HPCS-Lecture

[ Pobierz całość w formacie PDF ]
High Performance Computing Systems
Pawel Czarnul
pczarnul@eti.pg.gda.pl
Computer Architecture Department
Gdansk University of Technology, Poland
© 2011 Pawel Czarnul
Parallelization Goals
From Execution Speed-up to Better Reliability
© 2011 Pawel Czarnul
Parallelization Goals and Performance Measures
Common Parallelization Goals

Performance: Speed-up the execution

Performance measure – time in e.g. seconds

Usual architecture: local clusters, supercomputers

Reliability: Increase the reliability

Performance measure – time in e.g. seconds provided that the
probability that the application finishes is not smaller than a
predefined value

Usual architecture: local clusters

Functionality: Integrate your application(s) with services that are not
available locally

Pay for remote services e.g.:

File storage (pay for the number of MBdays or similar)

High performance resources e.g. very fast clusters or
supercomputers etc.

Usual architecture: WWW, geographically distributed resources
© 2011 Pawel Czarnul
General Trade-offs
in
Parallel and Distributed Processing
Flexibility vs. Performance vs. Cost
© 2011 Pawel Czarnul
Performance vs. Other Quality Attributes
Security
Flexibility
Scalability
Reliability
Efficiency
Interactive Features
Portability
Ease of Use
performance
© 2011 Pawel Czarnul
[ Pobierz całość w formacie PDF ]

zanotowane.pl

doc.pisz.pl

pdf.pisz.pl

chiara76.opx.pl