adrian daniel popescu, mohamed a. sharaf, cristiana amza mdm 2009 sla-aware adaptive on-demand data...

Adrian Daniel Popescu, Mohamed A. Sharaf , Cristiana Amza

MDM 2009

SLA-Aware Adaptive On-Demand Data

Broadcasting in Wireless Environments

1

OutlineIntroduction

Motivation

System model

Tardiness-aware scheduling -- SAAB-T

Utility-aware scheduling -- SAAB-U

Experiments

Conclusions

2

Introductionin a dynamic mobile environment

motivated us to use Service Level Agreements (SLAs) where a user specifies the utility of data as a function of its arrival time.

SLAs provide users with the flexibility to define the utility of delayed data

Broadcasting• Pull-based(on-demand) • Push-based

3

IntroductionOn-demand data broadcasting

More scableusers submit requests for data items

of interest and the broadcast server aggregates requests for the same data item and broadcasts it only once.

If a data item is highly popular, then broadcasting that data item to all interested users substantially reduces the number of transmissions.

4

Motivationoptimizing response time is not sufficient to

maximize data usability since it overlooks the user’s requirements and expectations

SLA-aware adaptive data broadcast (SAAB) scheduling policy for maximizing the system utility under SLA-based performance measures

minimizing response time or drop rate

5

System modelon-demand data broadcasting environment

6

System modelRequest1) Data Item (Ii): which is the data item

corresponding to request Ri,2) Arrival Time (Ai): which is the point of

time where request Ri is issued, and3) Deadline (Di): which is the soft deadline

associated with request Ri.

7

System modelScheduling Queue1) Service Time (Cj ): is the time required

for transmitting data item Ij on the downlink channel

2) Popularity (Pj ): is the number of pending requests for data item Ij , and

3) Requests Vector (Rj ): is a requests vector of length Pj , where each entry in Rj corresponds to one of the Pj pending requests for Ij (i.e., Rj = Rj1 Rj2 … RjPj )

8

Tardiness-aware scheduling -- SAAB-T

9

there is some pending request Rx for that data item Ix with deadline Dx

Slack :

assume two data items I1 and I2schedule of choice (1)X I1 first,then I2

(2)Y I2 first,then I1

)( xxx CtDS

Tardiness-aware scheduling -- SAAB-T

10

Under schedule XAssume that the number of these requests

is P1,def , where 0≤ P1,def ≤ P1.The sum of all requests which currently

have a negative slack, S1,j

tardiness of requests for item I1

defP

jjX ST

,1

)( ,1,1

Tardiness-aware scheduling -- SAAB-T

11

Assume that the number of these requests is P2,def , where 0≤ P2,def + P2,add ≤ P2

The sum of all requests which currently have a negative slack, S2,j

Pending requests to I2 had positive slack when the scheduling decision was made but that slack became negative

tardiness of requests for item I2

adddef PP

jjX SCT

,2,2

)( ,21,2

Tardiness-aware scheduling -- SAAB-T

12

adddefdef PP

jj

P

jjX SCST

,2,2,1

)()( ,21,1

adddefdef PP

jj

P

jjY SCST

,1,1,2

)()( ,12,2

total tardiness under schedule X and Y

Tardiness-aware scheduling -- SAAB-T

13

Example:C1 = 5, C2 = 10

2 R21 R22 R23

Deadline 7 5 13

Negative slack

-3 -5 3

add 2

1 R11 R12 R13

Deadline 3 8 18

Negative slack

-2 3 13

add 7

T1y = -(-3-5)=8T2y=-(-2)+(10-3) =9Ty = 17

T1x = -(-2)=2T2x=-(-3-5)+(5-3) =10Tx = 12

)( xxx CtDS

Tardiness-aware scheduling -- SAAB-T

14

If Tx < Ty,

priority:

)(1 ,

,,,

addiP

j

jiaddidefi

iri

C

SPP

CP

))1(

(1 ,

,,,

addiP

j

jiaddidefi

iri

Cn

SPP

CP

)(1

)(1 ,2,1

1

,2,2,2

22

,1,1,1

1 addadd P

j

jadddef

P

j

jadddef C

SPP

CC

SPP

C

Utility-aware scheduling -- SAAB-U

SLA -- Utility functionSAAB-U -- Utility function

15

Utility-aware scheduling -- SAAB-U

16

each request for I2 is delayed by the amount of time needed to transmit I1

general priority function

21

)()( ,21,2,1,1

P

jjj

P

jjjX ECuEuU

12

)()( ,12,1,2,2

P

jjj

P

jjjY ECuEuU

ii P

jjiji

P

jjijiri CnEuEuP ))1(()( ,,,,

Experiments

17

SAAB-TSlack factor (SF)

SJF = 1/Ci high loadEDF = 1/Di low loadMRF = Pi W-SJF = Pi/Ci

Experiments

18

SAAB-U

Conclusions

19

This request aggregation is efficient since it allows for fewer data broadcasts which saves bandwidth and reduces delays.

minimizing response time, or drop rate

SAAB is highly sensitive to workload conditions