Hardware Reference
In-Depth Information
Fig. 8.15
Reference fault and observation distances
injected faults, whenever the FTM is implemented at a level located between the
level of the targeted faults and the level where the faults are injected and thus inter-
cepts the error propagation paths. Indeed, assuming a perfect (100%) coverage for
the FTM, then representativeness (with respect to the targeted faults) would then be
null. This could be simply accounted for by introducing another distance parameter:
the distance
dm
separating the level where faults are injected from the level where
the FTM is acting.
Another typical concern is whether the error propagation process is monotonously
(usually upstream) directed or not; the first alternative prevents FTMs characterized
by negative values of
dm
from being exercised. For example, the SWIFI technique
can seldom be used to test hardware-implemented ECCs.
8.3.4
Comparative Study of Fault Injection Techniques
consequences of a fault (i.e., the error propagated) offers a pragmatic and sensible
means to address the representativeness issue. Two categories of approaches can
be distinguished, depending whether the analysis concerns the erroneous behaviors
provoked by (1) some specific fault injection technique with respect to a set of real
faults or (2) the concurrent application of several fault injection techniques.
The study considered hereafter - see
Arlatetal.
(
2003
) for additional details -
refers to the second approach. While the comparison of several fault injection tech-
niques provides only an indirect means for assessing their representativeness, such
an approach is well suited to obtain extensive error data sets from which useful in-
sights can be derived. Should the experiments using different injection techniques
lead to similar behaviors, then the techniques can be considered as “equivalent”, and
thus the one that exhibits the most suitable practical properties (e.g., reachability,
