Department of Informatics Technische Universität München Informatik X: Rechnertechnik und Rechnerorganisation / Parallelrechnerarchitektur Prof. Dr. Arndt Bode , Prof. Dr. Hans Michael Gerndt

Programming models for distributed architectures

Low-level parallel programming models can generally be divided into two main classes: models using message passing and models using shared memory. Both concepts are well established. The message passing approach, however, is easier to realize. It does not require a tightly coupled system, as each node has its own local memory and address space and all the accesses to remote data are explicitly implemented by sending messages over an interconnection network. This does not call for any additional hardware and also offers a clear programming environment. Because of its simplicity, this model is very well suited for systems like clusters of workstations and large scale MPPs. The drawback of this approach, however, is that the programmer has to explicitly define the data and work distribution across the different nodes. Further, data redistribution, if required, has to be explicitly contained in the code. This has the effect that the programming model is quite different compared to the traditional sequential one on single processor machines.

In contrast, models using shared memory offer a programming model that does not differ far from the sequential one. The programmer only has to deal with one address space. Conceptually, there is no distinction between local and global data; all data can be directly accessed by any processor. Hence, there is also no need for data redistributions allowing easier and more implicit implementations of load and work distribution. However, this ease of programmability comes at a price: the complexity of the implementation of such a system is rather large as each processor needs fine grain access to all the data. Due to this, shared memory programming models can normally be found on tightly coupled systems with small numbers of processors like symmetric multiprocessors (SMPs).

One of the goals of the SMiLE project is to deliver both programming models for the SCI cluster platform. This allows the optimal exploitation of all SCI features: low latency and high bandwidth message passing as well as remote memory access support. For more infomartion please select below for

• SMiLE Message passing, or
• SMiLE Shared memory programming.

In addition, the software efforts in the SMiLE project also include experiments with non conventional parallel programming paradigms. For more informartion please follow:

• Alternative parallel programming paradigms

Back to SMiLE homepage

Questions, suggestions, ideas, criticism ? Don't hesitate to send us an eMail!