A Performance Analyzer for Task Parallel Applications based on Execution Time Stretches
スポンサーリンク
概要
- 論文の詳細を見る
Performance loss in task parallel applications is contributed by 3 factors of thread idleness, parallelism overhead and work time stretch. Thread idleness is the time that threads have no work to do and parallelism overhead is the time that threads spend on extra instructions that would not be necessary in serialexecution. The third factor, work time stretch, refers to the surplus time by which the same application-levelcode takes longer in parallel execution than in serial execution. We believe that work time stretch is the most important factor in future multi-core systems. Therefore, we have developed a profiler that analyzes worktime stretch of task parallel applications. The profiler can clarify the contribution of work time stretch factorout of the other two, attributing stretched amount to specific code blocks so that programmers can know which parts of their programs are stretching. It also shows the surplus cache miss count that accompanies work time stretch.
- 2013-07-24
著者
-
Kenjiro Taura
The University of Tokyo
-
Jun Nakashima
The University of Tokyo
-
An Huynh
The University of Tokyo
関連論文
- Cynk: A Hybrid Rsync and SSH Filesystem for Cloud Computing
- A Preliminary Evaluation of Chapel with Molecular Dynamics Simulation
- A Preliminary Evaluation of Chapel with Molecular Dynamics Simulation
- A Preliminary Evaluation of Chapel with Molecular Dynamics Simulation
- A Preliminary Evaluation of Chapel with Molecular Dynamics Simulation
- Porting MassiveThreads Thread Library to FX10 Supercomputer System
- Porting MassiveThreads Thread Library to FX10 Supercomputer System
- A Power State Transition Algorithm for Power Controller of Energy Efficient Computer Clusters
- A Performance Analyzer for Task Parallel Applications based on Execution Time Stretches