Research of realistic parallel programming models
Simo Juvaste, University of Joensuu, Department of Computer Science
Some definitions of (the use of) the keywords
and terms are available.
Motivation for the research
-
- Parallel computation = computation + communication
- There is no free communication!
Goal
-
To construct a good shared memory parallel programming model with the
required algorithmic backround to be usable on different types of
parallel computers and application areas.
The model should be able to predict the real execution costs of
algorithms (i.e. not only computation costs, but also the communication
costs (in form of shared memory reference costs).
- (Virtual) shared memory
- Asynchronous execution
- A shared memory reference results a future
- Parameters describing the properties of the machine used
- Shared memory reference latency
- Shared memory bandwidth/processor (i.e. computation/communication ratio)
- Cost of (barrier) synchronization
Some of the above optimization/machinedependency possibilities make
problem complexity analysis (and its results) much more difficult, but
might remove random worst cases on algoritms, and make algorithm
analysis easier.
Last modified 21. June 1995 SJ