Just to be sure I'm understanding this right, in the formula on this slide, r refers to the resources used to make the one "fat" core with r=4, correct?
Then, in the (perf(r) + (n-r)) term, perf(r) would be the performance of the one big core, or perf(4). And (n-r) would represent the 12 cores with r=1.
Can someone verify this?
This comment was marked helpful 0 times.
eatnow
@aew that looks correct to me
However, considering the difficulties in using a heterogeneous system (as described in later slides), would this formula be a fair approximation for how performance actually scales in real machines? It seems to be a little too optimistic...
This comment was marked helpful 0 times.
kayvonf
@aew: correct.
@eatnow: perf(r) = sqrt(r) is a very crude approximation to reality that was used to make the graphs. The key point is that it is sublinear. If it was the case that perf(r) = r, then there would be little need for parallel processing as we'd use all our resources to linearly improve single-threaded performance, and not make software developers tackle the challenges of parallel programming.
Just to be sure I'm understanding this right, in the formula on this slide,
rrefers to the resources used to make the one "fat" core withr=4, correct?Then, in the
(perf(r) + (n-r))term,perf(r)would be the performance of the one big core, orperf(4). And(n-r)would represent the 12 cores withr=1.Can someone verify this?
This comment was marked helpful 0 times.
@aew that looks correct to me
However, considering the difficulties in using a heterogeneous system (as described in later slides), would this formula be a fair approximation for how performance actually scales in real machines? It seems to be a little too optimistic...
This comment was marked helpful 0 times.
@aew: correct.
@eatnow:
perf(r) = sqrt(r)is a very crude approximation to reality that was used to make the graphs. The key point is that it is sublinear. If it was the case thatperf(r) = r, then there would be little need for parallel processing as we'd use all our resources to linearly improve single-threaded performance, and not make software developers tackle the challenges of parallel programming.This comment was marked helpful 0 times.