Information Technology Reference
In-Depth Information
higher. Besides, when a RP receives much more events than other RPs, its matching
costs may also be higher. In this paper, we relieve stresses of overloaded RPs by
distributing matching overheads over more nodes.
When the RP related to the frequent itemsem
FI
is overloaded, its matching over-
heads are distributed to multiple nodes, and the number of newly increased nodes is
decided by the number of attribute in
FI
. The detailed policy is presented as follows.
If
FI
is selected as the attribute set for subscription installation, the subscriber ranks
attribute names in
FI
in different order, and
A
n
keys will be obtained, where
n
is the
number of attribute in
FI
. Then one key is randomly selected and the subscription
will be sent to the corresponding node in the overlay network. For example, when
the RP related to
is overloaded, 6 keys can be obtained by hashing string
A
1
A
2
A
3
,
A
1
A
3
A
2
,
A
2
A
1
A
3
,
A
2
A
3
A
1
,
A
3
A
1
A
2
,and
A
3
A
2
A
1
. As a result, subscriptions
that used to be stored on one RP will be distributed over 6 nodes, and the matching
overheads will be reduced on each node. However, in order to realize exhausting
matching, events that contain
FI
should be sent to all the newly increased nodes for
matching. So the load balancing strategy may increase the event publication cost
slightly.
{
A
1
,
A
2
,
A
3
}
4
Performance Evaluation
In this section, we evaluate the performance of DMPSS by comparing it with Ferry
and Eferry. A set of metrics are used to evaluate performances of the three systems,
including the total matching time on each RP, data volume of forwarded messages
on each node, and latency of the event. As mentioned above, overhead for matching
is affected by three factors: the number of subscriptions stored on the RP, the number
of events the RP receives and the matching complexity. For evaluating the match-
ing load comprehensively, we calculate the total matching time that is consumed on
each RP during the event publication process. Data volume of forwarded messages
indicates overhead for message transmission, and it also reflects the bandwidth con-
sumption in the network. Latency is one of the most important performance metric
for publish/subscribe systems, and it is denoted by the time duration from event
publication to subscriber reception.
4.1
Experimental Setup
We built DMPSS, Ferry, and Eferry on OverSim [20] platform, which has imple-
mented many DHT protocols. In order to realize equal comparison, Chord, which
was used in Ferry and Eferry, is applied in DMPSS. Moreover, the schema used in
our simulation is also the stock quotes model used in Ferry and Eferry. The defini-
tion of the schema is as follows:
Search WWH ::
Custom Search