Spring 2002 1 Efficient Dissemination of Enterprise Summary Data to Mobile Clients Mohamed A. Sharaf University of Pittsburgh
Dec 29, 2015
Spring 2002 1
Efficient Dissemination of Enterprise Summary Data to Mobile Clients
Mohamed A. SharafUniversity of Pittsburgh
2
Motivation
“…Currently handheld and palmtop computers are widely used for personal information management. In the near future they will also be used to access enterprise data…”, IBM Corp., 2000
“…For organizations to be successful in today's fast-paced digital economy, decision makers require access to all business-critical information on any platform. Wireless devices are quickly becoming alternative platforms for e-enabling the enterprise, as they provide instant access to relevant enterprise data for Mobile Decision Making…”, Hummingbird Communications Ltd., 2000
3
Outline
OLAP (On-Line Analytical Processing) Data ModelWireless OLAP ModelScheduling AlgorithmsSimulation ResultsConclusion
4
Multi-Dimensional Model [Codd93]Pro
duct
TV
VCRPC
Date 1Qtr 2Qtr 3Qtr 4Qtr
Cou
ntr
y
U.S.A
Canada
Mexico
Group-By (P,C,D),
Sum(Sales)
Dimensions
Measures
5
A Sample Data CubePro
duct
TV
VCRPC
Date 1Qtr 2Qtr 3Qtr 4Qtr
Cou
ntr
y
U.S.A
Canada
Mexico
G(P,C)
G(P)
Derivation Dependency
6
Traditional OLAP Server
Point to Point Access
7
Wireless OLAP Server
Broadcast
Uplink Channel
Power Consumption
8
Wireless Environments
Asymmetry in the communicationBroadcast for data dissemination Periodic (push-based) On-Demand (pull-based) Hybrid
A broadcast schedule determines what and when to broadcast Metrics Access Time = Wait + Tune Power Consumption = Active + Doze
9
On-Demand Scheduling Algorithms
First-Come First-Serve (FCFS)Shortest Service Time First (SSTF)
RxW: broadcast a page either because it is popular or because it has at least one long-outstanding request [Franklin 99]
Most Request First (MRF) [Ammar 86]
Summary Tables : 1) Heterogeneous2) Skewed Access3) Derivation Dependency
10
Broadcast Organization
Header Packet = Identifier + PointerA table TX is characterized by set of dimensional attributes X . TX subsumes TY, iff Y X, similarly, TY is dependent on TX
X is the dimensionality degree
100 G(Supp) 111 G(Supp, Prod, Cust) … Tune Wait Tune Tune
Target TableHeader Table
11
RxW Variants
Strict RxW/S: For each request QX for a summary table TX, the
server maintains the following values: R: The number of requests for TX. W: The age of the first request has for table
TX. S: The size of table TX.
Table with highest RxW/S is the one to broadcast.
Flexible RxW/S: Decision is same as RxW/S, but using “Derivation
Dependency” allows: Server to remove dependent tables from queue Client to tune to the first subsuming table
12
Controlling the Flexibility
Why ? Compromise between access time and
power consumption
How ? Integrate derivation dependency with
scheduling decision Classify dependents into beneficial and
impairing according to dimensionality
13
d3d1 d2 d4
d1,d2 d1,d3
d1,d2,d3 d1,d2,d4 Benefit(B)
Impairment(I)
d3d1 d2 d4
d1,d2 d1,d3
d1,d2,d3 d1,d2,d4 Benefit(B)
Scheduling Intuition
d1,d2,d3,d4,d5
d5
d4,d5
d3d1 d2 d4
d1,d2 d1,d3
d1,d2,d3 d1,d2,d4 d3,d4,d5
d1,d2,d3,d4 d2,d3,d4,d5d1,d2,d3,d5
QX
distance = X /2
14
Benefit/Impairment Scheduling (BI)
• The BI for Qi is computed as:
Bj Ikkiji
Bj Ikkji
Bj Ikkji
SSSS
WWWRRR
)(
)()(
• The highest BI value request is broadcast next and dependents B are removed
• A priority queue is used to store requests
15
Experiments• A synthesized six-dimension lattice.• Packet capacity = 10 attribute values• Each Mobile host poses 100 queries
according to a Zipf distribution• Each experiment was run 5 times• Metrics:
• Average Access Time in simulation ticks• Average Power consumption in doze units
• Active power = 20 times doze power• Fairness: Standard Deviation of requests’
stretch [Acharya 98]• stretch = access time/service time
16
Average Access Time
Number of Clients
0 50 100 150 200
Acc
ess
Tim
e (
Sim
ula
tion
Pa
cke
ts)
0
5000
10000
15000
20000
25000
SSTFFCFSRxWRxW/SFlex. RxW/SBI
17
Power Consumption
Number of Clients
0 100 200 300
Pow
er
Con
sum
ptio
n (D
oze
Uni
ts)
6000
8000
10000
12000
14000
16000
18000
RxWRxW/SFlex. RxW/SBI
18
Fairness
Number of Clients
0 50 100 150 200
Fa
irn
ess
Me
tric
1
10
100
1000
10000
SSTFFCFSRxWRxW/SFlex. RxW/SBI
19
Varying Skewness
Zipf Parameter (
0.0 0.2 0.4 0.6 0.8 1.0
Acce
ss T
ime
(S
imu
latio
n P
acke
ts)
0
5000
10000
15000
20000
25000
30000
35000
SSTFFCFSRxWRxW/SFlex. RxW/SBI
20
Conclusion
We introduced the new problem of scheduling objects with a derivation dependency propertyWe proposed a variety of scheduling algorithms that minimize access time and preserve power consumption
Load AAT PC
Low BI RxW/S
Med BI RxW/S
High Flex. RxW/S BI
65% less than RxW & 55%
less than RxW/S70% less than RxW & 55%
less than RxW/S77% less than RxW
15% less than RxW20% less than RxW
24% less than RxW
21
Future Work
We are planning to extend the research to include: Subscribe push environment Caching mechanisms More detailed cost model