Abstract

Abstract
Parallel systems are not necessarily special-purpose machines. Al present, most departmental servers already resort to shared memory multiprocessing as a means to increase performance, while a network of workstations can also be regarded as a distributed memory parallel system. This paper examines the support offered by the operating system to exploit such parallelism. After discussing the design of multi-threaded programs in a Unix environment, a comparison is made with their distributed counterparts. Some performance figures obtained on a SparcCenter 2000 multiprocessor, on a network of workstations and on a transputer-based system are presented.

Abstract
Complex applications in distributed-memory parallel systems often follow a demand-driven approach with domain decomposition. A uniform data distribution among the local memories at the processing elements may require frequent remote data access. To keep the processors busy while data is remotely fetched, concurrent application processes are assigned to each transputer-based processing element. Adding more concurrent application processes in a large-scale parallel system may degrade performance, due to the traffic increase with data requests and data block replies. A conditional broadcast mechanism is implemented during data requests, to limit this flow of messages. Monitoring strategies are proposed to further reduce the messages density, and a parameterized model to measure and evaluate global execution times is presented. Simulation data running the model with up to 35 transputers show that monitoring can reduce the performance degradation when more local concurrence is added. However, if too much data replication is present, the simulation data also show that the supply of communication services at each node still imposes a burden, requiring complementary monitoring strategies to allow removal of redundant reply messages.

Abstract
This paper describes CA/C++, Concurrency Annotations in C++, a language extension that regulates method invocations from multiple threads of execution in a shared-memory multiprocessor system. This system provides threads as an orthogonal element to the language, allowing them to travel through more than one object. Statically type-ckecked synchronous and asynchronous method invocations are supported, with return values from asynchronous invocations accessed through first claw future-like objects. Method invocations are regulated with synchronization code defined in a separate class hierarchy, allowing separate definition and inheritance of synchronization mechanisms. Each method is protected by an access flag that can be switched in pre and post-actions, and by a predicate. Both must evaluate to true in order to enable a thread to animate the method code. Flags and method predicates are independently redefinable along the inheritance chain, thus avoiding the inheritance anomaly.

Abstract
By adopting theories as primitive components of a logic and recognizing that formulae are just presentation details we arrive at the concept of topological institution. In a topological institution, we have, for each signature, a frame of theories, a set of interpretation structures and a satisfaction relation. More precisely, we have, for each signature, a topological system. We show how to extract a topological institution from a given institution and establish an adjunction. Illustrations are given within the context of equational logic. We study the compositionality of theories. Formulae are recovered when we establish a general technique for presenting topological institutions. Topological institutions with finitely observable theories are shown to be useful in temporal monitoring applications where we would like to be able to characterize the properties of the system that can be monitored. Namely, an invariant property (Gf) cannot be monitored because it cannot be positively established in finite time. On the contrary, a reactivity property (Ff) can be positively established in finite time. © Springer-Vedag Berlin Heidelberg 1995.

Abstract