Information Technology Reference
In-Depth Information
Algorithm 2. Event Publication
Require:
e //the event
1: set Att r e = ext ract ( e ) ;
2: for each i [ 0 , n 1 ] do
3: if IFContain ( Att r e , FIS [ i ] . it emset )== true then
4: key = hash ( LexiOrder ( FIS [ i ] . it emset )) ;
5: send ( sub , key ) ;
6: end if
7: end for
8: for each i [
0
, m
1
] do
9:
key
=
hash
(
LexiOrder
(
VFIS
[
i
] .
it emset
))
;
10:
send
(
sub
,
key
)
;
11: end for
3.6
Event Delivery
RPs deliver the event e to all the matched subscribers after matching processes. In
DMPSS, we use point-to-point communication to reduce cost on overlay routing and
transmission. When a subscriber installs a subscription, it also sends its IP address
to the RP. After the matching process finishes, the RP counts the number of matched
subscribers, e.g. l , and collects their IP addresses. Then it calculates a transmission
degree d as follows:
W
S e
where W is the current available upload bandwidth, and S e is the size of event e .
If l
d
=
d , the RP sends event e to all the matched subscribers through direct point-
to-point communication. If l
>
d , the RP divides all the matched subscribers into d
groups, and randomly selects one node from each group as the group leader. Then
it sends e and the IP list of the group to each leader. When a leader receives such
a message, it has two choices. One of them is sending e to all the nodes in the IP
list through direct point-to-point communication. The other is further dividing these
nodes into groups and delivering e and the corresponding IP list to each sub-leader.
A leader makes its choice also according to its available upload bandwidth.
3.7
Load Balancing Strategy
As mentioned above, the load on a node includes overheads for matching and over-
heads for message transmission. In DMPSS, load balancing performance for mes-
sage transmission is mainly affected by the routing policy of the DHT network, so
we only provide load balancing strategy for matching.
The non-uniform distribution of matching load on nodes results from the skewed
dataset of the real world. For example, in DMPSS, RPs related to frequent itemsets
with higher supports will store more subscriptions, and their matching costs may be
Search WWH ::




Custom Search