Hardware Reference
In-Depth Information
Tabl e 8. 5 Detailed analysis of the hardware EDMs activated by the injection techniques
Hardware EDMs
HI
PF
EI
SWIFI
# of CPU EDMs activated
7
4
5
5
# of NMI types (and
combinations)
13 (26)
17 (34)
10 (16)
0
# of exceptions activated
73
65
14
Data not available
segment provoke rather control flow errors that, to a large extent, better simulate
the consequences of hardware faults. SWIFI experiments targeting the data segment
generate mainly data flow errors that are different from the ones created by the other
techniques.
Nevertheless, it is worth noting that these experiments provide also a rather high
level of activation of the hardware EDMs in the NOAM configuration, i.e., when the
end-to-end mechanisms are inhibited. A closer examination of the results is reported
in Table 8.5 that summarizes some of the main differences observed.
Concerning CPU EDMs, although out of the eight mechanisms supported, a dif-
ferent number of mechanisms were activated, the same three mechanisms dominated
(bus error, address error and illegal opcode) for all the fault injection techniques.
The number of NMI types and their combinations (i.e., the simultaneous occur-
rences of several triggering events) vary significantly for the HWIFI techniques. 8
The results indicate that PF may be more effective than the other techniques in
exercising hardware EDMs located outside of the CPU chip. Moreover, the most
frequent NMIs observed differ: while “unavailable memory” significantly domi-
nates for HI (more than 60%), “memory parity” dominates for PF and EI (more than
50%). Both PF and EI also exhibited a significant proportion of NMIs triggered by
the TSC (more than 15%).
The differences observed are further exemplified by the variations in the number
of different types of exceptions (including CPU-related and NMI) activated during
the various experiments, out of the 255 possible exceptions. This is illustrated by
Fig. 8.17 that shows the distribution of the exceptions observed for the three HWIFI
techniques considered.
However, it is worth noting that for EI, most of the experiments exercised CPU
EDMs (especially when using the probe without antennas), which reveals the very
restricted spectrum of the type of errors generated by this technique. However, the
variation in the error set was somewhat enhanced when the antennas were used.
8.3.5
Some Lessons Learned
Fault injection is a powerful and suitable technique for objectively assessing the
efficiency of fault tolerance mechanisms. Based on the analysis reported here and
other related studies conducted at different levels, e.g., see Moraes et al. ( 2006 ), the
quest for the “unique best technique” looks somewhat unproductive.
8 Due to the small number of NMIs observed, such an analysis was not carried out for SWIFI.
 
 
 
Search WWH ::




Custom Search