|
What are typical overheads of sequential computations?
Sequential computation overheads, compared to parallel computation,
are caused by artificial limitations imposed on the application.
Two examples illustrate this point:
- artificial search constraints - to limit search to practical sizes, one
usually imposes by oneself artificial constraints. This is essential, otherwise
search paths may not terminate. Parallelization adds degrees of freedom,
exploring many paths in parallel, relaxing the severity of such constraints.
- memory limitations - in data intensive applications, sequential
computation uses the single processor memory, which is a hard limitation.
Parallel computation with several processors, automatically adjusts memory
resources to application needs.
Are there communication overheads in parallel computations?
Communication overheads in parallel computation, occur only when
parallelization analysis of the application was done improperly.
Parallel applications display computation/communication trade-offs .
Suitable analysis leads to good balance between them.
How many applications are best represented in parallel?
Many of them. A small sample of such applications is:
- geometric figures - problems dealing with geometric constraints on figures
- e.g. chip placement and routing problems in VLSI - see the references
in the Industry Solutions page.
- 3D space - applications embedded in physical space -
e.g. CAD-CAM problems.
- multi-dimensional - problems with many dimensions, e.g. imaging spaces,
possibly including a time axis.
Are there intrinsically parallel applications?
YES. There are applications that cannot be serialized,
without complete distortion or serious losses. For instance,
simulation of multiple participant games, such as basketball.
|
