TOP-C Description :

•       TOP-C especially distinguishes itself as a package to easily parallelize existing sequential applications. An application programmer can frequently convert a sequential application to a parallel application in less than a day, with no prior experience with TOP-C. TOP-C is both small and self-contained, even including its own MPI subset. Yet, it can also work well with your debugger, your own MPI, etc.
  

•       The design goals of TOP-C are that it should follow a simple, easy-to-use programmer's model, and that it should have high latency tolerance. This allows for economical parallel computing on commodity hardware, as well as high performance on the latest supercomputers. TOP-C hides the details of parallel programming, and presents the application programmer with a simple task-oriented interface, for which the application programmer need only define four callback functions. Yet, this simple model has been shown to readily adapt to a wide variety of algorithmic requirements.

•     TOP-C runs on most variants of UNIX/Linux. The source code is layered, making it easy to modify and easy to add a new communication module for a new hardware architecture. Current communication modules include one for distributed memory using sockets (e.g. networks of workstations), one for shared memory using POSIX threads, and one for a single CPU (useful for debugging). The same TOP-C application code runs with any of the three communication modules.

•     TOP-C has three fundamental concepts: the task (specified by the master and executed by a slave process), the global shared data, and the action chosen after a task is completed. Communication between processes occurs only through these three mechanisms. Yet, a TOP-C application is built around a single system call: TOPC_master_slave(). It takes four arguments consisting of four application-defined callback procedures:

GenerateTaskInput() -> input   DoTask(input) -> output   CheckTaskResult(input, output) -> action   UpdateSharedData(input, output)       (executed only if UPDATE action returned)

                                                Upon receiving the output of a task, the master then decides on one of four actions:

NO_ACTION UPDATE (update the shared data) REDO (in case the result of some              other task has altered the shared data) CONTINUATION(new_input_for_same_slave)

A task re-executed due to a REDO action can be executed more quickly, because it is executed in the same process as the original task. Hence any information from the previous task computation and previous update can be saved in a global variable and used to accelerate the re-computation of the task under the new task input.

This simple parallel model turns out to be surprisingly adaptable for parallelizing existing sequential software, as is demonstrated in the parallelization of a 1,000,000 line C++ program, Geant4 at CERN, for simulation of particle-matter interaction, with applications to physics, engineering and biomedicine. A version of TOP-C (called ParGAP) has also been used to parallelize GAP (Groups, Algorithms and Programming), and is distributed from their site.