Project Requirements

PPF is hoped to be a robust, efficient, and correct library to help in the parallelization of radiation particle transport packages. Specifically the goals of the PPF are to:

• Distribute workload of radiation particle transport problems semi-efficiently between nodes of large distributed supercomputers.
• Provide scalability of a problem across arbitrary nodes.
• Decompose the problem mesh across across multiple nodes transparently to the developer.
• Provide transparent particle passing between nodes/mesh sections.
• Eliminate the need for developers to learn the intricacies of distributed parallel programming.
• Provide a standard for particle physics module architectures.
• Reduce development and training time for radiation particle transport packages.
• Create a tracking system to aid in particle visualization.

Since the parallel particle framework is a library to be used by developers of large physics problems where correctness and efficiency is crucial, the library must have the following attributes.

• Ease of Use: PPF must provide a simple interface that is intuitive to a developer with no experience in parallel computing. All code dealing with parallelization technology must be transparent to the developer.
• Efficiency: Large scale problems running billions of iterations on thousands of nodes can continuously accumulate small efficiency problems eventually leading to a significant increase to the running time of the problem. For this reason the PPF must be extremely efficient in message passing and performance.
• Stability: Problems run by radiation transport particle packages may take days to complete. For this reason it is imperative that PPF be stable by ensuring no runtime errors nor memory leaks. This is more important for PPF than other small scale applications because a faulty run can cost the developers days at a time. In the advent of a crash PPF should be able to resume operation at the last time step completed.
• Scalability: The platforms used for large scale radiation transport problems are continuously growing in size, therefore in order for PPF to be of use in the future it must be scalable to an arbitrary number of nodes efficiently.
• Longevity: PPF must ensure the integrity of all its requirements for the life of the package which implements it. The KULL framework is expected to be a 20 year code, the PPF should retain functionality for at least this time period.

The following two lists are details of the costumers needs and wants for the Parallel particle frameworks functions.

Needs

• The library must be able to handle split meshes by indexing matched faces between mesh partitions.
• During runtime the PPF must be able to make the following decisions correctly:
  • when to stop tracking particles on a partial mesh and send particles frozen on a split surface to another domain
  • how many particles to buffer before sending
  • how often to perform checks whether or not to send buffered particles
  • when to ask if another domain is ready to send particles
  • how to handle load balancing if possible
  • what to do if one domain ends up with to many particles too fit in memory
• Particles must be sent as efficiently as possible.
  • particles should not be packaged with extra information for sending
  • data should only be sent to nodes which require it, no broadcasts
• Particle transfers between nodes must be accurate.
• The library's interface must be intuitive and not dependent on a prior knowledge in distributed computing.

• The library may handle any size mesh of type PolyMesh or PolyMesh2.
• The PPF may provide support for general domain decomposition and work sharing between nodes.
  • The mesh may be re-decomposed dynamically to handle load balancing. \item The PPF may use a domain replication scheme to fairly distribute work loads.
• The PPF may include shared memory parallelization within each node using pthreads or OpenMP.
• The library may include extensions for particle visualizations.
• The library may decompose any given mesh using a graph cutting package such as Parmetas.

• All code written for PPF must be compatible with both the egcs and KCC compilers
• Programming must conform to sound object oriented design, and make use of MPI2 if a reliable implementation is found for the sp2 architecture.
• PPF code must have general compliance with POSIX standards.
• There must be readable and comprehensive documentation in all stages of the development process.
• The library must compile and run on AIX 4.3 and DEC Unix.
• The library must be easily integratible into the make system of existing projects.