portant paradigms of parallelism. The Coir system is a Cϩϩ library that combines features of multithreading and object-oriented paradigms and provides abstractions to program these parallel constructs. It uses object-oriented abstractions to model a wide range of parallel systems from SMPs to distributed memory machines.
This paper is organized as follows: The next section introduces the Coir system. Section 3 describes abstractions for parallel architectures in the Coir model. Controlparallel abstractions using thread objects are introduced in Section 4. Data-parallel abstractions using rope objects are described in Section 5. A brief discussion on different classes of data-parallel applications appears in Section 6. Section 7 discusses related research work, and Section 8 describes the status of current work and proposes scope for future work.
2. THE Coir SYSTEM
The Coir system provides a programming model for both control and data-parallelism for shared and distributed memory machines. In Coir, the unit of control-parallelism is a thread. We define a thread as a sequence of instructions, with its own stack and program counter, in execution in a program. Coir threads are lightweight user-level threads which execute on top of kernel-level threads. Thus Coir provides lightweight application-level parallelism without kernel intervention. Threads in Coir can migrate across processes sharing the same memory arena.
Coir is designed at two levels. The lowest level consists of C libraries for multithreading and message-passing. The system is built so that any thread library that supports a ''reasonable'' subset of the pthreads standard can be plugged in at this level. Coir defines a communication mechanism that has a one-to-one mapping to MPI . Thus if the MPI implementation of a communication library avoids unnecessary overheads like buffer copies, then Coir does not introduce any. The second level of Coir is defined in Cϩϩ. Here the multithreaded control and communication mechanism is combined with inheritance, dynamic dispatch, and parameterized types. Using the interface provided at this level, user-specific parallel programming abstractions can be built.