A Data Mining Based Publish/Subscribe System over Structured Peer-to-Peer Networks - Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing - page 12

102400 events for matching. However, the RP that stores the most subscriptions
only received 51331 events for matching. As a result, the matching loads on RPs
are balanced in DMPSS.
Besides, Table 1 and Table 3 illustrate that the average matching time for an
event is mainly affected by the number of stored subscriptions on the RP. How-
ever, in DMPSS, RP 3 is an exception. It stores fewer subscriptions than RP 1and
RP 2, but its average matching time for an event is the longest. This is because RP 3
is responsible for frequent itemset

{ Symbol , Open } . Different with other template

subscriptions, in template subscription T 1 = { ( ,

Open is limited by a maximal value and a minimal value, and the matching com-
plexity for Open is higher than that for other attributes. As a result, the average
matching time on RP 3 is higher. Actually, the total matching time on RP 3isalso
the longest in DMPSS. The results verify that the overhead for matching is affected
by multiple factors, and it should be evaluated comprehensively.
In order to alleviate the matching load on RP 3, we apply our load balancing
strategy. From Table 1 we can see that subscriptions that contain frequent item-
set

Symbol = P 1 ) ∧ ( P 2 ≤ Open ≤ P 3 ) }

are distributed over one more RP. Fig. 2 shows that the load
balancing performance is improved obviously when the load balancing strategy is
applied.

{ Symbol , Open }

4.2.2 Overhead for Message Transmission

We calculate the data volume of forwarded messages on each node to evaluate the
overhead for message transmission and the bandwidth consumption in the network.
From Fig. 3 we can see that DMPSS reduces the data volume of forwarded messages
on each node obviously. This is because the number of events that are published
for matching is reduced in DMPSS. Besides, the point-to-point communication we
used for event delivery reduces cost on overlay routing and transmission, and the
overhead for message transmission is reduced further. When our load balancing
strategy is applied, more event messages are published for matching, and the data
volume of forwarded messages increases slightly, as shown in Fig. 2.

4.2.3 Latency

Fig. 4 plots the cumulative distribution functions of event latency. We can see that
DMPSS decreases the time duration from event publication to subscriber reception
dramatically. The reduction in latency mainly results from two reasons. First, the
number of events that are published for matching is reduced, and the throughput
in the network is reduced accordingly. Second, the point-to-point communication
in DMPSS avoids time for overlay routing and transmission, and the latency is re-
duced correspondingly. As described in [6], compared to Ferry-Rnd, Ferry-Pred can
reduce hops, latency and overhead, so Ferry-Pred has better latency performance
than Ferry-Rnd. However, it is still worse than DMPSS. Eferry combines the event
publication into event delivery. This strategy can reduce the number of messages in