qos-aware packet scheduling by looking ahead approach
DESCRIPTION
QoS-Aware Packet Scheduling by Looking Ahead Approach. 有助於提高服務品質的前瞻式封包排程機制 政治大學資訊科學所 指導教授:連耀南 學生:溫永全 2007.12.27. Outline. Introduction Related Work Our Approach Performance Evaluation Conclusion. Introduction. - PowerPoint PPT PresentationTRANSCRIPT
QoS-Aware Packet Scheduling by Looking Ahead Approach
有助於提高服務品質的前瞻式封包排程機制
政治大學資訊科學所指導教授:連耀南
學生:溫永全2007.12.27
Outline 2
Outline
Introduction Related Work Our Approach Performance Evaluation Conclusion
Introduction 3
Introduction
現今封包網路中, router 對於封包是以 FIFO 的方式處理,並未針對封包的時效性做適當的處理,對於遲到的封包而言,在目的地端幾乎已經沒有補救機會
如果能在傳遞過程中依封包的時效性及重要性做適度的 rescheduling ,讓過遲的封包提前送出,而將有時間餘裕的封包稍緩送出,如此截長補短,可提高網路效能及整體 QoS滿意度
Introduction 4
3GPP UMTS QoS 分類架構
即時性的應用服務,像是 VoIP 以及 VoD ,對於封包的 delay time 以及 jitter 具有嚴格的要求
不同 class 有不同的服務品質要求
Delay Sensitivity
Jitter Sensitivity
Packet Loss Sensitivity
EX.
Conversational High High Low VoIP
Streaming Medium High Low VoD
Interactive Medium Low High Telnet
Background Low No High E-Mail
Introduction 5
如何知道封包早到或遲到 在 ALL-IP Network 上,所有應用皆在同一
個封包交換網路上傳送 網路服務品質管理
QoS 管理架構 ( 例如: IntServ 及 Budget Based QoS) 封包均循有資源管理的指定路徑傳遞 可預定封包到達每一個 hop (router) 的時間及其
送出時間 當封包到達每個 router ,可以知道該封包
早到或遲到
Introduction 6
router 架構
packet forwarding 會依照封包所要傳送的目的地,選取適當的 output queue 將封包插入
我們的排程方法將對個別的 output queue 中的封包進行排程
Scheduler
Input Queue
Output Queue
PacketForwarding
Scheduler
Scheduler
Introduction 7
排程環境 我們的研究假設在兩種排程環境下進行
獨立排程 (Intra-Router Scheduling) 每個 router 根據自身所知的資訊而不參考其他 router
的狀態逕自進行排程 前瞻式排程 (Looking Ahead Scheduling)
router 會參考封包在後續路程上各 router 的狀態進行排程
packet
Router
Sender Receiver
Introduction 8
Our Motivation
per-hop QoS 排程是在每個 hop 內各自進行,如要達到最佳排程,各個 router 之間必須互相分享資訊 我們延續 A New Jitter Control Mechanism by Per-Hop
Packet Scheduling Approach 該研究中每個 router 各自做排程的決定
參考封包在後續路程上各 router 的狀態進行排程,進而改善為前瞻式排程 (example)
另外針對該研究中的 multiple queue 排程演算法,我們也提出改善的方法
Outline 9
Outline
Introduction Related Work Our Approach Performance Evaluation Conclusion
Related Work 10
Related Work文獻一 A New Jitter Control Mechanism by Per-Hop
Packet Scheduling Approach ( 後文簡稱 NJCM)
利用 profit function 的方式表達進入一個 router 的封包的時效性及重要性 當封包進入 router 時,可以得知此封包的服務等級及預
定送出時間 該研究提出與封包傳遞時間相關的四種 profit function ,
router 將視情況賦予該封包一個 profit function , scheduler 根據封包的時間參數及 profit function得到一個 profit ,再以該 profit 決定送出順序
Related Work 11
Four profit function
slope=S1
profit
timeTc
(current time)Td
(expected trans. time)
Ta(soft deadline)
1
profit=1 when Td≦ Taprofit=1+(Td-Ta)*S2 when Ta-(1/S2)≧ Td>Taprofit=0 when Td>Ta-(1/S2)
Tb(hard deadline)
slope=S2
profit
timeTc
(current time)Td
(expected trans. time)
Tb(hard deadline)
1
profit=1+(Td-Tb)*S1 when Td≦ Tbprofit=0 when Td>Tb
profit
timeTc
(current time)Td
(expected trans. time)
Ta(soft deadline)
1
profit=1+(Td-Ta)*S1 when Td≦ Taprofit=1+(Td-Ta)*S2 when Ta-(1/S2)≧ Td>Taprofit=0 when Td>Ta-(1/S2)
Tb(hard deadline)
(Slope-Down)
(Slope-Step) (Double-Slope)
slope=S1
slope=S2
profit
timeTc
(current time)Td
(expected trans. time)
Tb(hard deadline)
1
profit=1 when Td≦ Tbprofit=0 when Td>Td
(Step-Down)
hard local deadline
soft local deadline
pre soft deadline profit rate
post soft deadline profit rate
Related Work 12
兩種 output queue 架構 single preemptive queu
e router router 可將進入的封包插
到 output queue 的任意位置
multiple queue router output queue 是由數個 F
IFO queue 所組成, router只能將欲送出的封包插入其中一個 queue 的尾端
硬體較易實作 研究重點
Scheduler
Buffer Information
Packet In Packet Out
Pre-Scheduler Post-SchedulerPacket In
Buffer Information
Packet In
( 各有對應的排程演算法: SPQ-SA、MQ-SA)
Related Work 13
SPQ-SA (single preemptive queue)
Base case 封包可插入 queue任意位置
Scheduler 依封包 profit function 選擇適當的位置插入使封包送出順序獲得最大 total profit greedy algorithm
原本 queue 內封包順序維持不變,但進入的封包可任意插入 queue 之中
Related Work 14
MQ-SA (multiple queue)
可視為 SPQ 的 special case 只有固定位置可插入
Pre-Scheduler 依封包 profit function 選擇適當的 queue 插入尾端使封包送出順序獲得最大 total profit
Post-Scheduler 以單一 queue 為單位做輪詢 (round robin) 將單一 queue裡面的封包依序全部送出稱為對該 queue 服務
一次
Related Work 15
MQ-SA (cont.)
Post-Scheduler
Post-Scheduler
Related Work 16
Related Work文獻二 A Priority Based Packet Scheduler with
Deadline Considerations two sets of algorithms introduced
Static Priority with Deadline Considerations (SPD)
Dynamic Priority with Deadline Considerations (DPD)
not only simplify the complexity and overhead of a classical Earliest Deadline First (EDF) or Static Priority (SP) algorithm, but also provide a better QoS
Related Work 17
文獻二 (cont.) SPD
sorted according to their priorities
remaining deadline of the packet is checked discarded if the
packet has no remaining deadline line left
partial sorting DPD
modify the priority by remaining deadlines
Related Work 19
Related WorkSummary
NJCM 為獨立排程,並未納入後續 router負載因素;另一方面multiple queue 排程演算法尚有改善空間
其他相關研究沒有針對不同服務等級的服務提供不同優先次序的服務,在 router 對封包重新排程時,必須加以考慮不同類型封包的服務等級,讓資源依優先等級調配,如此可以提高整體滿意度。
Outline 20
Outline
Introduction Related Work Our Approach Performance Evaluation Conclusion
Our Approach 21
Our Approach
設計理念 改善 multiple queue 排程演算法
與 single preemptive queue 排程演算法比較,找出改善的方法
前瞻式排程 估計後續路程上 router 的負載 用以動態調整 profit function 參數,並找出最佳調
整參數
Our Approach 22
Design Challenges
Design Challenge of PMQ-SA 如何改善 multiple queue 排程演算法
Design Challenges of Looking ahead packet scheduling 後續網路狀態因素
如何估計後續 router負載 如何根據結果動態調整 profit function 參數
Our Approach 23
Objectives
改善 multiple queue 排程演算法 效能更逼近 single preemptive queue
前瞻式排程 反應網路狀況 各個 router 之間互相分享資訊以決定較佳的排程
因而提升網路資源使用率、提高 QoS 滿意度
Our Approach 24
Objective function (1/2)
整體 QoS 滿意度 (multiple class) total charge for multiple packet stream of class i
我們的目標就是要找出使 total charge(utility) 最大的 scheduling policy(profit function組合 )
)(3
1 1,,
i
N
jjiiji
j
Cqc
qi,j : quality index for the j-th flow of class iCi : unit price per byte of class iλi,j : number of bytes transmitted by the j-th flow of i (if i=1 or i=2) number of bytes received from the j-th flow of i (if i=3)Ni : number of flow of class i
Our Approach 25
Objective function (2/2)
quality index for a packet stream of class i
αi = 1 - (average delay time)i / Bi(1) (delay time satisfaction factor)
βi = 1 - (average jitter)i / Bi(2) (jitter satisfaction factor)
γi = 1 - (packet loss rate)i (loss rate satisfaction factor) ai (bi、 ci) : quality coefficient of delay time (jitter、 loss rate) for class iBi
(1) (Bi(2)、 Bi
(3)) : upper threshold of delay time (jitter、 loss rate) satisfaction factor
(3)ii
(2)ii
(1)ii
iii
Brate lossor
Bjitter averageor
Bdelay time averagefor 0
cba iii
iq
Our Approach 26
Design Challenge of PMQ-SA 如何改善 multiple queue 排程演算法 NJCM 中的 multiple queue
與 single preemptive queue比較,少最前面的位置以供插隊,導致非常緊急的封包也要排隊
priority multiple queue :讓其中一個 queue 的優先權提高以供較為緊急的封包插入
round-robin queue priority queue
Our Approach 27
如何改善 multiple queue 排程演算法 (cont.) PMQ-SA
round-robin queue作法跟 NJCM MQ-SA 一樣 只要 priority queue 中有封包就必須先服務
Pre-Scheduler Post-SchedulerPacket In
Buffer Information
Packet In
round-robin
Our Approach 28
Design Challenge 1 of Looking ahead packet scheduling如何估計後續 router負載 估計剩餘傳遞時間
利用 Routing Table 中的資訊 反應網路實際狀況及負載 進行後續路程封包排程
假設 router使用 link state routing protocol (e.g. OSPF) 提供幾項 routing metrics計算 link cost
流量負載 可用頻寬 delay time (RTT/2)
Our Approach 30
Design Challenge 2 of Looking ahead packet scheduling 如何動態調整 profit function 參數
IP 封包的傳遞時間起伏不定,使得封包已傳遞時間比預定的時間早到或遲到
當封包到每個 router 即可得知估計的剩餘傳遞時間 (by Routing Table) 比預定剩餘傳遞時間長或短
packet
預定已傳遞時間 預定剩餘傳遞時間
實際已傳遞時間 估計剩餘剩餘傳遞時間
Router
Sender Receiver
Routing Table
早到或遲到? 長或短?
Our Approach 31
封包傳遞中的幾種狀況
遲到
遲到
早到
早到 較預定剩餘傳遞時間長
較預定剩餘傳遞時間短
Router
packet
Sender Receiver
最不緊急:deadline往後延以補救其他封包
最緊急:deadline往前調以補救該封包
較預定剩餘傳遞時間長
較預定剩餘傳遞時間短
Our Approach 32
動態調整 profit function 的 local deadline
Step-Down & Slope-Down hard deadline += c. (dpre+dpost)
Slope-Down & Double-Slope soft deadline += c. (dpre+dpost)
以調整係數 c 決定調整程度 (o<c<1) 如何決定 c
以實驗找出最佳值
名字 ( 以最緊急狀況為例 ) 符號 定義已拖延的時間 dpre 預定已傳遞時間 – 實際已傳遞時間
後續可能延遲的時間 dpost 預定傳遞剩餘時間 – 估計剩餘傳遞時間
估計總拖延時間 dpre+dpost
調整係數 c 調整程度
Our Approach 33
profit profit
profit profit
time
time
time
time
hard local deadline
soft local deadline
soft local deadline
hard local deadline
hard local deadline
hard local deadline
Step-Down (urgent case)adjust hard local deadline backward
Double-Slope (urgent case)adjust hard local deadline backward
Step-Down (elastic case)adjust hard local deadline forward
Double-Slope (elastic case)adjust hard local deadline forward
Our Approach 34
Scheduler
class 1 packets
class 2 packets
class i packets
Step-Down
Slope-Down
Slope-Step
Double-Slope
Single Class (real time)
Multiple Class (real time)
Scheduling policy 1:Step-Down & Step-Slope
Scheduling policy 2:Step-Down & Slope-Down
Scheduling policy 3:Step-Down & Double-Slope
Scheduling policy 12:Double-Slope & Slope-Down
Packet order:maximize overall QoS
整體架構 排程演算法
SPQ-SA
MQ-SA
PMQ-SA
排程環境獨立排程前瞻式排程
Our Approach 35
Scheduling policy
四種 profit function 兩兩組合,共有六種組合,再賦予 conversational class 及 streaming class ,共有 12 種 Scheduling policy
Policy Conversational Streaming Policy Conversational Streaming
1 Step-Down Step-Slope 7 Step-Slope Step-Down
2 Step-Down Slope-Down 8 Slope-Down Step-Down
3 Step-Down Double-Slope 9 Double-Slope Step-Down
4 Step-Slope Slope-Down 10 Slope-Down Step-Slope
5 Step-Slope Double-Slope 11 Double-Slope Step-Slope
6 Slope-Down Double-Slope 12 Double-Slope Slope-Down
Outline 36
Outline
Introduction Related Work Our Approach Performance Evaluation Conclusion
Performance Evaluation 37
Performance Evaluation實驗設計
實驗步驟1. PMQ v.s. MQ
觀察改善 multiple queue 排程演算法的效能2. LA-PMQ v.s. PMQ
觀察前瞻式排程效能• 找出最佳調整係數 (c)
3. Multiple Class 觀察多等級服務的封包調配不同 profit function 間的行為與不同等
級服務間的差異化處理 研究重點在於如何調配 profit function給不同等級的服務
NJCMSPQ-SA
MQ-SA
Intra-Router Scheduling
PMQ-SAOur Approach PMQ-SAOur Approach
Looking Ahead Scheduling
Multiple Class
Performance Evaluation 38
Performance Evaluation (cont.)
實驗工具 NS-2 (ver 2.30)
Single Class 評估指標 Average jitter Packet loss rate Average delay time
Multiple Class 評估指標 Total charge
Performance Evaluation 39
實驗拓墣
Router Router
ReceiverUDP
Sender
UDPSender
UDPSender
UDPSender
TCPSender
RouterRouter
10MB/20ms
10MB/20m
s
10M
B/2
0ms
10MB/
20m
s
10M
B/2
0ms
5MB/20ms 5MB/20ms 5MB/20ms
Cross Traffic Cross Traffic Cross Traffic Cross Traffic
虛線部分代表實驗控制變數
Performance Evaluation 40
實驗 1 : PMQ v.s. MQ (1/4) 實驗目標
觀察改善 multiple queue 排程演算法的效能 控制變數
Traffic Load (1A) Hop Counts (1B) Queue Size (1C)
對照組 FIFO (DropTail) NJCM 中的 MQ、 SPQ
Performance Evaluation 41
實驗 1 : PMQ v.s. MQ (2/4)Result 1A :變動 Traffic Load 的影響
PMQ v.s. MQ (Step-Down)Average Jitter at Different Number of CBR flows
0
2
4
6
3 6 9 12 15 18
Number of CBR flows
Ave
rage
Jitt
er (m
s)
FIFO
SPQ
MQ
PMQ
PMQ v.s. MQ (Step-Down)Loss Rate at Different Number of CBR flows
020406080
3 6 9 12 15 18
Number of CBR flows
Los
s Rat
e (%
)
FIFO
SPQ
MQ
PMQ
PMQ v.s. MQ (Step-Down)Average Delay Time at Different Number of CBR flows
100
150
200
250
3 6 9 12 15 18
Number of CBR flows
Ave
rage
Del
ay (m
s)
FIFO
SPQ
MQ
PMQ
Simulation 1A
Profit Function Step-Down
Number of CBR flows 3,6,9,12,15,18
Hop Counts 4
Queue Size 20 packets
Performance Evaluation 42
實驗 1 : PMQ v.s. MQ (3/4)Result 1B :變動 Hop Counts 的影響
PMQ v.s. MQ (Step-Down)Average Jitter at Different Hop Counts
0
2
4
6
3 4 5 6 7 8
Hop Counts
Ave
rage
Jitt
er (m
s)
FIFO
SPQ
MQ
PMQ
PMQ v.s. MQ (Step-Down)Loss Rate at Different Hop Counts
15
35
55
75
3 4 5 6 7 8
Hop Counts
Los
s Rat
e (%
)
FIFO
SPQ
MQ
PMQ
PMQ v.s. MQ (Step-Down)Average Delay Time at Different Hop Counts
90190290390490
3 4 5 6 7 8
Hop Counts
Ave
rage
Del
ay (m
s)
FIFO
SPQ
MQ
PMQ
(加入一個 priority queue 的確可以使效能更逼近 SPQ)
Simulation 1B
Profit Function Step-Down
Number of CBR flows 12
Hop Counts 3,4,5,6,7,8
Queue Size 20 packets
Performance Evaluation 43
實驗 1 : PMQ v.s. MQ (4/4)Result 1C :變動 Queue Size 的影響
PMQ v.s. MQ (Step-Down)Average Jitter at Different Queue Size
0
2
4
6
15 20 25 30 35 40
Queue Size
Ave
rage
Jitt
er (m
s)
FIFO
SPQ
MQ
PMQ
PMQ v.s. MQ (Step-Down)Loss Rate at Different Queue Size
020406080
15 20 25 30 35 40
Queue Size
Los
s Rat
e (%
)
FIFO
SPQ
MQ
PMQ
PMQ v.s. MQ (Step-Down)Average Delay Time at Different Queue Size
100150200250300350
15 20 25 30 35 40
Queue Size
Ave
rage
Del
ay (m
s)
FIFO
SPQ
MQ
PMQ
(雖然可以改善 MQ ,但由於 multiple queue 的特性,queue size越大 loss rate越大 )
Simulation 1C
Profit Function Step-Down
Number of CBR flows 12
Hop Counts 4
Queue Size 15,20,25,30,35,40
Performance Evaluation 44
實驗 2 : LA-PMQ v.s. PMQ (1/6) 實驗目標
觀察前瞻式排程動態調整 deadline 的效能 找出最佳調整參數 (c)
控制變數 Traffic Load (2A) Hop Counts (2B) Queue Size (2C) Another Topology (2D)
對照組 獨立排程
Performance Evaluation 45
實驗 2 : LA-PMQ v.s. PMQ (2/6)Result 2A :變動 Traffic Load 的影響
LA-PMQ v.s. PMQ (Step-Down)Average Jitter at Different Number of CBR flows
0
1
2
3
3 6 9 12 15 18
Number of CBR flows
Ave
rage
Jitt
er (m
s)
PMQ
LA-PMQ c=0.1
LA-PMQ c=0.2
LA-PMQ v.s. PMQ (Step-Down)Loss Rate at Different Number of CBR flows
020406080
3 6 9 12 15 18
Number of CBR flows
Los
s Rat
e (%
)
PMQ
LA-PMQ c=0.1
LA-PMQ c=0.2
LA-PMQ v.s. PMQ (Step-Down)Average Delay Time at Different Number of CBR flows
108110112114116
3 6 9 12 15 18
Number of CBR flows
Ave
rage
Del
ay (m
s)
PMQ
LA-PMQ c=0.1
LA-PMQ c=0.2
Simulation 2A (Step-Down)
Profit Function Step-Down
Number of CBR flows 3,6,9,12,15,18
Hop Counts 4
Queue Size 20
Performance Evaluation 46
實驗 2 : LA-PMQ v.s. PMQ (3/6)Result 2B :變動 Hop Count 的影響
LA-PMQ v.s. PMQ (Step-Down)Average Jitter at Different Hop Counts
1
1.5
2
2.5
3 4 5 6 7 8
Hop Counts
Ave
rage
Jitt
er (m
s)
PMQ
LA-PMQ c=0.1
LA-PMQ c=0.2
LA-PMQ v.s. PMQ (Step-Down)Loss Rate at Different Hop Counts
1020304050
3 4 5 6 7 8
Hop Counts
Los
s Rat
e (%
)
PMQ
LA-PMQ c=0.1
LA-PMQ c=0.2
LA-PMQ v.s. PMQ (Step-Down)Average Delay Time at Different Hop Counts
50100150200250
3 4 5 6 7 8
Hop Counts
Ave
rage
Del
ay (m
s)
PMQ
LA-PMQ c=0.1
LA-PMQ c=0.2
( 根據 router 後續負載動態調整 local deadline 的確可以改善效能,尤其是 loss rate部分 )
Simulation 2B (Step-Down)
Profit Function Step-Down
Number of CBR flows 12
Hop Counts 3,4,5,6,7,8
Queue Size 20 packets
Performance Evaluation 47
實驗 2 : LA-PMQ v.s. PMQ (4/6)Result 2C :變動 Queue Size 的影響
LA-PMQ v.s. PMQ (Step-Down)Average Jitter at Different Queue Size
0
1
2
3
15 20 25 30 35 40
Queue Size
Ave
rage
Jitt
er (m
s)
PMQ
LA-PMQ c=0.1
LA-PMQ c=0.2
LA-PMQ v.s. PMQ (Step-Down)Loss Rate at Different Queue Size
0
20
40
60
15 20 25 30 35 40
Queue Size
Los
s Rat
e (%
)
PMQ
LA-PMQ c=0.1
LA-PMQ c=0.2
LA-PMQ v.s. PMQ (Step-Down)Average Delay Time at Different Queue Size
100105110115120
15 20 25 30 35 40
Queue Size
Ave
rage
Del
ay (m
s)
PMQ
LA-PMQ c=0.1
LA-PMQ c=0.2
Simulation 2C (Step-Down)
Profit Function Step-Down
Number of CBR flows 12
Hop Counts 4
Queue Size 15,20,25,30,35,40
(綜合實驗 2 ,整體觀察後我們選擇 0.1 當作我們的最佳調整係數,並進行後續實驗 )
Performance Evaluation 48
實驗 2 : LA-PMQ v.s. PMQ (5/6)Topology 2D :魚骨拓墣
Router Router
Receiver
RouterRouter
5MB/20ms
Cross Traffic
Router Router
5MB/20ms
5MB/20ms
5MB/20ms
5MB/20ms
5MB/20ms
Cross Traffic Cross Traffic Cross Traffic Cross Traffic Cross Traffic
UDPSender
UDPSender
UDPSender
UDPSender
10M
B/2
0m
s
10MB/20m
s
10M
B/20
ms
10M
B/2
0m
s
UDPSender
UDPSender
10M
B/20
ms
10M
B/2
0m
s
UDPSender
UDPSender
10M
B/20
ms
10M
B/2
0m
s
UDPSender
UDPSender
10M
B/2
0m
s
10MB/20m
s
UDPSender
UDPSender
10M
B/2
0m
s
10MB/20m
s
Performance Evaluation 49
實驗 2 : LA-PMQ v.s. PMQ (6/6)Result 2D :魚骨拓墣 變動 Queue Size對前瞻式排程效能的影響
LA-PMQ v.s. PMQ (Step-Down)Average Jitter at Different Queue Size
0
0.5
1
1.5
15 20 25 30 35 40
Queue Size
Ave
rage
Jitt
er (m
s)
PMQ
LA-PMQ
LA-PMQ v.s. PMQ (Step-Down)Loss Rate at Different Queue Size
020406080
15 20 25 30 35 40
Queue Size
Los
s Rat
e (%
)
PMQ
LA-PMQ
LA-PMQ v.s. PMQ (Step-Down)Average Delay Time at Different Queue Size
104
114
124
134
15 20 25 30 35 40
Queue Size
Ave
rage
Del
ay (m
s)
PMQ
LA-PMQ
Simulation 2D (Step-Down)
Profit Function Step-Down
Queue Size 15,20,25,30,35,40 packets
(魚骨拓墣中有的接收端較為接近接收端,如此這些封包較不緊急可以稍緩送出以補救其他較為緊急的封包 )
Performance Evaluation 50
實驗 3 :多等級服務 (1/7) 實驗目標
觀察多等級服務的封包調配不同 profit function 間的行為與不同等級服務間的差異化處理
研究重點:我們將藉由實驗找出適合的配方 實驗流程
四種 profit function共存 (3A) 控制變數
Traffic Load 雙 Real-Time 服務等級搭配不同 profit function (3B)
控制變數 VoIP 訊務流數量
Performance Evaluation 51
實驗 3 :多等級服務 (2/7)Result 3A :四種 profit function共存變動 Traffic Load 的影響
Four Profit Function Co-existAverage Jitter at Different Number of CBR flows
0
5
10
15
4 8 12 16 20
Number of CBR flows
Ave
rage
Jitt
er (m
s)
Step-Down
Step-Slope
Slope-Down
Double-Slope
Four Profit Function Co-existLoss Rate at Different Number of CBR flows
020406080
4 8 12 16 20
Number of CBR flows
Los
s Rat
e (%
)
Step-Down
Step-Slope
Slope-Down
Double-Slope
Four Profit Function Co-existAverage Delay Time at Different Number of CBR flows
110120130140150
4 8 12 16 20
Number of CBR flows
Ave
rage
Del
ay (m
s)
Step-Down
Step-Slope
Slope-Down
Double-Slope
Simulation 3A
Profit Function Four profit function co-exist
Number of CBR flows 4,8,12,16,20
Hop Counts 4
Queue Size 20 packets
( 大致上兩兩分為一組, Step-Down與 Step-Slope較差,而 Slope-Down與 Double-Down表現較佳 )
Performance Evaluation 52
實驗 3 :多等級服務 (3/7)3B :雙 Real-Time 服務等級Topology 3B
虛線部分代表實驗控制變數
Router Router
ReceiverUDP
Sender
UDPSender
UDPSender
UDPSender
TCPSender
RouterRouter
3MB/20ms
3MB/20m
s
3M
B/2
0m
s
3MB/
20m
s
3M
B/2
0m
s
3MB/20ms 3MB/20ms 3MB/20ms
Cross Traffic Cross Traffic Cross Traffic Cross Traffic
VoIPData rate:15.2 Kb/s
Packet size:78 Bytes
3MB/20ms
VoDData rate:
0.5 Mb/sPacket size:
1000 Bytes
FTPPacket size:
1000 Bytes
Performance Evaluation 53
實驗 3 :多等級服務 (4/7)3B :雙 Real-Time 服務等級 (cont.) 參數設定
q3=1Bi
(1)=150(ms) (upper threshold delay time satisfaction) Bi
(2)=15(ms) (upper threshold jitter satisfaction)B1
(3)=0.05 (upper threshold of loss rate satisfaction for Conversational class)
Conversational Streaming Background
i 1 2 3
Quality coefficient
ai 0.6 0.8
bi 0.2
ci 0.2 0.2
Unit Price
Ci 0.001 0.0003 0.0001
Number of flows
Ni 15,20,25,30,35,40
1 1
)(3
1 1,,
i
N
jjiiji
j
Cqc
(3)ii
(2)ii
(1)ii
iii
Brate lossor
Bjitter averageor
Bdelay time averagefor 0
cba iii
iq
Performance Evaluation 54
實驗 3 :多等級服務 (5/7)Result 3B :雙 Real-Time 服務等級變動 VoIP數量的影響 (1)
0
500
1000
1500
1 2 3 4 5 6 7 8 9 10 11 12
Total charge of different scheduling policy
15
20
25
30
35
40
Total charge
Scheduling policy
Number ofVoIP flows
Performance Evaluation 55
實驗 3 :多等級服務 (6/7)Result 3B :雙 Real-Time 服務等級變動 VoIP數量的影響 (2)
Policy 1Thoughput at different number of VoIP flows
020406080
15 20 25 30 35 40
Number of VoIP flows
Tho
ughp
ut (K
B/s
)
VoIP
VoD
FTP
Policy 6Thoughput at different number of VoIP flows
020406080
15 20 25 30 35 40
Number of VoIP flows
Tho
ughp
ut (K
B/s
)
VoIP
VoD
FTP
Policy 1 v.s. Policy 6Average Jitter at different number of VoIP flows
0
5
10
15
15 20 25 30 35 40
Number of VoIP flows
Ave
rage
Jitt
er (m
s)
VoIP-1
VoD-1
VoIP-6
VoD-6
Policy 1 v.s. Policy 6Average Delay at at different number of VoIP flows
100
110
120
130
15 20 25 30 35 40
Number of VoIP flows
Ave
rage
Del
ay (m
s)VoIP-1
VoD-1
VoIP-6
VoD-6
Performance Evaluation 56
實驗 3 :多等級服務 (7/7)Conclude 3B :雙 Real-Time 服務等級 實驗結果結果顯示在我們定義的評估指標下,只要是
將 conversational class賦予 Slope-Down 或 Double-Slope 的 policy 整體效能都有不錯的表現
而觀察整體結果,我們推薦 policy 6
slope=S1
slope=S2
profit
timeTc
(current time)Td
(expected trans. time)
Tb(hard deadline)
1
profit=1+(Td-Tb)*S1 when Td≦ Tbprofit=0 when Td>Tb
(conversational class) (streaming class)
slope=S1
slope=S2
profit
timeTc
(current time)Td
(expected trans. time)
Ta(soft deadline)
1
profit=1+(Td-Ta)*S1 when Td≦ Taprofit=1+(Td-Ta)*S2 when Ta-(1/S2)≧ Td>Taprofit=0 when Td>Ta-(1/S2)
Tb(hard deadline)
Outline 57
Outline
Introduction Related Work Our Approach Performance Evaluation Conclusion
Conclusion 58
Conclusion Real-time traffic 的訊務對於 jitter 以及 delay time 有著不同
的要求,但目前網路並未根據封包的服務等級差異性的處理,無法滿足 real-time traffic 的品質要求
我們在每個 router 對封包進行重新排程,並參考封包在後續路程 router 的負載,最後實驗結果顯示我們的方法能有效改善 real-time traffic 的整體效能。並且進一步研究在多服務等級之下對不同重要性的封包做差異化的處理,能有效提高整體滿意度
未來研究方向 其他 routing metric ,甚至其他 routing protocol 設計更多適用於不同網路環境的 profit function ,甚至可以
設計適合非 real-time 服務等級使用的 profit function ,以追求最大的整體滿意度