-
Notifications
You must be signed in to change notification settings - Fork 0
Home
PinT-TF is a light-weighted framework to fastly explore the feasibility and practicality of real world numerical simulations using time-space hybrid parallel method. It also can be used as a performance testing tool for large scale computing, or a base codes for parallel computing related education. The name is the abbreviation for Parallel in Time Test Framework.
In the current release, the time division scheme is parareal method, the space division scheme is the classical uniform mesh. The parareal algorithm was presented by Lions, Maday, and Turinici in 2001 year as a numerical method to solve evolution problems in parallel. The name was chosen to indicate that the algorithm is well suited for parallel real time computations of evolution problems whose solution cannot be obtained in real time using one processor only. The method approximates successfully the solution later in time before having fully accurate approximations from earlier times. The algorithm has received a lot of attention over the past few years, especially in the domain decomposition literature (MARTIN J. 2007).
As a path finder of applying PinT to specific real-world computing problems, the framework provided basic time-space division functionalities, you can freely divide the time domain and the space domain of the problem through a configuration file. The underlying uniform mesh can support 1/2/3 dimension automatically, and complete all common tasks related to space parallel, such as guard cell synchronization, applying boundary condition, and aggregating the result to HDF5 file.
As a performance testing tool to verify PinT in HPC environment, it provides a default performance counter based on the open-source library PMLib. Hardware counter, OpenMP/MPI hybrid mode, calculation/communication measurement are supported. You can activate the performance monitor as need.
As a teaching and experiment code base for parallel computing, it is designed very compactly and clearly. It is easy to be modified and extended. Students can work on higher concepts to implement parallel numerical algorithm without explicitly calling MPI routines, or tune the performance by directly modifying the MPI code and the OpenMP directives at lower concepts.
- Jian Xiao
- Seigo Imamura
- Mikio Iizuka
- Kenji Ono