Object-Oriented Parallel Programming

Object-oriented programming has become fairly common and popular and has been used in numerous computer-science projects. However, it has not yet been utilized to any degree in large-scale scientific computing. Nevertheless, we believe it is well suited to scientific computing, which frequently deals with well-defined, loosely interacting, physical objects. In particular, the paradigm is particularly useful in distributed, parallel computing because the objects help to encapsulate and clearly define the movement of data. The message interface maps well to the distributed memory model of parallel computing by constraining the mapping of data into memory. The additional features of inheritance and data abstraction also promise to reduce significantly the cost of software maintenance for large-scale scientific programs.

Distributed Computing

An important tool in defining the objects to be used in scientific computing comes from a mapping of the physical model into the computational environment. If the objects are made to correspond to physical elements in the underlying model, the paradigm fits very well. Since all of the information an object needs is stored internally, an object can also provide a good representation for a thread of control. This can greatly assist in the development of parallel applications in a way that is

independent of the hard ware on which it is being run. One can logically define the necessary number of threads (or processes) required for a given problem without worrying about the number of physical processors. This is analogous to not worrying about the size of vector registers in a supercomputer or the actual number of processors in a Thinking Machines Corporation Connection Machine.

A number of these ideas have been implemented in a distributed particle simulation code, which is reported on in the 1990 USENEX C++ conference proceedings (Forslund et al. 1990). The distributed environment for this code is the ISIS programming environment developed at Cornell University by Ken Birman (1990) and his colleagues.

Data Parallel Programming

Another area in which object-oriented programming has had success is data parallel programming. Rob Collins (personal communication), from the University of California, Los Angeles, has built an efficient C++ library for the Thinking Machines Corporation CM-2, called CM++. It gives full access to the C/Paris functionality without any loss in performance. This allows writing in a higher-level abstraction without sacrificing speed. Collins and Steve Pope of the Advanced Computing Laboratory at Los Alamos National Laboratory (personal communication) have been working on porting this library to a more general environment, called DPAR. Thus, the data-parallel paradigm has been abstracted and now runs on a workstation with good optimization. We are optimistic that this library could be made to work on a Cray Research, Inc., supercomputer, with comparable efficiency to that provided on the CM-2.

We are also trying to combine this data-parallel programming paradigm with the distributed environment we mentioned before, following some of the ideas of Guy Steele. In a paper entitled "Making Asynchronous Parallelism Safe for the World," Steele (1990) describes a programming style that tries to unify SIMD and MIMD computing. The essence of the proposal is to allow asynchronous threads to have only a restricted shared-memory access providing only commutative operations. This removes the dependence of the order of sibling parallel threads.

Threads then communicate (synchronize) only on their death. Complex operations are provided by a hierarchy of communications. This is the style we have used in our distributed particle code mentioned earlier. In one sense, this style is the simplest extension of the SIMD programming model.