision for a given architecture. To this end, we analyze the performance of a set of applications under different placement strategies and we try to relate the obtained results to characteristics of the applications such as the data footprint of each thread, the amount of data shared between threads, or the reuse distance.

To collect information about the characteristics of multi-threaded applications, we developed a set of tools based on the PIN dynamic binary instrumentation tools. PIN allows us to get information about all instructions executed and memory location accessed by each thread of an application during its execution, and this without modifying the source code of the application.

We used our PIN-based tools to study a representative set of applications taken from two well-known benchmark suites, namely the Mantevo benchmark suites (HPC applications based on OpenMP) and the Parsec benchmark suites (general-purpose applications based on pthreads). Analyzing the results of all our tests is an ongoing work.