Prof. S. Hambrusch, West Lafayette, IN

Lecture
Oesterreichisches Forschungsinstitut fuer Artificial Intelligence (OeFAI)
                      Schottengasse 3, A-1010 Wien
                       Tel.+43-1-53532810,5336112
-------------------------------------------------------------------------

                                VORTRAG
                                *******


Prof.Dr.Susanne HAMBRUSCH
Purdue University
West Lafayette, Indiana
USA


High Performance Computing:
Can We Achieve Scalability and Usability?
-----------------------------------------

High performance computing must meet seemingly contradictory goals: 
software development should be carried out in an architecture- and
technology-independent environment, but both algorithms and system
software should take full advantage of the hardware features of
potentially diverse MPP (Massively Parallel Processors) systems.
Distributed memory, message-passing MPPs, which include the Intel 
Paragon, IBM SP-2, and Cray T3D machines, have emerged as major high
performance architectures.

In this talk we discuss the impact of different MPP features on the 
performance of scalable algorithms. We show how parameters like the 
number of processors, the message size, cost of setting up a message, 
the bandwidth of the processors and the network, the latency, and the 
bisection width influence scalability. We describe a parallel model, 
the C^3-model, which gives an accurate prediction of algorithm per-
formance and supports a realistic scalability metric. We argue that 
scalability is achieved by the use of poly-algorithms which employ 
a parallel model to make algorithm selection and that usability is 
enhanced through libraries incorporating poly-algorithms. We validate 
our approach for regular and irregular communication operations and
discuss performance results for the Intel Paragon.


Zeit: Donnerstag, 1.August 1996, 18:30 Uhr pktl.
Ort:  OeFAI, Schottengasse 3, 1010 Wien 1.