DTIC ADA520325: An Energy-conscious Transport Protocol for Multi-hop Wireless Networks

0.47 MB·English

by  Defense Technical Information Center

An Energy-conscious Transport Protocol
for Multi-hop Wireless Networks
Technical Report BUCS-2007-014(Revises BUCS-2006-025)
∗
N. Riga I. Matta A. Medina C. Partridge J. Redi
ComputerScience BBNTechnologies
BostonUniversity Cambridge,MA,USA
{inki, matta}@cs.bu.edu {amedina, craig, redi}@bbn.com
ABSTRACT gunto see effortsto reduce the energyconsumedinthe
radio circuit, which draws most of the battery power.
Wepresentatransportprotocolwhosegoalistoreducepower
Low power [36], adjustable power [40], as well as
consumption without compromising delivery requirements
two-stage receiver radios [30] have been designed and
of applications. To meet its goal of energy efficiency, our
implemented in an effort to minimize the energy con-
transportprotocol(1)containsmechanismstobalanceend-
sumption while transmitting and receiving packets. In
to-endvs.localretransmissions;(2)minimizesacknowledg-
order to minimize the energy spent on “idle” listening
menttrafficusingreceiverregulatedrate-basedflowcontrol
ofthechannel,newMACprotocolshavebeenproposed
combinedwith selected acknowledgementsandin-network
that implement coordinatedwake-upschedules that al-
cachingofpackets; and(3)aggressivelyseekstoavoidany
lownodestobeturnedoffforlongperiodsoftime. Also
congestion-basedpacket loss. Within a recently developed
in an effort to minimize network control (e.g. routing)
ultralow-powermulti-hopwireless networksystem, exten-
traffic, controlled scoping of topology information [29]
sive simulations and experimental results demonstrate that
has been proposed.
our transport protocol meets its goal of preserving the en-
Asanexample,theJAVeLENsystem[11,26]achieves
ergyefficiencyoftheunderlyingnetwork.
dramatic results demonstratingnetworksthat consume
1. INTRODUCTION
100 times less energy for the same effective network
Motivation: Multi-hop wireless networks are plagued goodput, compared to a typical 802.11 multi-hop wire-
withuniquechallenges: contentionforthewirelessmedium, less network running the OLSR routing protocol.
time-varying topology due to the variable quality of These type of highly energy-efficientsystems present
links or mobility, and power constraints imposed by a challenge to network protocol designers. All high-
battery lifetimes. The focus of this paper is the last layer protocols need to be examined and, if necessary,
challenge: power constraints. We seek to minimize the redesigned to make sure they are conservative in their
usageofenergysoastoextendthelifetimeofbothindi- transmissions. Aparsimoniousmedia-accessprotocolis
vidualnodesandthenetworkasawhole,whilemeeting oflimitedbenefitifapplicationschoosetobechattyand
the requirements of applications. consume power with low-value transmissions.
Whileenergyefficiencyhasbeenrecognizedasachal- In this paper we examine the problem of designing
lengeforsometime[37],untilrecently,verylittleprogress a transport protocol that is energy conserving and is
has been made. Inthe pastcouple of yearswe havebe- suitable for deployment over energy-conscioussystems.
Extensive studies (reviewed in Section 7) have demon-
∗This work was funded by the Defense Advanced Research
strated the inadequacy of TCP to serve as a trans-
Projects Agency (DARPA)undercontract NBCHC050053.
port protocolin wireless environments. Later attempts
to enhance or redesign TCP for wireless scenarios are
Permission tomake digital orhardcopies ofall orpartofthis workfor mainly focusedon improvingperformance metrics such
personalorclassroomuseisgrantedwithoutfeeprovidedthatcopiesare as goodput and delay, but not on minimizing energy
notmadeordistributedforprofitorcommercialadvantageandthatcopies consumption.
bearthisnoticeandthefullcitationonthefirstpage.Tocopyotherwise,to
Designing an energy-conscious transport protocol is
republish,topostonserversortoredistributetolists,requirespriorspecific
permissionand/orafee. simultaneously simple and hard. The problem is sim-
Anabbreviatedversionofthistechnicalreportappearsintheproceedings ple because,forpracticalpurposes,wecanusea simple
ofCoNEXT’07,December10-13,2007,NewYork,NY,U.S.A.
.
1
Report Documentation Page Form Approved
OMB No. 0704-0188
Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and
maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information,
including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington
VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it

does not display a currently valid OMB control number.
1. REPORT DATE 3. DATES COVERED
2007 2. REPORT TYPE 00-00-2007 to 00-00-2007
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER
An Energy-conscious Transport Protocol for Multi-hop Wireless
5b. GRANT NUMBER
Networks
5c. PROGRAM ELEMENT NUMBER
6. AUTHOR(S) 5d. PROJECT NUMBER
5e. TASK NUMBER
5f. WORK UNIT NUMBER
7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION
Boston University,Computer Science,Boston,MA,02215 REPORT NUMBER
9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S)
11. SPONSOR/MONITOR’S REPORT
NUMBER(S)
12. DISTRIBUTION/AVAILABILITY STATEMENT
Approved for public release; distribution unlimited
13. SUPPLEMENTARY NOTES
14. ABSTRACT
see report
15. SUBJECT TERMS
16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18. NUMBER 19a. NAME OF
ABSTRACT OF PAGES RESPONSIBLE PERSON
a. REPORT b. ABSTRACT c. THIS PAGE Same as 14
unclassified unclassified unclassified Report (SAR)
Standard Form 298 (Rev. 8-98)
Prescribed by ANSI Std Z39-18
metrictominimize: thetotalnumberofnodetransmis- agate information—to avoid maintaining per-flow
sions for eachpacket, or eachconversation.1 The prob- state and maintain the modularity of the system.
lemishardbecauseverylittle isknownaboutminimiz-
ingtransmissionsacrossmultiplenodes. Thereisplenty • JTP exploits any energy-gain opportunities pro-
of research work on minimizing transmissions end-to- vided by the applications. Historically, transport
end,especiallylimitingthenumberofacknowledgments protocols have offered a particular reliability/QoS
sent,butalmostnothingonintermediatetransmissions. model and the application’s task was to pick the
And the one key piece of work in the area, Ludwig’s transport protocol whose model most closely met
work on the interaction of end-to-end and hop-by-hop the application’s needs (e.g. UDP, TCP, ITP [25],
retransmissions [22], gives a depressing result: namely RTP [13]). JTP is designed to act as a “generic”
in a world where one needs both end-to-end and hop- transportprotocoltailoredbytheapplicationbased
by-hop retransmissions, it is very easy to have the two on its specific QoS semantics. JTP uses the tol-
transmissionmechanismsinteracttocausemore,rather erance of the application to losses and limits the
than less, total transmissions. Our transport protocol network’seffortintryingtodeliverapacket,based
coordinatesend-to-endandhop-by-hopretransmissions on the packet’s individual importance as well as
byallowingtheendsoftheconnectiontoexplicitlycon- current energy costs (Section 3).
trol the amount of local retransmission effort, based
on application’s reliability requirement. JTP’s objec- • In JTP, the receiver is fully responsible for con-
tive of minimizing the number of total transmissions is trolling all transmission parameters; connection’s
not only beneficial in energy-constrainednetworks,but sendingrate,retransmissionrequestsformissing/lost
in any wireless network since limiting the overall num- packets, as well as the frequency of such controls.
ber of link-layer transmissions effectively increases the To the best of our knowledge, JTP is the first
network-wide capacity. transportprotocolthatsupportsvariabledestination-
Our Contribution: Our transport protocol mediates controlled feedback trying to keep feedback as low
between an application’s need to share information of as the stability and reliability of the network per-
varying importance over the network,and the system’s mits (Section 5).
goal of minimizing energy expenditure per successfully
• JTP implements a caching mechanism which en-
delivered bit. Given that JAVeLEN is the state of the
ables intermediate nodes along the path of a JTP
art in energy conserving network systems, our protocol
connection to temporarily store traversing pack-
uses JAVeLEN as the underlying architecture and thus
ets.2 This enables the recovery of lost packets as
we will refer to it as JTP (JAVeLEN Transport Pro-
close to the receiver as possible. These pipelines
tocol). Although JTP is implemented to work within
of caches along paths generalize the single-level
JAVeLEN, it is designed to operate within any wireless
cachingoftenemployedincellular-type(singlewire-
(or wireline) architecture that provides an interface for
less hop) networks [5]. Although a system that
JTP to both, control the number of node retransmis-
supports symmetric routes betweenhosts,like the
sions made by the media-access protocol, and get at
JAVeLEN system, would exploit caching benefits
leastindicationoftheavailablecapacityofthechannel,
toitsfullest,theopportunisticdesignofthecaching
and of the packet loss rate. JTP’s architecture strives
system would seize any chance for locally recover-
tomaximizethemodularityofthesystem. Information
ing lost packets,without interfering with the end-
ispropagatedthroughpacketheadersandalltheneces-
to-end semantics of each connection.
saryhop-by-hopoperationsareperformedbyaseparate
module that is used by the MAC. To allocate bandwidth fairly among flows in the
In this paper, we present a summary of our design presenceofin-networkcachingandretransmissions,
and the novel features of JTP as follows: a JTP sender backs off its sending rate to ac-
• To the best of our knowledge, JTP is the first countfor“internal”retransmissionstriggeredfrom
caches on its behalf (Section 4).
multi-hop wireless end-to-end transport protocol
designed to perform hop-by-hop soft-state oper-
• JTP also employs a (congestion-avoidance) rate-
ations to improve (goodput and energy) perfor-
based flow control—using ATM-like explicit rate
mance while preserving the end-to-end principle
[28] (Section 2). 2We note that efficiency achieved by caching and repairing
JTP employs mechanisms akin to the Dynamic errors earlier using such in-network caching does not con-
tradict theend-to-endargument of system design [28]—the
Packet State [33]—using packet headers to prop-
source does not delete its copy of a packet until it gets an
1This metric admittedly ignores energy costs for computa- acknowledgment from the final destination that it has suc-
tion and memory inthenodes. Theusual logic for ignoring cessfully received the packet. Furthermore, the soft-state
these costs is that the radio consumes far more energy. natureof caches provides resilience to route changes.
2
feedbackfromthenetwork—inanattempttoelim- how JTP actually works.
inate energy wastage associated with congestion- Before we delve into the details of JTP, we provide
induced packet drops. a brief description of the JAVeLEN system. JAVeLEN
deploysaTDMAMACwhichorchestratesnodes’trans-
• JTP is implemented as “sharedcode” that can be
missions by using pseudo-random schedules, providing
run either in simulation or on real radio nodes.
collision-free access to the channel and allowing nodes
Results from simulation and from a prototype of
to turn off their radios when they are not in use. Each
JAVeLEN radios (Section 6) confirm the premise
node also keeps statistics about link transmissions and
of JTP in reducing the energy consumed per de-
idle slots in order to provide estimates of the available
livered bit.
transmission rate and of the packet loss rate on every
link. JAVeLEN uses an energy conserving link-state
2. JTPDESIGN
routing algorithm [29], that provides each node with a
Aswesoughttodesignaprotocolthatminimizes the
local, possibly inaccurate, view of the network’s topol-
total number of node transmissions required to deliver
ogy.
application data, we chose to decompose that goalinto
6: Path Monitor
three subgoals: 8: Transmitter Wireless Network 7: Path controller
depends on 5 and 7
(1)Minimizeend-to-endretransmissions. Ludwig’swork
showedthatweneededtostrikeabalancebetweenend-
to-end (source) and local retransmissions. End-to-end
2: Intermediate Nodes 3: Destination
retransmissions effectively waste all the energy already 1: Source of of JTP Connection
JTP Connection
expendedongettingthepacketatleastpartwaytothe 4: JTP packet 9: Mid-path operations
destination by the initial transmission. So, while occa- 5: Header erro(rc arecchoinvge,r yd,e sflcehcetidounl,ing)
(controlled transmission
sionalretransmissionsfromthesourcearerequired(e.g. rpaatera, meneetregrsy:, SNACK)
due to intermediate node failure or topology changes),
we sought to do everything we could to retransmit a
lostpacketfromthe farthestdownstream,intermediate Figure 1: Elements of JTP
node along the packet’s path which had received the
packet successfully.
2.1 APacket-based ViewofJTP
(2) Minimize acknowledgments. Acknowledgments are,
We begin by describing JTP in terms of packets—
often, pure overhead—they carry no application data.
data and acknowledgment packets—and how they flow
Yet, they consume roughly as much energy as a data
between source and destination. Figure 2 shows the
transmission. So, consistent with reliability and other
packetformatfor JTP packets. While the JTP header,
goals, we would like to minimize acknowledgments.
shown in Figure 2(a), is attached to all packets, the
(3) Avoid congestion loss in the nodes. By design, a
ACKheader,showninFigure2(b),isoptionalandisin-
classicTCP-likecongestioncontrolinducespacketdrops,
cludedonlyinpacketsthatcarryfeedbackinformation.
since lossis the onlysignofcongestion. The energyex-
Note that this is an optimized version of JTP head-
pended by packets that are discarded, simply to signal
ers that only carry the necessary information. In our
congestion, is wasted. In a world where energy is the
prototype implementationthe JTP headersareslightly
key metric, congestion control should consume less en-
longer, mainly for debugging purposes.
ergy than what discarding a packet implies. Assuming
that JTP is effective in only transmitting data when
necessary,JTPaimstoavoidcongestioninsteadofcon- 2.1.1 Data Packets
trolling it. Datapacketstravelfromthesourcetothedestination
The result of these goals was the JTP architecture ofaJTPconnection. InJTP,datapacketscontainthree
shown in Figure 1. JTP uses rate-based transmission novel fields: available rate, loss tolerance, and energy
controlled by the destination. Intermediate wireless budget.
nodes reporton their conditionin data packetheaders. • Available rate: The available rate of a link, from a
A Path Monitor and Path Controller at the destina- node to its neighbor, represents its current available
tioncollectthispathperformancedataandadjustdata transmission capacity—inaTDMAMAC,liketheJAVe-
rates to avoid congestion. Intermediate nodes retrans- LEN MAC, that available rate is determined by the
mit packets on a per hop basis, cache packets, and ex- current rate of unused (idle) time slots during which
amine end-to-end acknowledgments. If an acknowledg- the neighbor is awake for reception; in CSMA/CA net-
mentindicatesapacketwaslostfurtheralongthepath, works,amethodsimilarto [20]canbeusedtoestimate
the node will retransmit this packet, thereby avoiding theavailablebandwidth. Ateachnodevisited,apacket
an end-to-end retransmission (a la the work of Balakr- isstampedwiththeminimumavailableratecollectedso
ishnanetal.[5]). Basedonthis design,we nextlookat far alongthe pathof the JTPconnection As described
3
used exceeds the energy budget of the packet. This
approach provides an energy-conscious mechanism for
dealingwithroutingloops(asopposedtothetraditional
hop-countTTL)and, asweelaboratelater,inconjunc-
tion with the loss tolerance field, creates a sturdy way
to manage the energy expenditure per packet.
Theavailableratefieldaddressesthegoalofavoiding
congestion loss. Because JAVeLEN provides a practi-
callycollision-freeMAClayer,iftheJTPflowcontroller
ensuresthatitdoesnotdrivethe availablerateto zero,
congestion-induced losses are avoided.3
The energy budget and loss tolerance fields manage
the expenditure of energy per packet. By limiting the
effort of each node to successfully deliver a packet, the
losstolerancefieldboundsthetotaltransmissionenergy
spent by each node. The energy budget on the other
hand,setsanupperboundontheenergythatthewhole
network might expend to deliver a packet to the desti-
nation. The deadline field is used by real-time traffic,
andalthoughthis is outofthe scope ofthis paper. it is
included for completeness.
2.1.2 Acknowledgment Packets
Figure 2: JTP packet format
JTP ACK packets carry data acknowledgments, as
well as transmission rate and energy budget informa-
in Section5, this availablerate is used by the flow con- tion from the receiverto the sender. The rate at which
troller at the destination to update the sending rate of ACKsare fedbackto the senderis regulatedby the re-
the source. Note that due to retransmissions that may ceiver,basedonthestabilityconditionsofthepath. For
be required to get the packet to the next hop, a packet a stable path, a minimum feedback rate is determined
may consume more than one MAC-level transmission bytheapplicationbasedonitsrequirements—forexam-
slot. So the availablerate value must be normalizedby ple, anapplicationwith amore stringentdelayrequire-
the average number of MAC-level transmissions. Lud- ment would require a higher feedback rate to achieve
wig [22] pointed out that allowing multiple MAC-level a more timely recovery of missing data. A lower feed-
transmissionscanincreasepacketdelayandalsoreduce back rate, if tolerated by the application, allows JTP
the effective bandwidth that an application perceives. to aggregate ACK information in a single packet, thus
As we will see later, by allowing applications to control reducing feedback load.
the number of MAC-level transmissions per link, JTP Itiswellknownthatrate-basedflowcontrolisvulner-
effectively gives applications controloverthe delay and able to the loss of feedback messages. As we elaborate
effective bandwidth on every link. inSection5,JTPovercomesthisproblembyhavingthe
• Loss tolerance: A source node encodes the desired receiverinformthesenderofitsfeedbackrate. Soifthe
end-to-end loss tolerance in packet headers. Each node sender does not get an ACK within the expected feed-
along the path pre-computes the maximum number of back delay, it backs off its transmission rate. Further-
transmission attempts to the next hop, given the re- more, short-term, and significant, variations in path’s
maininglengthofthepath(knownfromthenode’sview conditions (available rate or energy used) would be de-
ofthetopology)andpacket’slosstolerance. Thepacket tected by the path monitoring function at the receiver,
is dropped if this pre-determined maximum number of thus triggering an early feedback.
local attempts is exceeded. As described in Section 3, Inadditiontoreportingrateandenergyinformation,
beforeforwardingthe packet,the nodeupdatesthe loss an ACK packet carries both positive cumulative and
tolerance field so any left-over attempts (from the pre- negativeselectiveacknowledgments. Eachintermediate
determined maximum number) do not get used down- node on the path examines the SNACKs and retrans-
stream, thus reducing the variability in energy con-
sumption across nodes along the path. 3EveniftheunderlyingMACdoesnotprovidecollision-free
•Energybudget: Thesourceinitiallyassignseachpacket access, JTP’s operation will not be affected,since collisions
would only increase the link loss experienced by packets,
an energy budget value based on the energy the net-
thusincreasingthenumberoflink-layerretransmissionsper
work would typically expend to deliver the packet suc- packetandeffectivelyreducingthemeasuredavailableband-
cessfully. A packet is dropped whenever the energy width, which in turn forces the sources to back off.
4
mitsmissingpacketsifthesepacketsareintheinterme- Algorithm 1 PreXmit()
diate node’s cache—if they are, the node appropriately 1: increaseEnergyUsed(packet);
modifies the ACK packet so the sender is explicitly in- 2: if (packet.energyUsed > packet.energyBudget)
formed of such in-network retransmissions done on its then
behalf (cf. Section 4). Caching, and acknowledgments, 3: dropPacket(packet);
are discussed in more detail in Section 4. 4: else
2.2 JTP Components 5: if firstDataTransmission(packet)then
6: lossRate=getLinkLossRate(packet);
As mentioned earlier in Section 1, although JTP is
7: setMaxDataTransmissions(packet,lossRate);
an end-to-end transportprotocol, it also performs hop-
8: updateLossTolerance(packet);
by-hopoperationstoimproveits performance. Inorder
9: end if
to alleviate the overhead of the hop-by-hop operations
10: rate=getAvailableRate(packet);
and increase the protocol’s efficiency, JTP is divided
11: packet.rate=
intotwomaincomponents: theend-to-endJTP(eJTP)
12: MIN(packet.rate,rate/AvLinkLayerAttempts);
and the hop-by-hop (or, intermediate) JTP (iJTP).
13: end if
2.2.1 End-to-end JTP
The functionalities of eJTP are further categorized Algorithm 2 PostRcv()
into application-specific and network-specific modules. 1: if (packet.type==DATA) then
This modularity makes JTP a more versatile transport 2: cachePacket(packet);
protocol. In this work, we focus on bulk data transfers 3: else if (packet.type==ACK) then
over multi-hop wireless networks. 4: retransmitPackets(packet.SNACK)
•Theapplication-specificmoduleofeJTPisresponsible 5: updateACK(packet);
for fragmentationandreassemblyofapplication’sdata. 6: end if
Moreover,itexpressesthe QoSrequirements(e.g.relia-
bility level) to the network-specificmodule to influence
in-network packet handling decisions, flow control, and reflect its value for the remainder of the path.
retransmission requests only for those missing packets BesidesthefactthatiJTPmustprocessallJTPpack-
that are important to the application. ets that pass through a node, the above described soft-
•Thenetwork-specificmodulemanagesallJTPconnec- stateoperationsrequirethecraftingofcross-layerinter-
tionsandimplementsthepathmonitorandthe(congestion- actions with the MAC layer. In order not to compro-
avoidancerate-based)controllerofper-packettransmis- mise the performance ofa node, by redundantcopying,
sion parameters. context switching, and message passing between JTP
andtheMAClayer,weimplementediJTPasaseparate
2.2.2 Hop-by-hop JTP loadable plug-in module of the MAC protocol. iJTP is
iJTPperformsallthemid-path(intermediate)cache- invoked whenever a packet is received or about to be
related and soft-state operations. It manages the local transmitted over the air interface.4
cacheofeverynodebyensuringthatonlyvaluabledata Algorithms 1 and 2 present in pseudo-code the op-
packets are stored, data packets indicated in SNACKs erations of iJTP, which is invoked exactly before the
are retransmitted, and cached data packets are evicted transmission(Algorithm1)andexactlyafterthe recep-
based on the selected caching policy. tion (Algorithm 2) of a packet from the physical layer.
Soft-state per-packet operations performed by iJTP .
at each node, include:
3. ADJUSTABLE RELIABILITY FOR EN-
•Update the available rate field: iJTPisresponsiblefor
ERGYCONSERVATION
acquiring from the MAC layeran estimate of the avail-
ableratetoeveryneighbor,aswellasanestimateofthe Not all applications (e.g. voice, video, images [25])
packet loss rate on that link. It uses this information require full reliability to performwell. Given reliability
toestimatetheeffectiveavailableratetoeachneighbor. targets from applications (provided by the application
iJTP stamps each passing packet with the lowest effec- module), and knowledge of packet loss rates (provided
tiveavailablerate(throughput)observedthusfaralong bytheMAClayer5),JTPadjuststheeffortitputsinto
the path.
• Update the loss tolerance field: Based on the loss tol- 4AlthoughwithintheJAVeLENsystem,iJTPresidesinthe
MAC,iJTP operationscan beperformed byeJTP usingan
erance carried in the packet’s header and the link’s es-
overlay typeof architecture.
timated packet loss rate, iJTP sets the number of data 5The MAC layer keeps statistics about successful packet
transmission attempts on that link, as we elaborate in transmissionsandestimatestheaveragepacketlossrateex-
Section 3. The loss tolerance field is then updated to perienced overeach link.
5
delivering each packet, seeking to expend only enough links or on nodes with higher available energy—in this
energy to deliver “needed” packets. Specifically, iJTP paperweassumethatJTPattemptstoassignthesame
controls the number of node transmission attempts on q =q for all the links. From Equation (1) we get:
i
a per packet basis. In JTP, the reliability level is ex-
pressed in terms of the loss tolerance percentage (e.g. q =(1−lti)H1i (4)
10% of packets can be lost) encoded in the header of Bydynamicallyadjustingthehop-by-hopsuccessprob-
each packet. ability experienced by each packet, the end-to-end re-
Letle2e betheend-to-endlosstolerancerequestedby liability requirements are met even if the topological
the application. Let ni, i ∈ [0,H] be the nodes on the views at different nodes are inconsistent, or the path
path from the source n0 to the destination nH, where changes. Moreover,by assigning a loss tolerance target
H is the total number of links on the path. Let qi, i∈ to each individual packet, JTP enables the application
[0,H −1] denote the probability that a packet sent by to prioritize its packets (e.g. video frames of varying
node ni will be successfully delivered to the next node importance).
ni+1. In order to satisfy the end-to-end loss tolerance WetestedJTPunderdifferentreliabilitylevels. Three
of the application, the following equation should hold: different levels are considered 0% (jtp0), 10% (jtp10)
le2e =1−ΠHi=−01 qi (1) and 20% (jtp20).
The value of q changes depending on the number of
i (a) Total Energy spent for transfers of
node transmission attempts indicated to the MAC by 0.6 different reliability levels. 1.35x 104 (b) Data delivered to application.
ittfwJorohaTrtianaPclsh.mpnLawuiscemesktiebcoptaeninrofdnroceooflnmimnnokotpnedui(etnttehoit,erMnnapini+i+rs1oa1mbs)fa:.ia6sibTslsiiol,hinateynnadtqthtileea=mttMp1at−issdipnreeMignqlioeu,tenefsrottohdemdee Total Energy Spent(Joules)00000.....012345 2 3 4 5 6 7 jjjtttppp0128 00 dataDelivered(kb)111111...012...1123555 2 3 4 5 6 7 89800 % %9
netSize netSize
log(1−q ) (c) Maximum Number of link layer retranmissions for Node 3
M =max(1,min((cid:6) i (cid:7),MAX ATTEMPTS)) 5
whiereMAX ATTEMloPg(TpiS)isthemaximumnumber(2o)f #link layer rtx234 jjttpp1200
link transmissionsthat the MAC allows. The challenge 1 1020 1040 1060 1080 1100 1120 1140
time(sec)
istodynamicallyadjustthevaluesofM foreachpacket
i
in a flow so as to satisfy the desired le2e. Figure 3: Different reliability levels: jtp0, jtp10,
If the length of the path to the destination is known, jtp20 have loss tolerance of 0%,10% and 20%, re-
the values for q ’s canbe directly computed from equa- spectively.
i
tion (1), and encoded in the headers of packets. How-
Figures 3(a) and (b) demonstrate that by neither
ever, in a network setting where the topological views
overachieving (e.g. TCP-like full reliability for all), nor
atthenodesaretypicallynotaccurate,thepathlength
underachieving(e.g.UDP-likenoreliabilityforall),JTP
isestimatedbasedonanode’scurrentview. Therefore,
manages to save energy and still satisfies an applica-
JTP carries out the computation of q ’s at each node,
i
tion’s requirements, denoted by the horizontal dotted
asthe packettravelstowardthe destination,thususing
lines in Figure 3(b).7 (See Section 6 for simulation de-
increasingly more accurate views.
tails.) Figure3(c)showsthe maximumnumberofnode
Let l be the loss tolerance that is encoded in the
ti
retransmissionssetbyiJTPforeachpacketatthethird
packetwhenreceivedbynoden . LetH bethenumber
i i
node along a 4-node path. Packets that request higher
of links from n to the destination. Before the node
i
reliabilityareretransmittedmoretimesbythelinklayer
transmitsthepackettoitsnext-hop,itupdatestheloss
in cases when the link quality is not good enough, and
tolerance field as follows:
thusJTPdirectlyaffectstheamountofnetwork’seffort
Πi+Hi−1q =1−l ⇒q Πi+Hi−1q =1−l ⇒ (3) in delivering each packet.8
j=i j ti i j=i+1 j ti
1−l 4. CACHING
lt(i+1) =1− qi ti JTP employs in-network caching of data packets to
Althoughtherearemanydifferentstrategiesthatmight avoid end-to-end (source) retransmissions as much as
be employed to compute qi on each link—e.g. impos- 7The jtp20 plot is always above the application’s require-
inghighersuccessfuldeliveryrequirementonlessloaded mentsincetheaggregatelossrateofthepathisbelow20%.
P 8The flow that corresponds to 0% loss tolerance is omitted
6The success probability is equal to kM=i0−1pki(1−pi) = from this plot since iJTP will always assign its packets the
(1−pMi). maximum numberof nodeattempts.
i
6
possible. Caching can be viewed as a second line of de- 6x 10−5 0.105 jtp
fadtteiheontasnesttme.iniitpafotiaisrolalnmrMyimsAesaxiCnyc-gel,esrsveaiqevnluedaedtsatteregemarnaspdyttiasllotfifnaoeintleh;diefnoesdrelienbcxkayacqmhtuhepaedlleia,tppyda.puclTkeicehtatoes- energy per delivered bit (Joules/bit)12345 jjtnpc Energy Spent(joules)000...00000..078900.555891 23% mtoo rme ifdapira athll oncoadteiosn jnc
UponreceivingatraversingACKpacket,anodechecks 03 4 5 6 7 8 9 1 2 3 4 5 6 7
netSize Node ID
the Selective Negative ACK (SNACK) field to deter- (a) Total energy expended (b) Energy expended by each
mine whether any packet(s) requested for retransmis- perapplicationdatabitdeliv- node in a 7-node linear topol-
sionexists inthe node’s localcache. Requestedpackets ered. ogy.
found in the cache are forwarded downstream toward
Figure 4: Comparison of JTP against JTP with
the destination.
No Caching (JNC) in static linear topologies.
BesidestheSNACKfield,anACKpacketheaderalso
retransmitted from the source.
contains a locally-recovered packets field, used to indi-
Denote by S > k, the random variable representing
cate which of the packets requested for retransmission
the total number of packets sent by the source until
have been already locally retransmitted by some node.
k packets are successfully delivered at the destination,
Upstream nodes check this field to avoid multiple re-
transmissions of the same packets. When the source then E[S] = qek2e, where qe2e is the end-to-end success
probability.
of the transfer receives an ACK, it will only retransmit
When a packet is received at a node, the average
packets that remain in the SNACK field.
number of node transmissions that it triggers is:
If the cache of a node becomes full, to insert a newly
arriving packet, the packet evicted from the cache is E[TJNC]=(1−p)+2(1−p)p+···+n(1−p)pn−1+npn
theleastrecentlymanipulated,i.e.LeastRecentlyUsed l
1−pn
(LRU) policy. Themotivationisthatit isunlikely that =
1−p
those packets not recently requested for retransmission
wouldbeeverrequestedinthefuture. Adetailedstudy Giventhatthelinksuccessprobabilityisq =(1−pn),
of different cache replacement strategies is the subject theprobabilitythatapacketmakesitoverilinksisqi,9
of future work. whichthentriggersE[TJNC]nodetransmissions. Thus,
l
thetotalnumberofnodetransmissionsforJNCisgiven
4.1 AnalysisofIn-network CachingGain
by:
Inthis subsection,weprovideananalyticassessment
of the benefits of caching. H(cid:1)−1
• JTP with caching: In order to compute an upper E[TtJoNt C]= E[S]×qi×E[TlJNC]
bound on the gains achieved by caching, we assume i=0
a best-case scenario whereby cache sizes are infinite, = k(1−pn)(1−(1−pn)H) ≈ k×H
and the path is symmetric, thus each lost packet will (1−pn)H(1−p)pn (1−pn)H−1(1−p)
be recovered by the last downstream node which has (6)
successfully received it.
For H = 1, equation (6) degenerates to (5). Observe
Wecomputetheexpectedtotalnumberofnodetrans-
1
missions, denoted by E[TtJoTtP]. In the presence of in- that the cost of JNC is (1−pn)H−1 times higher than
networkcaching,eachpacketwilleffectivelyberetrans- that of JTP.
mittedovereachlinkforasmanytimesasneeded,until To confirm the energy gains of caching, we compare
it is successfully delivered to the next node. If p is the bysimulationJTPandJNCinFigure4. (SeeSection6
linklossprobability,theexpectednumberofnodetrans- for experimentaldetails.) Figure4(a)showsthe energy
missionsonalinklfollowsageometricdistributionwith expenditure per successfully delivered application bit.
mean E[TJTP]= 1 . As expected, as the path between the source and the
l 1−p
The expected totalnumber ofnode transmissionsre- destination increases,the energy gains achievedby em-
quiredbyJTPinordertodeliverkpacketsoverH hops ploying in-network caching also increase. Figure 4(b)
is given by: shows the per-node energy consumption in a linear 7-
node topology. Observe that JTP not only expends
1
E[TJTP]=k×H × (5) lesstotalenergy,asexpected, butalsodistributes more
tot 1−p
fairly the energyexpenditure amongnodes inthe path.
• JTP without caching: In the case of JTP with no
in-network caching (henceforth denoted by JNC), over 4.2 FairIn-network Caching
each link, the packet is transmitted, say, at most n
times. If its transmission still fails, then it must be 9qe2e =(1−pn)H.
7
(a)JTP with Backoff (b)JTP without Backoff
reception rate (pps)02468 1000 12FF0lloo0ww112400 1600 1800 reception rate (pps)0246810001200tim14e0(s0e1c6)00FF1ll8oo0ww012 tssJorheoafTnuteOtPedhrfn,cireoneeeevcgqJdoeouTrrfeneasPnnntithge,eeccne.wyctoTiacloltiotolnrnenawkc’tcsnsthhesspipiomscastfhetaiohsntmtshf,dhipeoJs,eulnTtepdchPahepmestaaitheersisn.anatsdtmaouhL-ttereieaooettdmem-naerxinsnpeni,ndotiidtmsdvcsueeeuoncolmimhyonftpfmateaohosvrintemnaaheiislsemnatnttabieihzttowlheeeef
time(sec) sampleofapath’smetric,x¯the estimatedaverage,and
s)3.5 s)3.5 R¯ the estimated deviation from the average. We use
p p
e (p 3 e (p 3 principles from statistical quality control [23] to detect
n rat2.5 n rat2.5 asignificant changeinthepath’sstate,whichthentrig-
ptio 2 Flow1 ptio 2 Flow1 gers the destination to send additional feedback signal,
ce Flow2 ce Flow2 inadditiontofeedbacksentregularlyatalowfrequency.
re11.5000 1200 1400 1600 re11.5000 1200 1400 1600 To that end, we estimate the EWMA’s x¯ and range
time(sec) time(sec) R¯ as follows:
Figure 5: Short-term (top row) and long-term
(bottom row) average of the reception rate for x¯ = (1−α)x¯+αxi, initially x¯=x0 (7)
two competingflows: (a)the source backs off for R¯ = (1−β)R¯+β |xi−xi−1 |, initially R¯ = x0
2
locally recovered packets; (b) it does not.
R¯ is used to estimate the deviation around x¯ and is
calculated only from samples x within the following
i
Enabling mid-path nodes to retransmit packets on
upper and lower control limits:
behalf of sources may cause a violation of sending rate
compliance imposed by the destination of a transfer. R¯ R¯
UCL=x¯+3 ; LCL=x¯−3 (8)
For the sake of fairness and congestion control, the 1.128 1.128
source node must incorporate in-network retransmis-
Under normal operation, stable EWMA filters are
sions in its sending rate calculations. To this end, JTP
employed, i.e. the weights α and β are small so short-
forces sources to back off in accordance to the extra
termvariationsarefilteredout. Aslongasx lieswithin
i
traffic that is induced by in-network retransmissions.
the control limits, the state of the connection’s path is
When an ACK packet is received, the source uses the
considered stable and feedback is only reported to the
locally-recovered packets fieldto adjustitssending rate. source at low frequency, say every T seconds. 10 Oth-
Letr(t)betherateindicatedtothesourcebyareceived
erwise, x is consideredanoutlier. A certainnumber of
i
ACKattime t. LetN bethe numberofpacketslocally
consecutive outliers is used as indication of significant
recovered within the network, and let s , j ∈ [1,N]
j and persistent change in the state of the path, which
be the sizes of these packets. The source computes an
wouldthentriggeranimmediatefeedbacktothesource
appropriate back-off period t as follows:
b node. At this point, eJTP at the destination switches
(cid:2)
N s toanagile EWMAfilterwherealargerαvalueisused,
t = j=1 j so that x¯ catches up with the actual value. Once x
b r(t) i
falls back again within the control limits, eJTP at the
Figure 5 shows the short- (top plots) and long- (bot- destination switches back to the stable filter for this
tom plots) term average of the packet reception rate connection. This usage of both stable and agile filters
(throughput)atthedestinationfortwocompetingflows. is known as a Flip-flop Filter [6].
(See Section 6 for simulation details.) Flow 1 does In our implementation, we set T as a function of the
notrequestpacketretransmissions(i.e.UDP-likeflow), sending rate and of the cache size used in the network.
while flow 2 requires that all its packets be delivered Specifically,
and thus invoking the local recovery mechanism of the T =max(TLower Bound,n× Sendin1gRate); n≥1.
in-networkcaches. Weobserveintherightplots,spikes Notice that this ensures that the destination does
inthe receptionrateofflow2whenit doesnotbackoff not send feedback/SNACK messages at a rate higher
its sending ratetoaccountforits additionalin-network than the sending rate. The value ofTLowerBound is de-
retransmissions—the unfairness introduced is more ev- pendent on the size of in-network caches, since if pack-
ident from the long-term average plots. ets requested for retransmissionby a feedback message
have already been evicted from the cache then the en-
5. DESTINATIONBASED CONTROL ergy savings achieved by infrequent feedback messages
10The value of T is used to set the sender’s timeout field in
5.1 Path MonitoringusingFlip-flopFiltering
theJTP ACK header shown in Figure 2(b).
8
wouldbeoffsetbytheenergyconsumedbypacketsthat
440 55
have to be retransmitted from the source. If C is the 420 50
caancdhReTsTizTLeo—iwseatrhBceoorunonfiudgnu≤dr-aCtbril×pe tpSiamernaedmfionertgetRhraeotfceot−nhneRencTteTitownotrhke—n energy spent(mJ)333342468000000 packet drops223344050505
300 evnareiragbyle s fpeeendtb wahcken using 15 queue drops when using variable feedback
2800 0.1 0.2 0.3 0.4 0.5 100 0.1 0.2 0.3 0.4 0.5
feedback rate (pkts/sec) feedback rate(pkts/sec)
(a) Total energy expended in (b) Total number of packet
thesystem. dropsinthequeuesofthesys-
tem.
Figure 7: The gains achieved by using variable
feedback rate.
the path of the JTP connection. Let A¯ denote the av-
Figure 6: The effect of cache size for various erage available path rate measured at the eJTP des-
network sizes. tination, and δ ≥ 0 indicate the target available path
rate. IfA¯>δ thenthe sourceincreasesits sendingrate
Figure 6 shows the effect of cache size on the per-
r in proportion to the current available capacity and,
formance of JTP for different network sizes and feed-
to improve fairness among competing flows, inversely
back rates. In this experiment one JTP flow is started
proportional to the current sending rate:
over linear networks. (See Section 6 for experimental
details.) The figure depicts the number of source re- A¯(t)
transmissions for increasing cache sizes. We observe r(t+1) = r(t)+K , 0<K <1 (9)
I I
r(t)
the sudden drop in the number of source retransmis-
sionsoncethecachesizeislargeenoughtoholdmissing On the other hand, if there is little available rate (A¯
packetsuntiltheyarerequestedforretransmissionfrom
< δ), then the source decreases its sending rate multi-
the caches. Increasing the cache size further does not
plicatively:
improve the performance significantly.
Figure 7 depicts the energy gains achieved by using r(t+1) = K r(t), 0<K <1 (10)
D D
variable-rate feedback instead of a constant rate. In
this experiment, a linear topology of 8 nodes is used, We note that the eJTP destination also limits the
with one long-lived flow competing with several short- sending rate by its delivery rate up the stack to the
lived flows. (See Section 6 for simulation details.) We receiving-side of the application.
varytherateofconstant-ratefeedback—asexpected,as
the feedback rate increases, the total energy consumed 5.2.2 StabilityAnalysisofPI2/MDRateController
(Figure7(a))increasessincemorefeedback/ACKpack-
Inordertoanalyzethestabilityofthecontroller,con-
etsaregenerated;andforlowfeedbackrates,weobserve
sider a single JTP flow adapting its sending rate overa
a high number of packet drops in the queues of inter-
fixed-capacity channel. For analytical tractability, let’s
mediate nodes (Figure 7(b)) because the sender of the
ignore the EWMA computation of the available rate,
long-lived flow does not back off fast enough causing
thatis,ifr(t)<C,thenthe JTPsourceadaptsitsrate
congestioninthe system. Using variable-ratefeedback,
as follows:
JTPnotonlyachieveslowenergyconsumption,butalso
minimizes packet drops in the system since whenever (C−r(t))
r(t+1) = r(t)+K × (11)
the system load increases, it sends a timely feedback I r(t)
forcing the sender to back off.
On the other hand, if r(t) > C, then the JTP source
5.2 CongestionAvoidance Mechanism
adapts its rate as follows:
When the flip-flop path monitor triggers a new feed-
back message, the path controller at the destination r(t+1) = K ×r(t) (12)
D
should set the transmission parameters (sending rate,
energy budget), to be used by the source for the subse- Observethatthesystemremainsnon-linear,withtwo
quent packets until a new feedback is received. operating regions determined by whether the sending
rate r(t) is less than or greater than the capacity C.
5.2.1 PI2/MDSendingRateController
We next consider each of these two regions, and prove
This controller at the destination sets the sending stability by showing that the value of a positive Lya-
rate using the minimum available rate collected along punov function V(r) decreases with each iteration.
9