多天線多通道多模多速率無線網狀網路之設計與實作-子計畫一：M4無線網狀網路之網路規劃及資源分配問題(I)

(1)

行政院國家科學委員會專題研究計畫期中進度報告

子計畫一：M4 無線網狀網路之網路規劃及資源分配問題

(1/3)

計畫類別：整合型計畫計畫編號： NSC94-2219-E-009-004- 執行期間： 94 年 08 月 01 日至 95 年 07 月 31 日執行單位：國立交通大學資訊工程學系(所) 計畫主持人：曾煜棋共同主持人：吳真貞計畫參與人員：王友群、林致宇、陳淑敏、談偉航、陳烈武、吳佩曄報告類型：完整報告報告附件：出席國際會議研究心得報告及發表論文處理方式：本計畫可公開查詢

中華民國 95 年 5 月 27 日

(2)

行政院國家科學委員會補助專題研究計畫

□ 成果報告

5 期中進度報告

多天線多通道多模多速率無線網狀網路之設計與實作-子計

畫一：M4 無線網狀網路之網路規劃及資源分配問題(1/3)

計畫類別：□ 個別型計畫

5 整合型計畫

計畫編號：NSC 94 － 2219 － E － 009 － 004 －

執行期間： 94 年 8 月 1 日至 95 年 7 月 31 日

計畫主持人：曾煜棋

共同主持人：吳真貞

計畫參與人員：王友群、林致宇、陳淑敏、談偉航、陳烈武、吳佩曄

成果報告類型(依經費核定清單規定繳交)：□精簡報告 5完整報告

本成果報告包括以下應繳交之附件：

□赴國外出差或研習心得報告一份

□赴大陸地區出差或研習心得報告一份

□出席國際學術會議心得報告及發表之論文各一份

□國際合作研究計畫國外研究報告書一份

處理方式：除產學合作研究計畫、提升產業技術及人才培育研究計畫、

列管計畫及下列情形者外，得立即公開查詢

□涉及專利或其他智慧財產權，□一年□二年後可公開查詢

執行單位：國立交通大學資訊工程學系

中華民國 95 年 5 月 25 日

(3)

中英文摘要

無線網狀網路的相關研究與應用在近幾年引起相當大的討論，此外 IEEE

802.11 WLAN 也已被廣範的接受了。在本計畫中我們將針對 IEEE 802.11-based

的無線網狀網路設計一些實用的通訊協定。為了提高無線網狀網路的效能，我

們將考量一個使用多重頻道、多重天線、多重傳輸速率的無線網狀網路，在第

一年中我們主要在設計適用於此種網路的媒介存取層協定，設計的重點在於頻

道的分配與媒介存取的機制。我們也已在今年實作一個適用於多頻道無線網狀

網路的鏈結層協定。

關鍵字：媒介存取協定、多重頻道、頻道分配、IEEE 802.11、無線網狀網路

IEEE 802.11 wireless local-area networks (WLANs) have been widely accepted

recently; besides, wireless mesh networks (WMNs) also have received a lot of

attention. In this project, we intend to design some practical communication

protocols for IEEE 802.11-based wireless mesh networks. In order to enhance the

performance of wireless mesh networks, when we design our protocols, the

advantages of using multi-channel, multi-antenna, and multi-rate should be taken

into consideration. In the first year, we focus on the design of medium access

control protocols for wireless mesh networks. The main issues are the channel

assignment and the access mechanism. In this year, we have implemented a

multi-channel link-layer protocol for multi-channel wireless mesh networks.

Keywords: Medium Access Control, Multi-Channel, Channel Assignment, IEEE

802.11, Wireless Mesh Network

(4)

一、前言

無線網狀網路(Wireless Mesh Network) [1] 就像一個由一群路由器(routers)組

成的網路，如圖一所示，路由器之間以無線網路的方式來連線，最近這幾年無

線網狀網路也獲得相當大的重視，無線網狀網路可提供無線寬頻服務的網路，

它融合了無線區域網路(Wireless LAN)和隨建即連(Ad Hoc)網路的優勢，無線

網狀網路採用無線傳輸的方式連結了許多的存取點(Access Point)，有了無線網

狀網路的架設可讓網路服務提供者透過少量的有線網路達成大範圍的服務區

域，且因為少了使用有線網路時纜線的建置過程，時間與成本都可大量的減少。

圖一、無線網狀網路示意圖

在本計畫中我們將針對 IEEE 802.11-based 的無線網狀網路設計一些實用的通

訊協定。為了提高無線網狀網路的效能，我們將考量一個使用多重頻道

(multi-channel)、多重天線(multi-antenna)、多重傳輸速率(multi-rate)的無線網狀

網路，在第一年中我們主要在設計適用於此種網路的媒介存取層協定，設計的

(6)

重點在於頻道的分配與媒介存取的機制。我們也已在今年實作一個適用於多頻

道無線網狀網路的鏈結層協定。

二、研究目的

在 IEEE 802.11[2]網路中有存在著數個不與其他頻道重疊(non-overlap)的頻

道，雖然在 IEEE 802.11 Infrastructure mode 中，在 IEEE 802.11 Ad-hoc mode

中如何有效的利用複數頻道增進網路效能仍是一個值得研究的領域，例如

IEEE 802.11 的媒介存取協定(Medium Access Control Protocol)的設計只針對單

一頻道的使用，因此造成效能無法進一步地提升。當複數個頻道被利用時，可

預期的是網路的吞吐量(Throughput)會增加，干擾可降低，空間中頻道的再使

用率(Spatial Reuse)能提高。

我們最終的目的是要能夠利用複數頻來提升無線網狀網路的吞吐量。由於在無

線網狀網路中，有一些特性是和傳統無線區域網路或無線隨建即連網路不同

的，例如存取點是不具移動能力的，此外在無線網狀網路中，通常會有一個連

接至有線網路的存取點(Gateway)，因此在網路協定的設計上會有別於傳統的

無線區域網路或隨建即連網路。例如由於網狀網路是以效能的提升(Throughput)

為考量，而非以省電或硬體成本為考量，因此通訊協定上的設計可以較為複

雜，而多頻道的使用也更為合理，因此針對無線網狀網路的環境，在頻道的切

換與管理方面我們預期提出一個可在鏈結層(Link Layer)實作的解決方法，以

便提供整個無線網狀網路最大的效能。

(7)

三、文獻探討

下面我們整理了相關文獻，我們針對使用複數頻道(multi-Channel)於無線網路

上的相關學術研究發整列表如下，並簡單地分析它們的優缺點。

協定 [3] [4] [5] [6][7] [8] [9]

頻道數

有限

無限

有限

MAC 協定

需新

MAC

802.11 相容

需新

MAC

需新

MAC

網路卡數

兩張

均可

單張

多張

單張

多張

時間同步

不需要

需要

不需要

演算法

分散式

集中式

分散式

應用網路 Ad-Hoc Mesh Mesh Mesh Ad

Hoc Ad-Hoc

表一、使用複數頻道的通訊協定比較

上表列出了相關複數頻道應用在無線網路的範疇，在[8]中假設可用頻道數

無限多，此假設在現實環境中並不合理，[3]為我們於多年前針對隨意網路所

提出的媒介存取協定，在[3]中需要設計一個新的媒介存取層協定(Medium

Access Protocol, MAC)，如此就無法使用已經被廣範使用的 IEEE 802.11 網路

介面卡，在[6][7]中網路卡數必須多於一張才能實行，[5]為單張網路卡，利用

不停的切換頻道來提高空間的再使用率，但必需時間同步，[4]中的演算法是

集中式演算法，必須知道整個網路的拓撲，[9]把複數頻道建立在繞徑(routing)

上，應用的網路類型屬於隨建即連(Ad-Hoc)式網路。

此外在目前的計畫成果中，我們有一篇期刊論文被接受並且有一篇相關的碩士

論文的發表，更多的文獻探討，可從這兩篇成果發表中發現，此兩篇論文將分

(8)

別在附錄一與附錄二中列出。

比較於之前的相關研究所提出的方法的優缺點，我們將鎖定不改變 IEEE

802.11 協定的條件下，討論於無線網狀網路上使用多重頻道(multi-channel)的

通訊協定，在計畫執行方面，我們將重心放在頻道的分配與一個鏈結層協定的

實作這兩個議題上，研究方法將在下一節做更詳盡的描述。

四、研究方法

在本計畫中我們針對多重頻道的使用，討論兩個相關的議題：（一）頻道分配

(Channel Assignment)問題，

（二）IEEE 802.11 相容的媒介存取機制。我們並且

在今年實作出了鏈結層(Link-layer)的通訊協定，實作的細節或在下面的報告以

及附錄中說明。

以格子狀為基礎(Grid-based)的頻道分配法

在第一個議題方面，我們觀察到頻道分配的原則是要盡量的提高空間中頻道的

再使用率(Spatial Reuse)，因此可知頻道的使用和空間是有相互的關係，基於

上面的觀察，在我們提出來的將空間切割成格子狀(Grid) ，一個 Grid 都會被

分配到一個頻道，如圖二所示：

(9)

圖二、Grid-based 頻道分配法，圖 a 的總頻道數為 9，圖 b 的總頻道數為 14

藉由事先就將空間上所應該使用頻道分配好，我們可以盡可能地確保空間中頻

道的再使用率(Spatial Reuse)，而這個方法有一個很重要的議題就是 Grid 的大

小問題，我們認為 Grid 的大小會跟節點的傳輸範圍 (Transmission Range) 有

關，我們假設一個 Grid 的大小為 (d×d) ，節點的傳輸範圍為 r。我們針對不

同的 r/d ratio 測量其效能。實驗的結果可於附錄一中查詢到。

在解決完 Channel Assignment 的問題之後，我們也提出了一個相對應的媒介存

取層協定(MAC Protocol)，此協定有點類似於我們之前所提出的 Multi-Channel

Mac Protocol[3] ，然而在頻道的選擇方面，我們採用了直接在空間上分配好頻

道使用的方法，以提高空間中頻道的再使用率(Spatial Reuse) 。此方法的細結

(10)

相容於 IEEE 802.11 的鏈結層媒介存取機制

由於我們之前所提出適用於多頻道環境的通訊協定（包含前面所提的

Grid-based Channel Assignment）

，大多需要修改到媒介存取層協定，也就是網

路介面可能需要重新的設計，這和已經相當普及的 IEEE 802.11 WLAN 是有所

抵觸的（因為使用者無法在使用原本的網路卡，而必須再另外購買新的網路介

面卡），因此我們也試著提出一個能相容於 IEEE 802.11 的鏈結層媒介存取機

制，並試著將此協定透過只需要更改網路卡驅動程式的方式於 Linux 平台上實

作此機制。下面將先簡單地敘述我們所設計的通訊協定，之後會說明設計此協

定時，哪些議題是要考良的，最後我們會敘述一下我們實作的方法。

我們發展了一套適用於無線網狀網路(Wireless Mesh Network)，而可實作在鏈

結層(Link layer)上使用複數頻道的頻道管理協定(Channel Management

Protocol)，這是一個以接收端的頻道(Receiver-based)做為傳輸頻道的方法，主

要想是是假設當有一個存取點(Access Point)A 要傳送資料給另一個存取點 B

時，A 就要切換到 B 所使用的頻道上進行通訊。我們假設每個存取點都會分

配到一個接收頻道(receiving channel)，此頻道應該與存取點的鄰居(Neighbors)

所使用的接收頻道(receiving channel)要盡量的不同。如圖三，存取點

A,B,C,D,E,F 都盡量的使用不同的頻道做為接收頻道(receiving channel)，例如 A

使用 Channel 3，B 使用 Channel 5，C 使用 Channel 4 等，假如 B 要傳輸資料

給 A，會用 Channel 3 去傳輸資料，同時 C 要傳資料給 D，會用 Channel 1 去

(11)

擾使網路吞吐量(Throughput)增加。

圖三、接收頻道示意圖

接收端為主(receiver based)的設計會產生一個問題，如果存取點 A 要傳送資料

給存取點 B，A 會切換到 B 的接收頻道(receiving channel)上，但假設此時 B 也

正要傳送資料給 C，則 B 會切換到 C 的接收頻道(receiving channel)上，如此有

可能形成死結(deadlock)的情形，為此我們加進“分時”的概念，我們把每個

存取點的時間軸切成一個一個的時槽(time slot)，我們假設每個存取點的時槽

的開始是同步的，並設定每 k 個時槽為一個週期(cycle)，然後重覆這 k 個時槽。

在這 k 個時槽裡，每個存取點（假設為 A）需指定好那些時槽是用來傳送資料

給其它存取點，那些時槽是用來接收其它存取點所送過來的資料，然後把這個

資訊廣播給其鄰居（假設為 B）

，當 B 收到這個資訊時，B 就能知道 A 的時槽

使用情況，如此 B 有資料要傳送給 A 時，B 能夠知道 A 何時可接收資料，何

時不能接收資料，當然 B 會選擇 A 可接收資料的時槽傳送資料給 A，除了傳

送時槽、接收時候和廣播時槽外，我們還選了一個接收時槽當固定式接收時槽

(12)

(Fixed Receiving Slot)，因為我們可以動態改變排程，所以接收時槽可能轉變為

傳送時槽，為了不讓所有的時槽都變成傳送時槽，固定式的接收時槽是不能變

成傳送時槽的，至於固定式接收時槽的選取，也是在鄰居區內要盡量不同。

圖四、Channel Model

如圖四：在這個例子中 k 值為 7，當存取點 B 要傳送資料給 A，因為 B 有收到

A 的廣播說 A 的時槽 3,4,5 是用來接收資料的時槽，因此 B 要送資料給 A 時，

B 會利用時槽 3,4,5 將頻道切換到 A 的接收頻道(receiving channel)，在此例中

為 Channel 3，來進行資料的傳輸。

另外要解決的問題是廣播(Broadcast)的問題，在複數頻道(Multi-channel)的網路

環境下，因為每個存取點可能正使用不同的頻道，如何做有效率的廣播就是一

個問題，相關研究中廣播的方法大多是複製多個廣播封包(Broadcast Packet)在

每個頻道都廣播出去，以便確保鄰居都能接收到此廣播封包，我們所採取的做

法是選擇第一個時槽裡當成廣播時槽(broadcast slot)，在這個廣播時槽裡，頻

(13)

道會切換到一個大家共同的頻道，其目的就是要把所有的存取點在這個時候同

時切換至此共同頻道上，如此一來，所有的存取點便能同時接收或傳送廣播封

包，因為我們並沒有改變 IEEE 802.11 的 MAC 協定，所以這時的接收和傳送

是經由 IEEE 802.11 的競爭機制在傳送。這樣做的好處在於每一次的廣播只需

廣播一次便所有的存取點都接收的到，並不需要在每個頻道上做廣播的動作，

也不需要複製多個廣播封包。有人會問：

「這樣不是會造成頻道擁塞(channel

congestion)？因為這個時候所有的存取點都切換到這個頻道上，造成封包過多

超過這個頻道所能負荷的量。」我們想這是可以避免的，因為我們定義的一個

週期的時槽數可經由廣播封包和一般封包的比例來設定，廣播時槽可以在一個

週期不一定只有一個，可以有兩個或三個，可視這個網路的特性去調整這個參

數。廣播時槽帶來的好處還不只這些，在複數頻道上同步是有困難的，因為所

有的人不在相同的頻道上，有了這個廣播時槽，順使可以在這個時槽發送信號

彈(beacon)來達成時間同步的效果。

我們所提出的頻道管理協定，會衍生出一些待解決的議題如下：第一，如何決

定每個存取點的接收頻道(receiving channel)？第二，如何分配每個存取點傳送

時槽(sending slot)和接收時槽(receiving slot)的比例？第三，如何分配傳送時槽

和接收時槽的順序？第四，進入某個傳送時槽時，要選擇傳送給那一個鄰居才

不會造成不公平？針對這些議題我們分別設計了一些簡單的演算法去解決，詳

細的演算法可在附錄二中查閱。而我們將演算法設計得較為簡單的目的是為了

可實作的考量，我們去修改網路卡的驅動程式以便將我們所設計的通訊協定實

(14)

為了在真的環境上面實作，我們去找尋有公開原始碼的驅動程式(Open Source

Driver)，Atheros 有公開晶片 Linux 驅動程式的原始碼，使用 Atheros 晶片的網

卡都可以使用這個驅動程式來驅動。所以我們選擇了幾台筆記型電腦，每台上

面均裝有 D-link DWL-AG650 的網路卡，它是使用 Atheros 的晶片，這讓我們

可以修改公開的原始碼來把我們的管理協定實作在上面。

圖五、實測環境

圖五為我們實作的環境，我們利用這些筆記型電腦當成是一個個的存取點，讓

它們形成隨建即連網路(Ad-Hoc Network)，並固定其位置模擬無線網狀網路

(Mesh Network)，無線網狀網路和隨建即連網路有很大的共通性，差別在於無

線網狀網路有閘道可以連上網際網路(Internet)，且無線網狀網路沒有行動性

(Mobility)，我們讓這些筆記型電腦模擬一個無線網狀網路的雛形(Prototype)來

當成我們要的環境。

(15)

圖六、測試監控程式介面

圖六為我們的介面，為了便於展示我們的系統，我們開發了此一介面，利用此

介面我們可以觀看每台筆記型電腦使用頻道的狀況、收送封包的狀況等等，同

時我們也可利用此介面去做送封包的動作，次外我們也做了一些簡單的實驗來

證明使用複數頻道的確可以帶來好處，詳細的實驗方法與數據可參閱附錄二。

五、結果與討論

無線網狀網路

融合了無線區域網路(Wireless LAN)和隨建即連(Ad Hoc)網路的

優勢，同時低佈建成本也可進一步地促進無線網路的普及，然而其必須能提供

足夠的頻寬以滿足使用者的需求，在此計畫中我們利用多重頻道

(multi-channel)、多重天線(multi-antenna)等的方式設法提高無線網狀網路的效

能。

在第一年中我們著重在設計適用於多重頻道環境的通訊協定

，首先我們設計了

一個 Grid-based 的 Channel Assignment 及其對應的媒介存取層協定，藉由實驗

(16)

我們發現適當地在空間上做切割之後再分配頻道的方法的確可提高空間中頻

道的再使用率(Spatial Reuse) ，如此可讓使用多重頻道的好處更進一步地展現

出來。然而此方法需要修改原有的網路卡設計，並不太符合現實面上的考量，

因此我們在今年發展了一個相容於 IEEE 802.11 的媒介存取機制，並在鏈結層

上實作出來，最後我們也以實作的平台驗證使用多重頻道的確可提升網路的效

益。

本計畫目前的研究成果為兩篇論文如下：

附錄一：Y.-C. Tseng, S.-L. Wu, C.-M. Chao, and J.-P. Sheu, "An Efficient MAC

Protocol for Multi-Channel Mobile Ad Hoc Networks Based on Location

Information”, Int’l Journal Communication Systems (to appear)

附錄二：陳威碩，“IEEE 802.11 無線網狀網路之分散式時槽分割式多重頻道

協定”，碩士論文(指導教授：曾煜棋教授)，民國九十五年六月

未來工作計畫

然而要提升無線網狀網路的效能，不能只單靠鏈結層或媒介存取層的改進就能

獲得大幅的改善，其它如繞路協定(Routing Protocol)，用來確保服務品質的允

入控制機制等都有改善的空間，在繞路協定方面，在此計劃的第二年我們將試

著設計一個適用於多重天線(multi-antenna)、多重頻道(multi-channel)的繞路協

定，並且在繞路的選擇上去考量多重傳輸速率(multi-rate)的影響，此外我們也

將思考一些無線網狀網路上資源分配的問題。而我們最終的目的則是要實作出

一個無線網狀網路的 Prototype。

六、參考文獻

[1] I. F. Akyildiz, X. Wang, and W. Wang. Wireless mesh networks: a survey. Comput. Netw. ISDN Syst., 47(4):445-487, 2005.

[2] IEEE Std 802.11b-1999. Supplement To IEEE Standard For Information Technology- Telecommunications And Information Exchange Between Systems- Local And Metropol- itan Area Networks- Speci¯c Requirements- Part 11: Wireless LAN Medium Access Control (MAC) And Physical Layer (PHY) Speci¯cations: Higher-speed Physical Layer Extension In The 2.4 GHz Band.

[3] S.-L. Wu, C.-Y. Lin, Y.-C. Tseng, and J.-P. Sheu. A new multi-channel MAC protocol with on-demand channel assignment for multi-hop mobile ad hoc networks. In Proc.

(17)

ISPAN, Dec. 2000.

[4] A. Raniwala, K. Gopalan, and T.-C. Chiueh. Centralized Channel Assignment and Routing Algorithms for Multi-Channel Wireless Mesh Networks. In Mobile Computing

and Communications Review, pages 50-65, Apr. 2004.

[5] P. Bahl, R. Chandra, and J. Dunagan. SSCH: Slotted seeded channel hopping for

capacity improvement in IEEE 802.11 ad-hoc wireless networks. In Proc. MobiCom, Sept. 2004. [6] P. Kyasanur and N. H. Vaidya. Routing and interface assignment in multi-channel

multi-interface wireless networks. In Proc. WCNC, New Orleans, U.S.A., Mar. 2005.

[7] Pradeep Kyasanur and Nitin H. Vaidya, "Routing in Multi-Channel Multi-Interface Ad-Hoc Wireless Networks", Technical Report, December 2004

[8] Sheng-Hsuan Hsu, Ching-Chi Hsu, Shun-Shii Lin, and Ferng-Ching Lin, “A Multi-Channel Mac Protocol Using Maximal Matching for Ad Hoc Networks” in ICDCSW, 2004

[9] M. X. Gong and S. F. Midki®. Distributed channel assignment protocols: A cross-layer approach. In Proc. WCNC, New Orleans, U.S.A., Mar. 2005.

(18)

附錄一：

An Efficient MAC Protocol for

Multi-Channel Mobile Ad Hoc

Networks Based on Location

Information

Y.-C. Tseng, S.-L. Wu, C.-M. Chao, and J.-P. Sheu, "An

Efficient MAC Protocol for Multi-Channel Mobile Ad

Hoc Networks Based on Location Information”, Int’l

(19)

An Eﬃcient MAC protocol for

Multi-Channel Mobile Ad Hoc Networks

Based on Location Information

∗

Yu-Chee Tseng

1

, Shih-Lin Wu

2

, Chih-Min Chao

3

, and Jang-Ping Sheu

4

1

_{Department of Computer Science and Information Engineering}

National Chiao-Tung University, Taiwan

2

_{Department of Electrical Engineering}

Chang Gung University, Taiwan

3

_{Department of Information Management}

Tamkang University, Taiwan

4

_{Department of Computer Science and Information Engineering}

National Central University, Taiwan

Email: [email protected]

Abstract

This paper considers the channel assignment problem in a multi-channel MANET environment. We propose a scheme called GRID, by which a mobile host can easily determine which channel to use based on its current location. In fact, following the GSM style, our GRID spends no communication cost to allo-cate channels to mobile hosts since channel assignment is purely determined by hosts’ physical locations. We show that this can improve the channel reuse ratio. We then propose a multi-channel MAC protocol, which integrates GRID. Our protocol is characterized by the following features: (i) it follows an “on-demand” style to access the medium and thus a mobile host will occupy a channel only when necessary, (ii) the number of channels required is independent of the net-work topology, and (iii) no form of clock synchronization is required. On the other hand, most existing protocols assign channels to a host statically even if it has no intention to transmit [3, 10, 12], require a number of channels which

(20)

is a function of the maximum connectivity [3, 8, 10, 12], or necessitate a clock synchronization among all hosts in the MANET [12, 27]. Through simulations, we demonstrate the advantages of our protocol.

Keywords: channel management, communication protocol, location-aware protocols,

medium access control (MAC), mobile ad hoc network (MANET), mobile computing, wireless communication.

1 Introduction

A mobile ad-hoc network (MANET) is formed by a cluster of mobile hosts without the infrastructure of base stations. Two mobile hosts can communicate with each other indirectly in a multi-hop manner. Since no base station is required, one of its main advantages is that it can be rapidly deployed. The applications of MANETs appear in places where pre-deployment of network infrastructure is difficult or unavailable (e.g., fleets in oceans, armies in march, natural disasters, battle fields, festival field grounds, and historic sites).

A MAC (medium access control) protocol is responsible of resolving the communi-cation contention and collision among hosts. Many MAC protocols have been proposed for wireless networks [4, 7, 13, 15, 20, 21], which assume a common channel shared by mobile hosts. We call such protocols single-channel MAC protocols. The widely ac-cepted standard IEEE 802.11 [1] follows such model. One common problem with such protocols is that the network performance will degrade quickly as the number of mobile hosts increases, due to higher contention/collision.

One approach to relieving the contention/collision problem is to utilize multiple channels. The idea of using separate control and data channels was ﬁrst proposed in [28]. We thus deﬁne a multi-channel MAC protocol as one which allows mobile hosts to dynamically access more than one channel in a MANET environment. Us-ing multiple channels has several advantages. First, while the maximum throughput of a single-channel MAC protocol will be limited by the bandwidth of the channel, the throughput may be increased immediately if a host is allowed to utilize multiple channels. Second, as shown in [2, 25], using multiple channels will experience less

(21)

the normalized propagation delay is deﬁned to be the ratio of the propagation time over the packet transmission time. Therefore, this reduces the probability of collisions. Third, QoS routing may be supported [22].

Here, we use “channel” upon a logical level. Physically, a channel can be a frequency band (under FDMA), or an orthogonal code (under CDMA). How to access multiple channels is thus technology-dependent. Disregard of the transmission technology, we categorize mobile hosts’ channel access capability as follows:

• single-transceiver: A mobile host can only access one channel at a time. The

transceiver can be simplex or duplex. Note that this is not necessarily equivalent to the single-channel model, because the transceiver is still capable of switching from one channel to another.

• multiple-transceiver: Each transceiver could be simplex or duplex. A mobile host

can access multiple channels simultaneously.

In this paper, we propose a new multi-channel MAC protocol for a MANET in which each mobile host is equipped with a positioning device, such as GPS. A multi-channel MAC typically needs to address two issues: channel assignment and medium access. The former is to choose proper channels to send/receive for hosts, while the later is to resolve the contention/collision problem when using a particular channel. These two issues are sometimes addressed separately, but eventually one has to integrate them to provide a total solution. Our channel assignment, called GRID, is characterized by two features: (i) it exploits location information by partitioning the physical area into a number of squares called grids, and (ii) it does not need to transmit any message to assign channels to mobile hosts since channel assignment is purely determined by a host’s physical location. Several channel assignment schemes have been proposed earlier [8, 9, 12, 25, 27], but none of them try to exploit the location information. Our medium access protocol is characterized by the following features: (i) it follows an “on-demand” style to access the medium and thus a mobile host will occupy a channel only when necessary, (ii) the number of channels required is independent of the network topology, and (iii) no form of clock synchronization is required. On the other hand, most existing protocols assign channels to a host statically even if it has no intention to

(22)

transmit [3, 10, 12], require a number of channels which is a function of the maximum connectivity [3, 8, 10, 12], or necessitate a clock synchronization among all hosts in the MANET [12, 27]. A centralized scheme is proposed in a recent work [34]. Similar to hexagonal cellular systems, all channel assignment in a cell is controlled and allocated by the cell leader located at this cell. Since a cellular structure is assumed, location information is needed by each station. Contrary to [34], our GRID scheme is fully distributed and no traﬃc overhead is incurred for channel allocation. A detail review will be given in Section 2.1. For an overview, Table 1 gives a comparison on existing and our protocols.

Since a MANET should operate in a physical area, it is very natural to exploit location information in such an environment. Indeed, location information has been exploited in several issues in MANET (e.g., routing [11, 14, 16, 17, 18, 19, 24, 33], broadcasting [26], and power saving [30]), but not in channel assignment. GSM (Global System for Mobile Communications) is an instance which uses location information to exploit channel reuse, but MANET has quite diﬀerent features — there is no base station, and thus channel assignment has to be done more dynamically in an in-band manner. Since the concept of “channel reuse” is highly related the area where a channel is used, exploiting location information, as we do in this work, on channel assignment could eﬀectively solve this problem.

Outdoor positioning can be solve satisfactorily by GPS (global positioning systems) or DGPS (diﬀerential GPS). Both the price drop of GPS and the recent discontinuation of SA (Selective Availability) motivate us to conduct this research. However, for indoor positioning there is no satisfactory solution at this point.

The rest of this paper is organized as follows. Section 2 discusses some existing channel assignment schemes and our GRID scheme. Section 3 presents our MAC pro-tocol by integrating the GRID assignment. Analysis and simulations are in Section 4. Conclusions will be drawn in Section 5.

(23)

2 Channel Assignment

As mentioned earlier, a multi-channel MAC needs to address two issues: channel as-signment and medium access. In this section, we will consider the channel asas-signment problem. We will ﬁrst review some existing protocols, which are all non-location-aware. Then we will present our location-aware channel assignment.

2.1 Non-Location-Aware Schemes

In this section, we review some channel assignment schemes that do not utilize the location information of mobile hosts. These schemes can be further divided to static and dynamic. The simplest static approach is to assign channels to mobile hosts when the system is ﬁrst set up. For instance, channel i can be statically assigned to those hosts with ID’s such that i = ID mod n (supposing that we number channels as 0, 1, . . . , n − 1).

A scheme based on Latin square is proposed in [12], which assumes a TDMA-over-FDMA technology. Each channel is divided into ﬁxed-length frames. Each host is statically assigned to a time slot in each frame belonging to a frequency band. Since TDMA is used, clock synchronization among all hosts is necessary. Furthermore, each host has to be equipped with a number of transceivers equal to the number of frequency bands, so this approach is quite costly. Also, this scheme needs to know in advance the maximum number of mobile hosts as well as the maximum degree of the topology formed by the MANET.

The schemes in [3, 5, 6, 10, 23] are for channel assignment in the traditional packet radio network. Partial or even complete network topology has to be collected to perform channel assignment. These approaches can basically be classiﬁed as static, although some can handle dynamic failure of base stations. Since these schemes are not designed for MANET, which is typically characterized by high host mobility, they do not ﬁt our need.

A protocol based on dynamic channel assignment is in [8]. It is assumed that the channel assigned to a host must be diﬀerent from those of its two-hop neighbors. To maintain this condition, a large amount of update messages will be sent whenever a

(24)

host determines any change on channel assignment in its two-hop neighbors. This is ineﬃcient in a highly mobile system. Further, this protocol is “degree-dependent” in the sense that it dictates a number of channels equal to an order of the square of the maximum degree of the MANET. So the protocol is inappropriate for a crowded environment.

A “degree-independent” protocol called multichannel-CSMA protocol is proposed in [25]. Suppose that there are n channels. The protocol imposes that each mobile host must have n receivers which concurrently listen on all n channels. Also, there is only one transmitter which will hop from channel to channel and, if necessary, will send on any detected idle channel. Again, this protocol has high hardware cost. Further, since no RTS/CTS is used, the hidden-terminal problem may easily occur. A hop-reservation MAC protocol based on very-slow frequency-hopping spread spectrum is proposed in [27]. Its channel assignment employs RTS/CTS dialogue to reserve a channel. The protocol is also degree-independent but requires clock synchronization among all mobile hosts, which is diﬃcult when the network is dispersed in a large area. Recently, Wu et al. [31] propose a new protocol, called Dynamic Channel

Assign-ment (DCA), which possesses the following characters: (i) it follows an “on-demand”

style to access the medium and thus a mobile host will occupy a channel only when necessary, (ii) the number of channels required is independent of the network topology, and (iii) no form of clock synchronization is required. DCA uses one dedicated channel for control packets, and other channels for data. The purpose of the control channel is to assign data channels to mobile hosts or schedule the use of data channels among hosts’ while data channels are used to transmit data packets and acknowledgements. Reference [32] combines DCA and power control to further improve channel reuse. However, because there is no location information, DCA cannot maintain an eﬃcient channel reuse pattern.

In Table 1, we summarize and compare existing schemes with our yet-to-be-presented GRID scheme.

(25)

Table 1: Comparison of channel assignment schemes (n is the number of hosts, and m is the maximum network degree.

scheme assignment no. channels info. collected loc.-aware assgn. cost transceivers [3, 5, 6, 10, 23] static deg.-dep. global no O(nk), k ≥ 2 1

[12] static deg.-dep. none no 0 m [8] dynamic deg.-dep. 2-hop no O(n3) 2 [25] dynamic deg.-indep. none no 0 m [27] dynamic deg.-indep. none no O(n) 1 ours dynamic deg.-indep. none yes 0 2

2.2 Our Location-Aware Channel Assignment: GRID

Next, we introduce our location-aware channel assignment scheme. The MANET envi-ronment is the same, except that each mobile host must be installed with a positioning device, such as GPS receiver. So our protocol is more appropriate for outdoor envi-ronment. As will be seen later, our approach will assign a channel to a host once the host knows its current location. As a result, in addition to the positioning cost, there is no communication cost for our channel assignment (no message will be sent for this purpose).

We will refer to our scheme as GRID. The MANET is assumed to operate in a pre-deﬁned geographic area. The area is partitioned into 2D logical grids as illustrated in Fig. 1. Each grid is a square of size d × d. Grids are numbered (x, y) following the conventional xy-coordinate. To be location-aware, a mobile host must be able to determine its current grid coordinate. Thus, each mobile host must know how to map a physical location to the corresponding grid coordinate.

Our channel assignment works as follows. We assume that the system is given a ﬁxed number, n, of channels. For each grid, we will assign a channel to it. When a mobile host is located at a grid, say (x, y), it will use the channel assigned to grid (x, y) for transmission. One can easily observe that if we assign the same channel to two neighboring grids, then there will be high chance that the transmission activities on these two neighboring grids will contend, or even interfere, with each other. Thus, we should assign the same channel to grids that are spatially separated by some distance, but will exploit the largest frequency reuse.

(26)

Figure 1: Assigning channels to grids in a band-by-band manner: (a) n = 9 and (b)

n = 14. In each grid, the number on the top is the channel number, while those on the

bottom are the grid coordinate. Here, we number channels from 1 to n.

The above formulation turns out to be similar to the channel arrangement in the GSM system. In the following, we propose a way to assign channels to grids. Let

m = √n . We ﬁrst partition the grids vertically into a number of bands such that

each band contains m columns of grids. Then, for each band, we sequentially assign the n channels to each row of grids, in a row-by-row manner. In Fig. 1, we illustrate this assignment when n = 9 and n = 14. It can readily be seen that when n is a square of some integer, each channel will be regularly separated in the area.

2.2.1 Grid Size vs. Transmission Range

Let r be the transmission range of an antenna. Suppose the value of r is ﬁxed. In this section, we discuss an important design issue: the relationship between r and the side length of grids, d. Below, we discuss several possibilities. For simplicity, let’s assume that m =√n is an integer.

• d r: This means many hosts will stay in a grid and thus contend with each

other on one channel. When d = ∞, this degenerates to the case of one single channel.

(27)

Figure 2: The eﬀect of r/d ratio on channel co-interference when n = 25.

• d > 2r/(m − 1): This is the case that the transmission activities from two hosts

choosing the same channel will never interfere with each other. As illustrated in Fig. 2(a), hosts A and B (both choosing the same channel) are located in the nearest possible locations, but their signals will not overlap in any location.

• d = 2r/m: This is the case that the transmission activities from two hosts which

choose the same channel and which are each located in the center of a grid will not interfere with each other. This is illustrated in Fig. 2(b).

• d = r/m: This represents the minimal value of d such that two hosts (located

at the grid centers) using the same channel will not hear each other. This is illustrated in Fig. 2(c). By simple calculus, we can ﬁnd that each receiver of these two hosts will have a probability of 0.396 being interfered by the signals from the other sender. The value is the ratio of the intersection area that is covered by both hosts A and B to the area that is covered by either host A or host B.

• d ≈ 0: This means that the grid size is inﬁnitely small. This degenerates to the

case that a mobile host will randomly choose a channel to transmit its packets, and thus little channel reuse can be exploited.

The above analysis has indicated some tradeoﬀs. This concept will be captured by the ratio r/d. If the ratio is too large, then the chance of co-channel interference will be high. On the other hand, if the ratio is too small, although co-channel interference can be reduced, the channel reuse will be reduced too since a channel will be unavailable in

(28)

Figure 3: Tests of blocked sender-receiver pairs at diﬀerent r/d ratios: (a) n = 36 and (b) n = 81.

many locations. Thus, we need to carefully adjust the r/d ratio for the best network performance. This will be further investigated through simulations in Section 4.2.

2.2.2 _{Some Experiments on the r/d Ratio}

At this point, it deserves to predict, under ideal situations, how much beneﬁt our location-aware channel assignment can oﬀer over a non-location-aware one. We de-veloped a simple simulation without concerning the details of medium access, such as collision, timing, etc. (this will be explored in Section 4). We simulated an area of size 1000× 1000. On this area, we randomly generated a sender A and then randomly generated a receiver B in the circle of radius r = 100 centered at A. A transmitted using a channel selected by two methods: (i) a static one based on host ID (referred to as SCA, static channel assignment), and (ii) our GRID approach. We then repeated this process to generate more sender-receiver pairs. However, for each pair generated, we tested whether this transmission will interfere any earlier ongoing pairs. If so, the current pair will be deleted; otherwise, it will be granted.

Through this ideal experiment, we intend to observe how many more sender-receiver pairs can be generated in the physical area by GRID than SCA. This will verify whether

(29)

Figure 4: A snapshot of our experiment in Fig. 3 when n = 36 and r/d = 3.0: (a) GRID and (b) SCA. The snapshots are taken on a 1000× 1000 area, and each circle means a sender-receiver pair.

GRID has a better channel reuse. Another important issue we would like to explore here is: what is best ratio r/d to maximize channel reuse?

Fig. 3 shows our ﬁrst experimental results. The x-axis is the number of sender-receiver pairs generated. The y-axis shows the number of pairs that fail and thus are deleted. For our GRID, we tested diﬀerent r/d ratios. Fig. 3(a) uses a total number of

n = 36 channels, and Fig. 3 (b) uses n = 81. Indeed, some r/d ratios are better than

SCA, while some are worse. In Fig. 3(a), we see that the r/d ratios 2.5, 3.0, and 3.5 will outperform SCA, while in Fig. 3(b), the r/d ratios 4.0, 4.5, and 5.0 will outperform SCA.

We conclude from the above experiments that when r/d ≈ √n/2, our GRID will

perform well. The reason is as follows. Let’s consider any channel. At this ratio, it is more likely that we can place most circles (which represent transmission activities of this channel) in a physical area, while incurring the least overlapping among circles (which represents co-channel interference). This is how our GRID can oﬀer better channel reuse. Fig. 4 shows a snapshot in our experiment when n = 36 and r/d = 3.0 on the use of channel 1. Clearly, the placement of circles by GRID is denser and more regular than that of SCA.

(30)

Figure 5: Tests of blocked sender-receiver pairs at various n’s.

In Fig. 5, we further vary the value of n to observe the trend. In this ﬁgure, we have picked the best r/d ratio for each n. The number of sender-receiver pairs generate is 2000. As can be seen, the best ratios are all very close to√_{n/2, as we have predicted.} Also, with more channels, there are less pairs being blocked by both GRID and SCA. But the gain of GRID over SCA will enlarge as a larger n is used.

3 The MAC Protocol

This section presents the medium access part of our protocol by integrating the channel assignment part in the previous section. The channel model is as follows. The overall bandwidth is divided into one control channel and n data channels D1, D2, . . . , Dn.

Each channel, including control and data ones, is of the same bandwidth. The purpose of data channels is to transmit data packets and acknowledgements. Control channel serves in many important management purposes: (i) to synchronize the use of data channels among hosts, (ii) to broadcast beacons periodically, and (iii) to search for routes. Note that beacons can help mobile hosts to discover which hosts are currently neighbors. Hosts can always communicate with others through the control channel, but they can only communicate with each other through data a channel if they switch to the same one. Route discovery and routing functions are beyond the scope of this paper and will not be elaborated, but can be supported by the control channel.

In our protocol, the channel assignment should be done in advanced. We think that the organization, e.g. city governments or corporations, should take the responsibility

(31)

of channel allocation if it wants to use GRID in its district such that the best perfor-mance can be got. It is something like that FCC regulates the use of radio spectrum to satisfy the communications needs without interference.

Each mobile host is equipped with two half-duplex transceivers:

• control transceiver: This transceiver will operate on the control channel to

ex-change control packets with other mobile hosts and to obtain rights to access data channels.

• data transceiver: This transceiver will dynamically operate on one of the data

channels, according to our channel assignment, to transmit data packets and acknowledgements.

Each mobile host X maintains the following data structure.

• CUL[ ]: This is called the channel usage list. Each list entry CUL[i] keeps

records of how and when a host neighboring to X uses a channel. CUL[i] has three ﬁelds:

– CUL[i].host: a neighbor host of X.

– CUL[i].ch: a data channel used by CUL[i].host.

– CUL[i].rel time: when channel CUL[i].ch will be released by CUL[i].host.

Note that this CUL is distributedly maintained by each mobile host and thus may not contain the precise information.

The main idea of our protocol is as follows. For a mobile host A to communicate with host B, A will send a RTS (request-to-send) to B. This RTS will also carry the channel number that A intends to use in its subsequent transmission. Then B will match this request with its in CUL[ ] and, if granted, reply a CTS (clear-to-send) to A. All these will happen on the control channel. Similar to the IEEE 802.11 [1], the purpose of the RTS/CTS dialogue is to warn the neighborhood of A and B not to interfere their subsequent transmission, except that a host is still allowed to use the channels diﬀerent from that indicated in the RTS and CTS packets. Finally, transmission of a data packet will occur on the data channel.

(32)

Table 2: Meanings of variables and constants used in our protocol.

TSIF S length of short inter-frame spacing

TDIF S length of distributed inter-frame spacing

TRT S time to transmit a RTS

TCT S time to transmit a CTS

Tcurr the current clock of a mobile host

TACK time to transmit an ACK

NAVRT S network allocation vector on receiving a RTS

NAVCT S network allocation vector on receiving a CTS

Ld length of a data packet

Lc length of a control packet (RTS/CTS)

Bd bandwidth of a data channel

Bc bandwidth of a control channel

τ maximal propagation delay

The complete protocol is shown below. Table 2 lists the variables/constants used in our presentaiton.

1. On a mobile host A having a data packet to send to host B, it ﬁrst checks whether the following two conditions are true:

a) B is not equal to any CUL[i].host such that

CUL[i].rel time > Tcurr+ (TDIF S+ TRT S+ TSIF S + TCT S).

If so, this means B will still be busy (in using data channel CUL[i].ch) after a successful exchange of RTS and CTS packets.

b) Suppose A determines that its current data channel is DA. Then for each

i = 1..n,

(DA= CUL[i].ch) =⇒ (CUL[i].rel time ≤ Tcurr+(TDIF S+TRT S+TSIF S+TCT S)).

If so, this means A’s data channel is either not currently being used by any of its neighbors, or currently being occupied by some neighbor(s) but will be released after a successful exchange of RTS and CTS packets. (Fig. 6 shows how the above timing is calculated.)

(33)

DA Sender(A) B RTS Receiver(B) RTS CTS CTS S Other (A,B) Communication NAVCTS D NAVRTS0 NAVRTS1 Time B = Backoff D = DIFS S = SIFS Tcurr Trel_time

Figure 6: Timing to determine whether a channel will be free after a successful exchange of RTS and CTS packets.

If the above two conditions are true, proceed to step 2; otherwise, A must wait at step 1 until these conditions become true.

2. Then A can send a RT S(DA, Ld) to B, where Ld is the length of the

yet-to-be-sent data packet. Also, following the IEEE 802.11 style, A can send this RTS only if there is no carrier on the control channel in a TDIF S plus a random backoﬀ

time period. Otherwise, it has to go back to step 1.

3. On a host B receiving the RT S(DA, Ld) from A, it has to check whether the

following condition is true for each i = 1..n:

(DA= CUL[i].ch) =⇒ (CUL[i].rel time ≤ Tcurr+ (TSIF S + TCT S)).

If so, DA is either not currently being used by any of its neighbors, or currently

being used by some neighbor(s) but will be released after a successful transmission of a CTS packet. Then B replies a CT S(DA, NAVCT S) to A, where

NAVCT S = Ld/Bd+ TACK + 2τ.

Then B tunes its data transceiver to DA. Otherwise, B replies a CT S(Test) to

A, where Test is the estimated time that B’s data channel DA will change minus

the time for an exchange of a CTS packet:

Test = max{∀i CUL[i].ch = DA, CUL[i].rel time} − Tcurr− TSIF S − TCT S.

4. On an irrelevant host C = B receiving A’s RT S(DA, Ld), it has to inhibit itself

from using the control channel for a period

(34)

This is to avoid C from interrupting the RTS → CTS dialogue between A and

B. Then, C senses channel DA for a period of τ to determine whether this

communication is success or not. If so, it appends an entry CUL[k] to its CUL such that:

CUL[k].host = A CUL[k].ch = DA

CUL[k].rel time = Tcurr+ NAVRT S1

where

NAVRT S1 = Tcurr+ Ld/Bd+ TACK+ τ.

5. Host A, after sending its RTS, will wait for B’s CTS with a timeout period of

TSIF S+ TCT S+ 2τ . If no CTS is received, A will retry until the maximum number

of retries is reached.

6. On host A receiving B’s CT S(DA, NAVCT S), it performs the following steps:

a) Append an entry CUL[k] to its CUL such that

CUL[k].host = B CUL[k].ch = DA

CUL[k].rel time = Tcurr+ NAVCT S

b) Send its DATA packet to B on the data channel DA.

On the other hand, if A receives B’s CT S(Test), it has to wait for a time period

Test and go back to step 1.

7. On an irrelevant host C = A receiving B’s CT S(DA, NAVCT S), C updates its

CUL. This is the same as step 6a) except that

CUL[k].rel time = Tcurr+ NAVCT S+ τ.

(35)

Figure 7: An example that the control channel is fully loaded and the data channel D4

is not utilized.

8. On B completely receiving A’s data packet, B replies an ACK on DA.

To summarize, our protocol relies on the control channel to negotiate the transmis-sions among hosts using the same data channel. Also, note that although our protocol will send timing information in packets, these are only relative time intervals. No absolute time is sent. So there is no need of clock synchronization in our protocol.

4 Analysis and Simulation Results

4.1 Arrangement of Control and Data Channels

One concern in our protocol is: Can the control channel eﬃciently distribute the com-munication jobs to data channels? For example, in Fig. 7, we show an example with 5 channels, one for control and four for data. For simplicity, let’s assume that the lengths of all control packets (RTS, and CTS) are Lc, and lengths of all data packets

Ld= 6Lc. Then Fig. 7 shows a scenario that the control channel can only utilize three

data channels D1, D2, and D3. Channel D4 may never be used because the control

channel can serve at most three data channels. Although Ld is typically larger than

Lc by an order of at least tens or hundreds, it still deserves to analyze this issue to

understand the limitation.

The above example shows that how to arrange the control and data channels is a critical issue. In the following, we consider two bandwidth models.

• ﬁxed-channel-bandwidth: Each channel (data and control) has a ﬁxed bandwidth.

(36)

• fixed-total-bandwidth: The total bandwidth offered to the network is fixed. Thus,

with more channels, each channel will have less bandwidth.

We comment that the first model may reflect the situation in CDMA, where each code has the same bandwidth, and we may utilize multiple codes to increase the actual bandwidth of the network. On the other hand, the second model may reflect the situation in FDMA, where the total bandwidth is fixed, and our job is to determine an appropriate number of channels to best utilize the given bandwidth.

We will show how to arrange the control and data channels under these models so as to well utilize a given bandwidth. Let’s consider the ﬁxed-channel-bandwidth model ﬁrst. Apparently, since the control channel can arrange a data packet by sending 2 control packets of total length 2Lc, the maximum number of data channels should be

limited by

n ≤ Ld

2× L_c. (1)

Also, consider the utilization U of the total given bandwidth. Since the control channel is actually not used for transmitting data packets, we have

U ≤ n

n + 1. (2)

From Eq. (1) and Eq. (2), we derive that

U 1− U ≤ n ≤ Ld 2× L_c =⇒ U ≤ Ld 2× L_c_{+ L}_d. (3) The above inequality implies that the maximum utilization is a function of the lengths of control and data packets. Thus, decreasing the length of control packets or increasing the length of data packets will improve the utilization. Since the maximum utilization is only dependent of Ld and Lc, it will be unwise to unlimitedly increase the number

of data channels.

Next, we consider the fixed-total-bandwidth model. Suppose that we are given a fixed bandwidth. The problem is: how to assign the bandwidth to the control and data channels to achieve the best utilization. Also, how many data channels (n) will be most efficient? Let the bandwidth of the control channel be Bc, and that of each data

(37)

channel Bd. Again, the number of data channels should be limited by the assignment

capability of the control channel:

n ≤ Ld/Bd

2× L_c_/B_c. (4)

Similarly, the utilization U must satisfy

U ≤ n × Bd n × Bd+ Bc.

(5)

Combining Eq. (4) and Eq. (5) gives

UBc Bd− UBd ≤ n ≤ LdBc 2× L_c_B_d =⇒ U ≤ Ld 2× L_c_{+ L}_d. (6) Interestingly, this gives the same conclusion as that in the ﬁxed-channel-bandwidth model. The bandwidths Bc and Bd have disappeared in the above inequality, and the

maximum utilization is still only a function of the lengths of control and data packets. Thus, decreasing the length of control packets or increasing the length of data packets may improve the utilization. To understand how to arrange the bandwidth, we replace the maximum utilization into Eq. (5), which gives

Ld 2× L_c_{+ L}_d = n × Bd n × Bd+ Bc =⇒ Bc nBd = 2Lc Ld . (7)

Thus, to achieve the best utilization, the ratio of the control bandwidth to the data bandwidth should be 2Lc/Ld. Furthermore, since the maximum utilization is

indepen-dent of the value of n, theoretically once the above ratio (2Lc/Ld) is used, it does not

matter how many data channels that we divide the data bandwidth into. (Thus, one can even adjust the value of n according to the number of mobile hosts or host density.) Finally, we comment on several minor things in the above analysis. First, if the control packets are of diﬀerent lengths, the 2Lc can simply be replaced by the total

length of RTS, and CTS. Second, the Ld has included the length of ACK packets. So

the real data packet length should be Ld minus the length of an ACK packet. Last,

we did not consider protocol factors (such as propagation delay, SIFS, DIFS, collisions of control and data packets, backoffs, etc.) in the analysis and hence the bandwidth considered above is not “effective” bandwidth. In reality, these factors will certainly affect the performance. In the next section, we will explore this through simulations.

(38)

4.2 Experimental Results

We have implemented a simulator to evaluate the performance of our GRID protocol. We mainly used the SCA protocol as a reference for comparison. SCA only differs from our GRID in its channel assignment strategy. Specifically, in SCA, the overall bandwidth is still divided into one control channel and n data channels. But each host is statically assigned to only one data channel. To use its data channel, a host must go through a RTS/CTS exchange with its intending receiver before using the data channel. Since both SCA and GRID use the same channel model and medium access approach, we believe that the experiment can give a clear indication how much more channel reuse that GRID can offer. Also, whenever appropriate, we will include the performance of IEEE 802.11, which is based on a single-channel model, to demonstrate the benefit of using multiple channels.

The parameters used in our experiments are: physical area = 1000× 1000, trans-mission range r = 200, hosts = 400, DIF S = 50µsec, SIF S = 10µsec, backoﬀ slot time = 20µsec, control packet length Lc = 100 bits. A data packet length Ld is a

multiple of Lc. Packets arrived at each mobile host in an Poisson distribution with

arrival rate λ packet/sec. For each packet arrived at a host, we randomly chose a host at the former’s neighborhood as its receiver. Both of the earlier bandwidth models are used. If the ﬁxed-channel-bandwidth model is assumed, each channel’s bandwidth is 1 Mbps/sec. If the ﬁxed-total-bandwidth model is assumed, the total bandwidth is 1 Mbps/sec. In the following, we make observations from four aspects.

A) Eﬀect of the r/d Ratios: In this experiment, we change the r/d ratio to observe

the eﬀect. We use n = 16 data channels and Ld/Lc = 200. Fig. 8 shows the network

throughput under diﬀerent loads under the ﬁxed-channel-bandwidth model. We can see that both SCA and GRID have similar throughput curves. When r/d = 0.5, 1.0, and 1.5, our GRID protocol is worse than the SCA protocol. When r/d ≥ 2.0, our GRID will outperform SCA. At r/d = 3.5, GRID will deliver the highest throughput, which is about 25% more than the highest throughput of SCA. After r/d > 3.5, GRID will saturate and degrade slightly, but still outperform SCA. It is worth to mention that according to our earlier ideal analysis in Section 2, the best performance of GRID

(39)

0 5 10 15 20 25 30 0 1 2 3 4 5 6 7 8 9 10

Arrival rate (packets/sec/host)

GRID(r/d=0.5) GRID(r/d=1) GRID(r/d=1.5) GRID(r/d=2) GRID(r/d=2.5) GRID(r/d=3) GRID(r/d=3.5) GRID(r/d=4) GRID(r/d=8) GRID(r/d=16) GRID(r/d=30) GRID(r/d=100) SCA Throughput (Mbps)

Figure 8: Arrival rate vs. throughput under the ﬁxed-channel-bandwidth model at diﬀerent r/d ratios with n = 16.

will appear when r/d = √n/2 = 2. This ratio is somewhat smaller than the ratio 3.5

that we obtain here. We believe that this is because in this experiment we have taken timing factors (such as different packet arrival time and different backoff intervals) into consideration, while in Section 2 we have disregarded this factor. Thus, different sender-receiver pairs may be time-differentiated, and thus more pairs may coexist. In fact, this is a favorable result to GRID because a higher r/d ratio means more signal overlapping, and thus higher channel reuse.

Fig. 9 shows the similar experiment under the fixed-total-bandwidth model. Again, the best r/d ratio appears at around 2.5 to 4. The trend is similar to that of the fixed-channel bandwidth model. Also, as a reference point, this figure contains the performance of IEEE 802.11.

B) Eﬀect of the Number of Channels: In this experiment, we still use Ld/Lc = 200,

but vary the number of channels n, to observe its effect. Fig. 10 shows the result under the fixed-channel-bandwidth model. Note that in this figure we have picked the best

r/d ratio (through experiments) for each given n for our GRID protocol. We see that

both SCA’s and GRID’s throughputs will increase as more data channels are used. This is quite reasonable because under the ﬁxed-channel-bandwidth model, a larger n means more total bandwidth being provided. As n enlarges, the gap between GRID and SCA will increase slightly.

(40)

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0 0.5 1

GRID(r/d=0.5) GRID(r/d=1) GRID(r/d=1.5) GRID(r/d=2) GRID(r/d=2.5) GRID(r/d=3) GRID(r/d=3.5) GRID(r/d=4) GRID(r/d=8) GRID(r/d=16) GRID(r/d=30) GRID(r/d=100) SCA 802.11 Throughput (Mbps)

Figure 9: Arrival rate vs. throughput under the ﬁxed-total-bandwidth model at dif-ferent r/d ratios with n = 16.

0 5 10 15 20 25 30 35 40 45 0 2 4 6 8 10 12 14 16 18 20 Arrival rate (packets/sec/host)

GRID(n=9,r/d=2.5) GRID(n=16,r/d=3.5) GRID(n=25,r/d=4) GRID(n=49,r/d=6) SCA(n=9) SCA(n=16) SCA(n=25) Throughput (Mbps)

Figure 10: Arrival rate vs. throughput under the ﬁxed-channel-bandwidth model with diﬀerent numbers of data channels.

(41)

0 0.5 1 1.5 2 2.5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Arrival rate (packets/sec/host)

GRID(n=4,r/d=1.5) GRID(n=9,r/d=2.5) GRID(n=16,r/d=3.5) GRID(n=49,r/d=6) SCA(n=4) SCA(n=9) SCA(n=16) SCA(n=49) IEEE 802.11 Throughput (Mbps)

Figure 11: Arrival rate vs. throughput under the ﬁxed-total-bandwidth model with diﬀerent numbers of data channels.

Fig. 11 shows the same simulation under fixed-total-bandwidth model. The trend is similar. One important observation is that the best performance for both SCA and GRID will appear at around n = 4 data channels. With more channels, the throughput will degrade significantly. Also, as comparing GRID and SCA, we see that when n is too large (e.g., n = 49), The gap between GRID and SCA will decrease significantly. This may due to two reasons: either the control channel is overloaded, or the control channel has not been fully loaded but there are too few mobile hosts to fully utilize these data channels.

C) Eﬀect of the Ld/Lc ratios: As discussed earlier, the performance of GRID will

be limited by the use of the control channel. One way to increase performance is to increase the data packet length in order to reduce the load on the control channel. To understand this issue, observe Fig. 12(a), which assumes Ld/Lc = 50 and the number

of hosts = 1600 under the fixed-channel-bandwidth model. Comparing the curves in this figure, we see that there is a large performance improvement between using n = 9 channels and n = 25 channels. However, the improvement reduces significantly from using n = 25 to using n = 49 channels. When using n = 100 channels, the gain relative to using n = 49 is very limited (note that under the fixed-channel-bandwidth model, this means much bandwidth being wasted). To resolve this problem, in Fig. 12(b), we increase Ld/Lc to 200. Now the improvements all enlarge. This has justified our

(42)

0 10 20 30 40 50 60 70 80 0 1 2 3 4 5 6 7 8 9 10

Throughput (Mbps) GRID(n=9,r/d=2.5) GRID(n=25,r/d=4) GRID(n=49,r/d=6) GRID(n=100,r/d=9) GRID(n=144,r/d=11) GRID(n=196,r/d=13) 0 20 40 60 80 100 120 0 1 2 3 4 5 6 7 8 9 10

Throughput (Mbps) GRID(n=9,r/d=2.5) GRID(n=25,r/d=4) GRID(n=49,r/d=6) GRID(n=100,r/d=9) GRID(n=144,r/d=11) GRID(n=196,r/d=13)

Figure 12: Arrival rate vs. throughput under the ﬁxed-channel-bandwidth model at diﬀerent numbers of data channels: (a) Ld/Lc = 50 and (b) Ld/Lc = 200.

10 11 12 13 14 15 16 17 50 100 200 400 800 1200 Ld/Lc Maximum throughput (Mbps) 17 17.2 17.4 17.6 17.8 18 18.2 18.4 18.6 200 400 800 1200 1600 2000 Ld/Lc Maximum throughput (Mbps)

Figure 13: Ratio Ld/Lc vs. maximum throughput under the ﬁxed-channel-bandwidth

model with n = 9: (a) bit error rate = 10−6 and (b) bit error rate = 5× 10−6.

argument. As a result, given an n, one has to wisely adjust the ratio Ld/Lc so as to

get the best throughput.

D) Eﬀect of Transmission Error Rates: In the previous experiment, we have made

a strong assumption: the transmission is error-free. To take this into consideration, we further assume a bit error rate during transmission. Under the ﬁxed-channel-bandwidth model with n = 9 channels, Fig. 13(a) and (b) show our simulation results under the transmission bit error rates of 10−6 and 5× 10−6, respectively. Under an error rate of 10−6_{, L}_d_/L_c = 800 has the best maximum throughput. With a larger error rate of 5× 10−6_{, the best maximum throughput will appear at the smaller ratio L}_d_/L_c = 400.

(43)

5 Conclusions

We have developed a new MAC protocol for a multi-channel MANET. Our channel assignment is characterized by location awareness capability and it incurs no communi-cation cost to conduct the assignment. This is a significant breakthrough compared to existing protocols which require clock synchronization and/or which dictate a number of channels which is a function of the network degree. Our simulation results have also indicated that it is worthwhile to consider using multiple channels under both the fixed-channel-bandwidth model and the fixed-total-bandwidth model.

In this paper, we focus on the scenario where hosts are randomly deployed. In such an environment, GRID is a simple yet eﬃcient solution. For larger areas where users have geographical locality, the GRID-B proposed in [29] tries to explore channel borrowing to make an eﬃcient use of channels. However, due to its channel relocation behavior, GRID-B involves higher complexity. The purpose of this paper is to develop a light-weight MAC protocol that is suitable for an ad hoc environment.

We believe that there are many open research problems from this work. In our simulations, we have used a number of data channels (n) which is a square of some integer. Other values of n deserve investigation. In practice, the best r/d ratio may change due to many factors, such as system load, which also deserves studies. While GPS is widely available, indoor positioning is still an open issue. Since our work relies on physical locations to assign channels, for indoor environment pre-assignment of channels to each location may be necessary.

References

[1] IEEE Std 802.11-1999: Wireless LAN Medium Access Control (MAC) and Physical

Layer (PHY) specifications. Institute of Electrical and Electronics Engineers, Inc., New

York, USA, 1999.

[2] M. Ajmone-Marsan and D. Roﬃnella. Multichannel Local Area Networks Protocols.

IEEE Journal on Selected Areas in Communications, 1:885–897, 1983.

[3] A. Bertossi and M. Bonuccelli. Code Assignment for Hidden Terminal Interference Avoidance in Multihop Radio Networks. IEEE/ACM Trans. on Networks, 3(4):441– 449, August 1995.