Participant : Franck Cappello [ correspondant ] .
Distributed applications can be programmed on the Grid using workflow languages, object oriented approaches (Proactive, IBIS, etc), RPC programming environments (Grid-RPC, DIET), component based environments (generally based on Corba) and parallel programming libraries like MPI.
For high performance computing applications, most of the existing codes are programmed in C, Fortran and Java. These codes have 100,000 to millions of lines. Programmers are not inclined to rewrite then in a "non standard" programming language, like UPC, CoArray Fortran or Global Array. Thus environments like MPI and OpenMPI remain popular even if they require hybrid approaches for programming hierarchical computing infrastructures like cluster of multi-processors equipped with multi-core processors.
Programming applications on the Grid add a novel level in the hierarchy by clustering the cluster of multi-processors. The programmer will face strong difficulties in adapting or programming a new application for these runtime infrastructures featuring a deep hierarchy. Directive based parallel and distributed computing is appealing to reduce the programming difficulty by allowing incremental parallelization and distribution. The programmer add directives on a sequential or parallel code and may check for every inserted directive its correction and performance improvement.
We believe that directive based parallel and distributed computing may play a significant role in the next years for programming High performance parallel computers and Grids. We have started the development of OpenWP. OpenWP is a directive based programming environment and runtime allowing expressing workflows to be executed on Grids. OpenWP is compliant with OpenMP and can be used in conjunction with OpenMP or hybrid parallel programs using MPI + OpenMP.
The OpenWP environment consists in a source to source compiler and a runtime. The OpenWP parser, interprets the user directives and extracts functional blocks from the code. These blocks are inserted in a library distributed on all computing nodes. In the original program, the functional blocks are replaced by RPC calls and calls to synchronization. During the execution, the main program launches non blocking RPC calls to functions on remote nodes and synchronize the execution of remote functions based on the synchronization directives inserted by the programmer in the main code. Compared to OpenMP, OpenWP does not consider a shared memory programming approach. Instead, the source to source compiler insert data movements calls in the main code. Since the data set can be large in Grid application, the OpenWP runtime organize the storage of data sets in a distributed way. Moreover, the parameters and results of RPC calls are passed by reference, using a DHT. Thus, during the execution, parameter and result references are stored in the DHT along with the current position of the datasets. When a remote function is called, the DHT is consulted to obtain the position of the parameter data sets in the system. When a remote function terminates its execution, it stores the result data sets and store a reference to the data set in the DHT.
We are evaluating OpenWP from an industrial application (Amibe), used by the European aerospace company EADS. Amibe is the mesher module of jCAE(project page: http://jcae.sourceforge.net ). Amibe generates a mesh from a CAD geometry in three steps. It first creates edges between every patch of the CAD (mesh in one dimension), then generates a surface mesh for every unfolded patch (mesh in two dimensions) and finally adds the third dimension to the mesh by projecting the 2D mesh into the original CAD surfaces. The first and third operation cannot be distributed. However the second step can easily be distributed following a master/worker approach, transferring the mesh1d results to every computing node and launching the distributed execution of the patches.