Peer-Based Collaborative Caching and Prefetching in Mobile Broadcast - Advanced Operating Systems and Kernel Applications - page 255

experimental results

ACP has the largest number of broadcast hits: most
of the queries are answered using the broadcast
channel rather than local cache or neighbors'
cache. CPIX is between them.
The interesting finding here is that the num-
ber (or the ratio) of local cache hits and neighbor
cache hits is not the key factor that determines
the schemes' performance (average response
time). Figure 3(c) shows the key factor, the aver-
age response time of broadcast hit s. We see that
DGCoca experiences the longest average waiting
time for the broadcast hits (neighbor cache misses),
and ACP experiences the shortest average waiting
time for the broadcast hits. This is why DGCoca
has the longest average response time although
it has the largest number of local cache hits and
neighbor cache hits, whereas ACP has the best
response time although it has the largest number
of broadcast hits.
Since the response times for local cache hits
and neighbor cache hits are really short, the ac-
cess latency caused by neighbor cache misses
(i.e. the requested data object is in neither local
cache nor neighbors' cache) -- the product of the
number of broadcast hits and the average waiting
time for a broadcast hit -- determines the average
waiting time of a scheme. Although DGCoca has
fewer broadcast hits, it takes a longer time to get
a broadcast hit if a query is not answered by local
cache or neighbors' cache. On the contrary, in
ACP, although more queries are answered after
broadcast hits, a mobile peer waits a shorter time
before the required data object appears on the
broadcast channel.
Recall that DGCoca does not consider the data
availability on broadcast channel, CPIX consid-
ers the data availability on broadcast channel,
and ACP further exploits the dynamics of data
availability on broadcast channel. The key idea
we have in CPIX and ACP is that both the data
availability from neighborhood and the data
availability on broadcast channel are important
for making caching or prefetching decisions. For
data objects with similar access probability, the

The performance metric used in the experimental
study is the average response time (access latency)
measured in broadcast unit. We first study the
performance of the schemes under the default
parameter setting. Then we investigate the effects
of some system parameters on the schemes.

Basic Performance Study

Here we investigate the performance of the
schemes under the default parameter settings.
The objectives are twofold: one is to compare
individual schemes with cooperative schemes; the
other is to study the performance differences of
the cooperative schemes and to find the reasons
behind the differences.
Figure 3(a) compares the response times of
the five caching management schemes, namely
PIX, DGCoca, CPIX, PT, and ACP. Recall that
PIX is an individual caching scheme, PT is an
individual prefetching scheme, DGCoca and
CPIX are two cooperative caching schemes,
and ACP is a cooperative prefetching scheme.
From the figure, we first confirm that prefetch-
ing schemes perform better than demand-driven
caching schemes: PT performs better than PIX,
and ACP performs better than both DGCoca and
CPIX. We also confirm that cooperative schemes
perform better than individual schemes: DGCoca
and CPIX perform better than PIX, and ACP
performs better than PT.
Among the cooperative schemes, CPIX per-
forms better than DGCoca, and ACP has the best
performance. Figure 3(b) and Figure 3(c) reveal
the underlying reasons.
Figure 3(b) shows the breakdown of query hits
in each scheme. We observe that among the three
cooperative cache management schemes DGCoca
has the largest number of local cache hits and
neighbor cache hits, and the smallest number of
broadcast hits. This shows that DGCoca utilizes
the local cache and neighbors' cache very well.