Information Technology Reference
In-Depth Information
Kumar et al. examine the optimization of two different objective functions, one minimiz-
ing Energy and the other EDP, assuming either oracle heuristics for core switching or realistic
dynamic switching heuristics. Under the oracle heuristics the operating system simply chooses
the core for each time quantum that optimizes the objective function. Switching overhead is
not taken into account.
Lack of oracle knowledge dictates that realistic heuristics must discover for themselves
which core optimizes the objective function. Every hundred time quanta, cores are tested for five
consecutive quanta to assess their effectiveness in optimizing the objective function. Switching
overhead is accurately modeled in the simulations. Thus, the effectiveness of a core on the first
of the five test quanta is ignored to discount the effects of cold caches, cold TLBs, and cold
branch predictors.
The four realistic heuristics studied by Kumar et al. differ on which cores are selected for
testing:
neighbor : a neighbor in the performance continuum to the core that is running is selected
at random;
neighbor-global : similar to the above but the selected core is the one that best optimizes
the objective function in the application's execution thus far;
random : any core is selected at random;
all : all cores are tested.
In all cases, the authors report substantial gains in energy or EDP. For the oracle
heuristics, optimizing for energy yields up to 60% energy reduction with a 5% performance
loss [ 147 ]. EDP is reduced by up to 63%. Translating these results into ED 2 P shows that core
switching can even outperform chip-wide DVFS—which can, at best, only break even on the
ED 2 Pmetric.
Results are also exceptionally good for the realistic heuristics with all switching overhead
accounted. EDP for three of the four heuristics ( neighbor , neighbor-global ,and random )iswithin
90% of the oracle heuristics. The “all” heuristic tends to fall behind because of its greater testing
overhead.
Overall, core switching is an important work steering technique that can be used not only
for reducing power consumption, but also for reducing power density (by migrating activity
from hot and busy cores to idle and cold cores). In this last incarnation, the technique is known
as activity migration [ 97 ].
pitfalls in optimizing EDP : Kumar et al. noticed a peculiar behavior in their interval-
based approach to optimize EDP. Even when choosing the optimal EDP per interval, with
Search WWH ::




Custom Search