|The GRIPs research interests concentrated on the following topics:
Parallel implementation of combinatorial optimization algorithms
Several projects pursued in the LITH focus on the parallelization of combinatorial optimization algorithms for distributed memory parallel computers. This problem is being tackled from several viewpoints:
Performance prediction of multi-processor programs
Performance prediction is an important aspect of parallel computation. It is a key issue for compiling techniques, for the design of programming languages as well as for scheduling and load-balancing. After preliminary works, conducted for the design of the PARCEL language, a new FNRS project was started in mid-1995 explicitly on performance evaluation. Excellent results have been obtained and a thesis has been presented at EPFL on the subject as well as several papers and a useful prediction tool has been developed.
Object-based parallel programming languages
Development of innovating programming languages is an important part of the LITH activities in parallel computing. Two different languages and compilers have been developed. In the PASTIS language (in the P.Kuonen PhD thesis), dedicated to the implementation of combinatorial optimization algorithms, asynchronous communications were investigated. In the PARCEL language, dedicated to neural networks and artificial intelligence, synchronous programming is investigated. Both projects produced implementations of operational multi-processor compilers. Compiling techniques are also investigated in the ADD project on automatic data distribution.
Automatic data distribution
The ADD project aims at developing techniques to achieve automatically the distribution of data on distributed memory parallel computers. The project notably addresses the problem of the alignment of data, and the development of efficient data redistribution algorithms.
Commodity based supercomputers
Commodity parallel computing becomes more and more popular, as the specialised supercomputing companies stop their activity. In this new concept, a user is confronted with a machine constructed with mass produced fully equipped workstations or PCs which are interconnected through some high speed, low latency network. In collaboration with CAPA the LITH was strongly involved in the Swiss-Tx project . The final objective of this was to build a low cost, parallel commodity computer delivering one Teraflop/s. A first machine of this type, the T1, was delivered in June 2000, its power is half a Teraflop.
Until recently, the world of High Performance Computing (HPC) was mainly involved in solving huge numeric problems using matrix calculation (number-crunching). But, since the beginning of the nineties, distributed computing is emerging and some prototypes of parallel applications running on less costly local networks of workstations appeared. Following the development of Local Area Networks (LAN) during the eighties, Internet technology has popularized Wide Area Networks (WAN). Message passing programming, first localized on LAN and on massively parallel computers, is now moving to WAN. Habits of users are rapidly changing from a program centric vision to a service centric vision. Future users will require the realization of a given service in the most efficient environment presently available to them through a WAN. This evolution in the user's needs has led to the creation of a new high performance computing paradigm called metacomputing. In the framework of the ADOOC project (Asynchronous Distributed Object Oriented Computing, LITH joined in 1999 the international WOS (Web Operating System) project. This project aims at providing users with the possibility to submit a parallel/distributed service request without any prior knowledge about the service (where it is available, at what cost, under which constraints) and to have the service request fulfilled within the user's desired parameters (time, cost, quality of service, etc.)