Upload
sara-mcdonough
View
215
Download
0
Tags:
Embed Size (px)
Citation preview
On Scheduling Vehicle-Roadside Data Access
Yang Zhang Jing Zhao and Guohong Cao
The Pennsylvania State University
2
VANET 2007, Sept 10th, Montreal, Canada
The Big Picture
Vehicular Ad-hoc Networks - VANET Moving Vehicles RoadSide Units(RSU)
Local broadcasting infostations 802.11 access point
Applications Commercial Advertisement Real-Time Traffic Digital Map Downloading
Task Service Scheduling of Vehicle-
Roadside Data Access
3
VANET 2007, Sept 10th, Montreal, Canada
Challenges
Bandwidth Competition All requests compete for the
same limited bandwidth
Time Constraint Vehicles are moving and they
only stay in the RSU area for a short period of time
Data Upload/Download The miss of upload leads to
data staleness
Service Queue
Wireless Channel
4
VANET 2007, Sept 10th, Montreal, Canada
Assumptions and Performance Metrics
Assumptions Location-aware and Deadline-aware The RSU maintains a service cycle Service non-preemptive
Performance Metrics Service Ratio
Ratio of the number of requests served before the service deadline to the total number of arriving requests.
Data Quality Percentage of fresh data access
5
VANET 2007, Sept 10th, Montreal, Canada
FCFS, FDF, SDF
First Come First Serve (FCFS): the request with the earliest arrival time will be served first.
First Deadline First (FDF): the request with the most urgency will be served first.
Smallest Datasize First (SDF): the data with a small size will be served first.
workload
6
VANET 2007, Sept 10th, Montreal, Canada
D*S Scheduling
Intuition Given two requests with the same deadline, the one
asking for a small size data should be served first Given two requests asking for the data items with same
size, the one with an earlier deadline should be served first
Basic Idea: Assign each arrival request a service value based on its
deadline and data size, called DS_value as its service priority weight
DS_value=(Deadline-CurrentClock)*DataSize
7
VANET 2007, Sept 10th, Montreal, Canada
The Implementation of D*S
Dual-List Search from the top of D_list Set MinS and MinD Search D_List and S_list
alternatively
Stops when the checked entry goes across MinD or MinS, or when the search reaches the halfway of both lists.
8
VANET 2007, Sept 10th, Montreal, Canada
Download Optimization: Broadcasting
Observation some requests may ask for downloading the
same data item wireless communication has the broadcast
capability Basic Idea
delay some requested data and broadcast it before the deadlines, then several requests may be served through a single broadcast
the data with more pending requests should be served first
DSN_value=(Deadline-CurrentClock)*DataSize/Number
9
VANET 2007, Sept 10th, Montreal, Canada
D*S/N: The Selection of Representative Deadline
When calculating their DSN value, we need to assign each pending request group a single deadline to estimate the urgency of the whole group.
10
VANET 2007, Sept 10th, Montreal, Canada
The Problem of D*S/N
Data Quality!!
DSN_value=(Deadline-CurrentClock)*DataSize/Number For upload request, it is not necessary to maintain several
update requests for one data item since only the last update is useful
Number value of update requests is always 1, which makes it not fair for update requests to compete for the bandwidth
D*S/N can improve the system service ratio but sacrifice the service opportunity of update requests, which degrades the data quality for downloading
11
VANET 2007, Sept 10th, Montreal, Canada
Upload Optimization: 2-Step Scheduling
Basic Idea: two priority queues: one for the update requests and the
other for the download requests. the data server provides two queues with different
bandwidth (i.e., service probability)
Benefits of Using Two Separate Priority Queues we only need to compare the download queue and
update queue instead of individual updates and downloads
update and download queues can have their own priority scheduling schemes
12
VANET 2007, Sept 10th, Montreal, Canada
Step I: Update Queue or Download Queue
Service_Profit: the sum of the profit gained from upload and download
Service_Profit = Update_Profit
+ FreshDownload_Profit
+ StaleDownload_Profit
Assume one update request can contribute the same profit as one download request with fresh data.
Assume the profit of download with stale data will degrade with α
Bandwidth Allocation (ρ): the download requests share ρ of the bandwidth and the update requests share the rest, 1- ρ
Goal: set ρthus achieving the balance between update and download
13
VANET 2007, Sept 10th, Montreal, Canada
Step I (cont.)
ru and rd : service rate of update and download requests
Update profit rate depends on: service rate, ru, and
allocated bandwidth, 1- ρ
Update_Profit≈ ru (1- ρ) t Download profit rate relies on
service rate, rd,
bandwidth allocation ρ, and data quality for each download.
FreshDownload_Profit≈ rd ρ (1- ρ) t
StaleDownload_Profit≈ rd ρ2 α t
14
VANET 2007, Sept 10th, Montreal, Canada
Step I (cont.)
Service_Profit ≈ ru (1- ρ) t
+rd ρ (1- ρ) t
+ rd ρ2 α t (0≤ ρ ≤1)
)rr(1)α(1
rr 0
)rr(1)α0(1),α)r2(1
rrmin(
ρ
du
du
dud
ud
15
VANET 2007, Sept 10th, Montreal, Canada
Step II: D*S/N and D*S/R
D*S/N for download queue D*S/R for update queue
Given two update requests with the same d*s value, the request that updates hot data should have a higher service priority
R is the service rate of the requests on corresponding data item in the download queue
DSR_value = (Deadline-CurrentClock)*DataSize/R
16
VANET 2007, Sept 10th, Montreal, Canada
Implementation of 2-Step Scheduling
Start a new adaptation window;Adjust ρ
Is adaptation window valid?
Generate a random number Κ, 0≤ Κ ≤ 1
Κ < ρ ?
Scheduling in download queue[D*S/N Scheme]
Scheduling in update queue[D*S/R Scheme]
Service Processing
Yes
Yes No
No
INITIATE
Is download queue empty?
Is update queue empty?
No NoYes Yes
Step I
Step II
The workload is examined with a time periodτ (adaptation window)
At the beginning of eachτ, ρ is re-calculated
17
VANET 2007, Sept 10th, Montreal, Canada
Simulation Setup
NS-2 based 400m*400m square street
scenario One RSU server is located at
the center of two 2-way roads 40 vehicles randomly deployed
on each lane Each vehicle issues request
with a probability p Access pattern of each data
item follows Zipf distribution
18
VANET 2007, Sept 10th, Montreal, Canada
Performance Evaluation: Effect of Workload
As workload increases, D*S/N can achieve the highest service ratio while its data quality degrades dramatically
19
VANET 2007, Sept 10th, Montreal, Canada
Performance Evaluation: Effect of Access Pattern(θ)
• Change of θ does not have too much impact on the performance of FCFS, FDF, SDF and D*S• D*S/N and 2-Step can benefit from the skewness of the data access pattern with the increase of θ
20
VANET 2007, Sept 10th, Montreal, Canada
Performance Evaluation: Effect of Access Pattern(Download/Update Ratio)
21
VANET 2007, Sept 10th, Montreal, Canada
Performance Evaluation: Adaptivity to Workload Condition Change
•2-Step scheme can achieve good performances in almost all scenarios.• ρ adapts quickly when workload condition changes
22
VANET 2007, Sept 10th, Montreal, Canada
Conclusion
We addressed some challenges in vehicle-roadside data access
We proposed a basic scheduling scheme called D*S to consider both service deadline and data size when making scheduling decisions.
To make use of the wireless broadcasting, we proposed a new scheduling scheme called D*S/N to serve multiple requests with a single broadcast.
We also proposed a Two-Step scheduling scheme to provide a balance between serving download and update requests.
Simulation results show that the Two-Step scheduling scheme outperforms other scheduling schemes. Further, the Two-Step scheduling scheme is adaptive to different workload scenarios.