Information Technology Reference
In-Depth Information
list of critical variables
list of critical
event locations
CodeSurfer
instrumented
program
source
*.c *.h
program
source
*.c *.h
monitoring
costs
Model
Generator
Instrument
Script
yices
yices model
speciication
instrumentation
scheme
clang
llvm-opt
CFG with estimated
execution times and weights
Fig. 6.
HyRV instrumentation toolchain for C applications
Tabl e 1.
Monitor cost configurations [clock cycles]
Configuration
c
hist
c
ET
c
TT
c
E→T
c
T→E
1
50 100 100 100
100
2
50 100 100 150
150
3
50 150 150 100
100
4
50 150 150 150
150
5
50 250 250 100
100
6
50 250 250 150
150
changing variables. Instructions that potentially change the value of these vari-
ables form the set of critical instructions monitored in the experiments. For each
program, the monitoring overheads were measured using the cost configurations
(listed in Table 1) and associated instrumentation schemes. The cost configura-
tions depend on the implementation of the monitor (e.g., running on the same
processor, distributed). We use the configurations in Table 1 to demonstrate that
the instrumentation schemes may change as a result of the relative differences
in the elementary monitoring costs.
4.2 Experimental Results
We classify the results of our experiments based on the generated instrumenta-
tion scheme and runtime overhead:
1. The first class consists of cases, where our ILP model suggests a hybrid
monitor and the monitor indeed significantly outperforms an ET or TT
monitorinpractice(seeFigure7).
2. The second class consists of cases where the ILP model suggests either an
ET or TT monitor and the suggested solution indeed outperforms other
monitoring modes (see Figure 8).
3. The third class consists of cases where the solution to the ILP model ei-
ther exhibits slight improvement over other monitoring modes or it slightly
underperforms in practice (see Figure 9).
