Hardware Reference
In-Depth Information
before, timing spent in different kernels during cycle-accurate simulations is profiled
and fed back to HLSim. By this timing back-annotation, we remain accurate enough
(
<
20% error) between HLSim and cycle-accurate simulations w.r.t. execution time
and power consumption. Further, Pareto operating points obtained by HLSim are also
validated with cycle-accurate simulator. By using such a two-simulator approach,
DSE phase can be done efficiently and accurately to reach optimum operating points
in a short amount of time. Note that using such multi-level simulators to reduce
number of simulations is orthogonal to sophisticated modeling techniques (e.g., Re-
sponse Surface Modeling techniques presented in Chap. 4) supported by DSE tools
[
1
].
To give an idea of the time saved by the proposed methodology we obtain that:
The
exhaustive exploration
(i.e., the simulation of all 335886 configurations) with
TLMsim
would take 1,343,544 h; almost
153 years
.
Performing the
NSGA-II optimizations
using
TLMsim
would require
16 months
.
The
exhaustive exploration
with
HLSim
would take about
6 months
.
Performing the
NSGA-II optimizations
using
HLSim
requires about
36 h
.
9.4.2
Run-Time Resource Management
9.4.2.1
Simulation of an Urban Environment
To generate dynamic QoS requirements for the RRM, the overall ACSS has been
simulated for several reasonable driving patterns and conditions [
4
]. Although we
simulated different patterns, for conciseness and clarity we will report results ob-
tained for a specific pattern reproducing an urban scenario. The car speed and the
frame rate requirements are shown in Fig.
9.6
.
The vehicle, starting from a stationary position with speed equal to 0 km/h starts
to accelerate a few before the 35th second. The vehicle reaches a maximum speed
of 60 km/h around the 50th second and it keeps a constant speed for a while. Then,
at about the 60th second, the vehicle starts decelerating to reach again a stationary
condition at about the 70th second (Fig.
9.6
a). All three live-views are active on the
dashboard at the beginning of the simulation. The live-view of the central camera is
deactivated from the driver at around 50 s and the frame rate of the corresponding
video stream is reduced to 15
fps
as required from the CSU (Fig.
9.6
c). During the
simulation two vehicles pass within the proximity area on the left side. The frame
rate requirement of the corresponding camera increases to 30
fps
(Fig.
9.6
b).
9.4.2.2
Run-Time Resource Management Behaviors
Whenever the CSU sets new frame rate requirements, the RRM is invoked to globally
select one operating configuration for each MPEG4 encoder with the goal of fitting
Search WWH ::
Custom Search