Intro Flashcards
1
Q
weak strong scaling
A
weak: problem size per HW thread/core etc stays the constant
strong: general size stays constant
2
Q
amdahl law
A
(i-f)T + fT/p
3
Q
parllel eficiency
A
speedup/p
4
Q
what can limit parallelization
A
- comm and synch
- resource bottlenecks
- network conflicts
- multiple processes on same node
- load imbalance
5
Q
superlinear speedup
A
more bandwisth and fit in cache
6
Q
data and function parallelism
A
data: SPMD
less overhead, but load imbalance
functional: pipelining
good load balance, but overhead
