W&M Featured Events
This calendar presented by
William & Mary
[PAST EVENT] Understanding Irregular Application Performance with Analytical Models
October 23, 2015
4pm - 5pm
Understanding Irregular Application Performance with Analytical Models
Analytical (white box) performance modeling has been an important methodology for diagnosing bottlenecks, optimizing systems, and designing large-scale machines. Irregular applications -- increasingly important for high performance computing -- resist common white-box modeling techniques. As a result, analysts may use statistical or black box models that provide less insight.
This talk presents examples of analytical modeling in the context of irregularity. We model one-sided network communication to quantify and pinpoint network contention in applications based on one-sided programming models. We model two-sided communication protocols to develop an energy efficient MPI runtime for irregular applications. Knowing the protocol an MPI runtime invokes accurately predicts whether it will be beneficial to incur the cost of applying a power saving feature. We model graph applications and network topology to predict the potential impact of silicon photonics on analytics workloads that require a "rack of memory." We conclude with ongoing work to ease the burden of creating analytical models.
Nathan Tallent is a computer scientist in the Advanced Computing, Mathematics, and Data Division at Pacific Northwest National Laboratory. His research is at the intersection of tools, performance modeling and analysis, and parallelism. He was an original author of Rice University's HPCToolkit performance tools.
Analytical (white box) performance modeling has been an important methodology for diagnosing bottlenecks, optimizing systems, and designing large-scale machines. Irregular applications -- increasingly important for high performance computing -- resist common white-box modeling techniques. As a result, analysts may use statistical or black box models that provide less insight.
This talk presents examples of analytical modeling in the context of irregularity. We model one-sided network communication to quantify and pinpoint network contention in applications based on one-sided programming models. We model two-sided communication protocols to develop an energy efficient MPI runtime for irregular applications. Knowing the protocol an MPI runtime invokes accurately predicts whether it will be beneficial to incur the cost of applying a power saving feature. We model graph applications and network topology to predict the potential impact of silicon photonics on analytics workloads that require a "rack of memory." We conclude with ongoing work to ease the burden of creating analytical models.
Nathan Tallent is a computer scientist in the Advanced Computing, Mathematics, and Data Division at Pacific Northwest National Laboratory. His research is at the intersection of tools, performance modeling and analysis, and parallelism. He was an original author of Rice University's HPCToolkit performance tools.