Outline o o o Introduction Related Work Our

Outline o o o Introduction Related Work Our Approach Performance Evaluation Conclusion Outline 2

3 GPP UMTS Qo. S分類架構 Delay Jitter Packet Loss Sensitivity EX. Conversational High Low Vo. IP Streaming Medium High Low Vo. D Interactive Medium Low High Telnet Background Low No High E-Mail o o 即時性的應用服務，像是Vo. IP以及Vo. D，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 Introduction 4

Outline o o o Introduction Related Work Our Approach Performance Evaluation Conclusion Outline 9

Related Work 文獻一 o A New Jitter Control Mechanism by Per -Hop Packet Scheduling Approach (後文簡稱NJCM) n 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 o 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 o 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 10

hard local deadline soft local deadline Four profit function (Step-Down) Related Work (Slope. Step) pre soft deadline profit rate post soft deadline profit rate (Slope. Down) 11 (Double. Slope)

兩種output queue架構 o single preemptive queue router n o router可將進入的封包插 到output queue的任意 位置 multiple queue router n n n output queue是由數個 FIFO 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 (各有對應的排程演算法：SPQ-SA、MQ-SA) Related Work 12

SPQ-SA (single preemptive queue) o Base case n o 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 n greedy algorithm o 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 13

MQ-SA (multiple queue) o 可視為SPQ的special case n o o 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 n n 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 14

MQ-SA (cont. ) Related Work 15

Related Work 文獻二 o A Priority Based Packet Scheduler with Deadline Considerations n two sets of algorithms introduced o o n 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 Qo. S Related Work 16

文獻二 (cont. ) o SPD n n sorted according to their priorities remaining deadline of the packet is checked o n o discarded if the packet has no remaining deadline left partial sorting DPD n modify the priority by remaining deadlines Related Work 17

Outline o o o Introduction Related Work Our Approach Performance Evaluation Conclusion Outline 20

Design Challenges o Design Challenge of PMQ-SA n o 如何改善multiple queue排程演算法 Design Challenges of Looking ahead packet scheduling n n 後續網路狀態因素 o 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 22

Objective function (1/2) o 整體Qo. S滿意度 (multiple class) n n 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) qi, j：quality index for the j-th flow of class i Ci：unit price per byte of class i λi, 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 24

Objective function (2/2) n 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 i Bi(1) (Bi(2)、Bi(3))： upper threshold of delay time (jitter、loss rate) satisfaction factor Our Approach 25

Design Challenge of PMQ-SA 如何改善multiple queue排程演算法 o NJCM中的multiple queue n o 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 Our Approach 26

如何改善multiple queue排程演算法 (cont. ) o PMQ-SA n n 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 Our Approach 27

Design Challenge 1 of Looking ahead packet scheduling 如何估計後續router負載 o 估計剩餘傳遞時間 n n 利用Routing Table中的資訊 o 反應網路實際狀況及負載 o 進行後續路程封包排程 假設router使用link state routing protocol (e. g. OSPF) o 提供幾項routing metrics計算link cost n n n Our Approach 流量負載 可用頻寬 delay time (RTT/2) 28

Our Approach 33

Scheduling policy o 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 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-Down 10 Slope-Down Step-Slope 5 Step-Slope Double-Slope 11 Double-Slope Step-Slope 6 Slope-Down Double-Slope 12 Double-Slope-Down Our Approach 35

Outline o o o Introduction Related Work Our Approach Performance Evaluation Conclusion Outline 36

Performance Evaluation (cont. ) o 實驗 具 n o Single Class 評估指標 n n n o NS-2 (ver 2. 30) Average jitter Packet loss rate Average delay time Multiple Class 評估指標 n Total charge Performance Evaluation 38

實驗 1：PMQ v. s. MQ (1/4) o 實驗目標 n n 觀察改善multiple queue排程演算法的效能 控制變數 o o o n Traffic Load (1 A) Hop Counts (1 B) Queue Size (1 C) 對照組 o o Performance Evaluation FIFO (Drop. Tail) NJCM中的MQ、SPQ 40

實驗 1：PMQ v. s. MQ (2/4) Result 1 A：變動Traffic Load的影響 Simulation 1 A Performance Evaluation Profit Function Step-Down Number of CBR flows 3, 6, 9, 12, 15, 18 Hop Counts 4 Queue Size 20 packets 41

實驗 1：PMQ v. s. MQ (3/4) Result 1 B：變動Hop Counts的影響 Simulation 1 B Performance Evaluation Profit Function Step-Down Number of CBR flows 12 Hop Counts 3, 4, 5, 6, 7, 8 Queue Size 20 packets (加入一個priority queue的確可以使效能更逼近SPQ) 42

實驗 1：PMQ v. s. MQ (4/4) Result 1 C：變動Queue Size的影響 Simulation 1 C Performance Evaluation Profit Function Step-Down Number of CBR flows 12 Hop Counts 4 Queue Size 15, 20, 25, 30, 35, 40 (雖然可以改善MQ，但由於multiple queue的特性， queue size越大loss rate越大) 43

實驗 2：LA-PMQ v. s. PMQ (1/6) o 實驗目標 n 觀察前瞻式排程動態調整deadline的效能 n n 控制變數 o o n 找出最佳調整參數 (c) Traffic Load (2 A) Hop Counts (2 B) Queue Size (2 C) Another Topology (2 D) 對照組 o Performance Evaluation 獨立排程 44

實驗 2：LA-PMQ v. s. PMQ (2/6) Result 2 A：變動Traffic Load的影響 Simulation 2 A (Step-Down) Performance Evaluation Profit Function Step-Down Number of CBR flows 3, 6, 9, 12, 15, 18 Hop Counts 4 Queue Size 20 45

實驗 2：LA-PMQ v. s. PMQ (3/6) Result 2 B：變動Hop Count的影響 Simulation 2 B (Step-Down) Performance Evaluation Profit Function Step-Down Number of CBR flows 12 Hop Counts 3, 4, 5, 6, 7, 8 Queue Size 20 packets (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) 46

實驗 2：LA-PMQ v. s. PMQ (4/6) Result 2 C：變動Queue Size的影響 Simulation 2 C (Step-Down) Performance Evaluation 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 ) 47

實驗 2：LA-PMQ v. s. PMQ (5/6) Topology 2 D：魚骨拓墣 Performance Evaluation 48

實驗 2：LA-PMQ v. s. PMQ (6/6) Result 2 D：魚骨拓墣 變動Queue Size 對前瞻式排程效能的影響 Simulation 2 D (Step-Down) Performance Evaluation 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 ) 49

實驗 3：多等級服務 (2/7) Result 3 A：四種profit function共存 變動Traffic Load的影響 Simulation 3 A Performance Evaluation Profit Function Four profit function coexist 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，51 而Slope-Down與Double-Down表現較佳 )

實驗 3：多等級服務 (3/7) 3 B：雙Real-Time服務等級 Topology 3 B 虛線部分代表實驗控制變數 Performance Evaluation 52

實驗 3：多等級服務 (4/7) 3 B：雙Real-Time服務等級 (cont. ) o 參數設定 i Conversational Streaming Background 1 2 3 Quality coefficient ai 0. 6 bi 0. 2 ci 0. 2 0. 8 0. 2 Unit Price Ci 0. 001 0. 0003 0. 0001 Number of flows Ni q 3=1 Bi(1)=150(ms) Bi(2)=15(ms) B 1(3)=0. 05 Performance Evaluation 15, 20, 25, 30, 35, 40 1 1 (upper threshold delay time satisfaction) (upper threshold jitter satisfaction) (upper threshold of loss rate satisfaction for Conversational class) 53

實驗 3：多等級服務 (5/7) Result 3 B：雙Real-Time服務等級 變動Vo. IP數量的影響 (1) Performance Evaluation 54

實驗 3：多等級服務 (6/7) Result 3 B：雙Real-Time服務等級 變動Vo. IP數量的影響 (2) Performance Evaluation 55

實驗 3：多等級服務 (7/7) Conclude 3 B：雙Real-Time服務等級 o o 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 (conversational class) Performance Evaluation (streaming class) 56

Outline o o o Introduction Related Work Our Approach Performance Evaluation Conclusion Outline 57