Information Technology Reference
In-Depth Information
50000
ET-only
TT-only (
SP
=10
,LSP
)
HyRV (
SP
=10
,LSP,Cx→y
= 100
)
HyRV (
40000
SP
=10
,LSP,Cx→y
= 150
)
TT-only (
SP
=20
,LSP
)
HyRV (
SP
=20
,LSP,Cx→y
= 100
)
30000
HyRV (
SP
=20
,LSP,Cx→y
= 150
)
20000
10000
0
C
ET
=
C
TT
= 100
C
ET
=
C
TT
= 150
C
ET
=
C
TT
= 250
Monitoring Cost
Fig. 7.
Monitoring overhead of
crc
for three monitoring modes under all cost
configurations
In the rest of this section, we will discuss the experimental results and focus
on one program from each class. We note that the three other programs not
specifically discussed in this section exhibit similar results.
Hybrid Monitor with Significant Improvement.
The program represent-
ing this class (i.e.,
crc
with CFG of the size 65 vertices and 82 arcs) has two
characteristics: it has (1) two tight loops, each containing one critical instruction,
and (2) a relatively large initialization function that contains only non-critical
instructions. Intuitively, if the program is monitored by an ET monitor, then
the tight loops in the program will cause monitor invocations for each iteration.
This is an instance where a burst of events creates a large overhead over a short
period of time (similar to the timeline in Figure 1). In such cases, an ET monitor
suffers.
On the contrary, the large initialization function does not contain critical
events; hence, a TT monitor would suffer from redundant sampling overhead.
We hypothesize that the combination of these two characteristics can exploit
the benefits of employing a hybrid monitor. The graph in Figure 7 validates
our hypothesis. As can be seen, in all cost configurations, the hybrid monitor
incurs significantly less overhead than both the ET monitor and TT monitor op-
erating with the same sampling period. Another interesting observation is that
increasing the cost of ET and TT monitor invocations does not greatly increase
the overhead of the hybrid monitor. This is because the hybrid monitor only
samples when the program reaches its tight loop, which reduces the cost of mon-
itoring frequently occurring critical events by buffering them into memory before
sampling. In addition, the monitoring scheme reduces the number of redundant
samples by letting the monitor run in ET mode when critical events are infre-
quent. In such cases, the behavior of a hybrid monitor is quite robust when the
cost of monitor invocation increases.
Time-Triggered Monitor with Significant Improvement.
The common
characteristic of the member programs of this class (i.e.,
bs
,
fibcall
,
insertsort
,
