Database Reference
In-Depth Information
root causes of execution problems. Namely, they attempt to uncover
underlying
interaction
problems.
3. Troubleshooting Interactive Complexity
A number of recent analysis techniques in sensor network debugging
literature aim to uncover root causes of errors resulting from anomalous
interactions
among large numbers of components. Modern networked
sensing systems feature heterogeneity and tight interactions between
computation, communication, sensing, and control. Tight interactions
breed interactive complexity; the primary cause of failures and vulner-
abilities in complex systems [76]. While individual devices and subsys-
tems may operate well in isolation, their composition might result in
incompatibilities, anomalies or failures that are typically very dicult
to troubleshoot. On the other hand, software re-use is impaired by the
customized nature of application code and deployment environments,
making it harder to amortize debugging, troubleshooting, and tuning
cost.
Techniques discussed in this section are grounded in the assumption
that today's professional developers are very good at debugging
individ-
ual components
. Hence, in large systems, unresolved problems arise not
from individual component failures, but from unexpected complex inter-
actions that evade developer imagination (and defeat scalability limits
of pre-run-time analysis tools).
Data mining offers a solution to the troubleshooting problem by em-
pirically uncovering those event sequences or patterns that led to the bad
states. Rather than exploring the entire space of possible states, data
mining solutions use ecient pruning techniques to focus the search on
those patterns that are correlated with anomalous behavior. The data
mining approach further exhibits two important qualities that makes it
suitable for the debugging techniques presented in this section:
Exploiting non-reproducible behavior:
Most hard-to-find bugs are
hard to reproduce. Data mining approaches are good at exploit-
ing non-determinism to improve understanding of system behav-
ior. For example, discriminative mining requires examples of both
good and bad system behavior to be able to isolate conditions cor-
related with good and bad. Non-reproducible bugs are thus inher-
ently suited for analysis using data mining approaches as the lack
of reproducibility itself and the inherent system non-determinism
improve the odds of occurrence of suciently diverse behavior ex-
amples to help the troubleshooting system understand the relevant
correlations and identify causes of problems.
