A Study on Data Aggregation and Reliability in Managing Wireless Sensor Networks

Transcription

1 A Sudy on Da Aggregion and Reliabiliy in Managing Wireless Sensor Neworks S Brown Deparmen of Compuer Science Nional Universiy of Ireland, Maynooh sephenbrown@nuimie C J Sreenan Mobile & Inerne Sysems Laborory, Universiy College Cork, Ireland csreenan@csuccie Absrac Using da aggregion o reduce power consumpion is a well sudied area In his paper we show a new resul: da aggregion can also be used o increase reliabiliy In scenarios where sus is being colleced from every node (eg nework managemen), one measure of reliabiliy is he compleeness (nodes included) Using some of he power savings from aggregion o increase he reransmission limi for aggreged da gives significan improvemens in compleeness The conribuion of his paper is in providing a deailed comparison of he compleeness and energy efficiency for aggreged and nonaggreged da Resuls, derived by numerical analysis and simulion, show h wih da aggregion and increased reransmission limis, reliabiliy can be subsanially improved, while sill saving power These resuls are applicable o designing nework managemen proocols for Wireless Sensor Neworks, as well as for oher scenarios where high compleeness is required for nework-wide da collecion Inroducion Deermining he collecive sus of all he nodes in a WSN (wireless sensor nework) is a fundamenal building block for nework managemen For large scale neworks, he individual managemen of nodes is impracical (and inefficien) The moivion for he work presened here is he monioring of node sofware upde sus as par of he sofware upde process, as discussed in [] and [2], bu he resuls presened here are applicable o oher nework-wide Suppored by SFI/Tyndall Nional Insiue under gran NAP-8 managemen and da-collecion aciviies requiring near-% compleeness In his paper we invesige he problem of reliably collecing sus informion from all he nodes in a WSN based on a reques from a managemen sion This form of sus collecion differs from sensor even collecion in several ways: i does inherenly require a flow of da from each node; i is infrequen; i is riggered by a sensor even; i requires a resul h is colleced from near- % of he nodes; and i needs o work on a wide range of WSNs (hus implying he need for scalabiliy) There is a lack of exensive sudies on he relionship beween da aggregion and reliabiliy (as measured by compleeness) despie he imporance of aggregion for collecing da in Wireless Sensor Neworks The conribuion of his work is o provide a deailed sudy, allowing various design radeoffs o be clearly idenified 2 Reled Work Da Aggregion is a key mechanism for WSNs, and reduces energy use hrough reduced ransmissions by processing raw da en roue and delivering jus he resuls [3] Efficiency issues are addressed in SPIN [4] which uses mea-da communicions o elimine redundan da ransmission, and LEACH [5] is a rouing proocol for low energy, hierarchical clusers, aggreging da and ransmiing i direcly o he base sion In [6] a new aggregion qualiy meric DAQ is proposed, and used o propose wo new proocols for aggregion: Enhanced LEACH and Clusered PEGASIS, where da aggregion akes place in he cluser heads In [7] i is shown h da disribuion and collecion asks can be performed opimally on ree neworks In [8] he performance of a greedy /7/$25 27 IEEE Auhorized licensed use limied o: The Library NUI Maynooh Downloaded on March 3, 29 2:9 from IEEE Xplore Resricions apply

2 algorihm o build rees in neworks of differing densiies is evalued Our work does examine he spanning-ree generion or reques phases, as suppored for example in Direced Diffusion [9], SNMS/Drip [], and REAR[] TAG [2] inroduces he idea of aggregion as a high-level, applicion-independen service o realize hese performance benefis I also inroduces he idea of an aggregion funcion This is developed furher in [3], which exends he classes of da aggregion from simple ypes such as SUM and COUNT o include MEDIAN, Consensus values, Hisograms, and ranges Reransmissions can be used o improve reliabiliy: as shown for PSFQ in [4], hop-by-hop error recovery is significanly more effecive han endo-end recovery in a wireless environmen For a WSN environmen, da aggregion does effec reliabiliy in erms of he number of repors received a roo node from a source (he expeced value) when compared o he non-da aggregion case [5] Bu i does increase he variabiliy 3 Compleeness We define compleeness o be a measure of he reliabiliy of he response o a query made o muliple nodes I is defined as he rio of he acual number of responses received o he expeced number of responses A C ( A, E) = E where: A = {x a response from node x is received} E = {x a response from node x is expeced} A E The compleeness depends on he effeciveness of wo facors: he query propagion, and he response delivery In his paper we address he impac of da aggregion on he response delivery effeciveness The compleeness of a nework-wide managemen response depends on he end-o-end reliabiliy of each node in he WSN There are wo way o ry and ensure maximum end-o-end reliabiliy in a WSN: sending he same dum muliple imes in a sream, hus increasing he probabiliy h leas one of he insances will be received [9][]; and sending he dum once ( one-sho ) wih acknowledgemens and local re-ransmissions [4] o improve he probabiliy h an undupliced da insance will be received As we show in our resuls, a one-sho response resuls in significanly less raffic In he examples addressed here, we consider a WSN scenario where all he nodes repor aggregable da o a roo node By aggregable, we mean da h allows for he aggregion of muliple values from muliple nodes wihou increasing he space requiremens in he packe(s) used o ransfer he da: i is hus scaleable Neworkwide sus informion such as SUM, COUNT, minima, and maxima would fall ino his cegory [3] A nework-managemen example, based on he original moivion for his sudy, would be in deermining he number of nodes h are running a new sofware version following he acivion of a sofware upde [] The compleeness of an aggreged repor measures he number of responses incorpored ino h repor 4 Analysis The parameers used are defined as follows: a: aggregion degree in a packe (response coun) d: spanning-ree deph (in hops from he sink) f: spanning-ree fan-ou (number of children) n: number of reporing nodes p: per-link probabiliy of single ransmission success q: per-link probabiliy of a packe being received successfully wih reransmissions : ransmi limi (for non-aggreged da) he maximum number of reransmissions=(-) u: modified ransmi limi for aggreged da (see he policies defined in secion 5) v: he acual number of ransmissions used on a link (including da & acks) For a se of responses sen up a ree from a node (Fig A) he number of repors received he sink is he same wheher each repor is sen individually, or all are aggreged ino a single packe [5] Sink Source (A) Single Source Sink Source N Source 2 Source (B) Muliple Sources Figure Single & muliple da sources We now examine he siuion where every node in he ree is a source (Fig B) There are wo cases: each node sends i own repor, or each node Auhorized licensed use limied o: The Library NUI Maynooh Downloaded on March 3, 29 2:9 from IEEE Xplore Resricions apply

3 aggreges is own da as he repor from Source propages up he ree 4 Non aggreged case Pr(receive from source deph i)= q i A = n i= q i Packes=v+(v+qv)+(v+qv+q 2 v) - n erms = v ( n + ( n ) q + ( n 2) q 2 + ) - n erms Pr(% success)=q n *q n- *q n-2 *q=q n(n+)/2 42 Aggreged case A = q ( + q( + q( + q())) = q i= Packes=v+v+ +v n erms =vn Pr(% success)=q n Thus he aggreged case provides he same average compleeness, wih a higher probabiliy of % compleion, using fewer packes Thus we can improve he compleeness by using some of he saved packes for exra reransmissions Due o he complexiy of he equions we explore his numerically 5 Numerical analysis In he analysis shown below, figures of p=7 and =3 are used, giving a base per-link packe success re (q) of 973 (wih reransmissions) The resuls for a balanced spanning-ree of 258 nodes are shown in Table and Fig 2 These show he improvemen in he compleeness of he response for five differen aggreged ransmi limi policies: a) non-aggreged da (ransmi limi ) b) aggreged da (modified ransmi limi u=) c) aggreged da (u=2) d) aggreged da (u=) e) aggreged da (u=) f) aggreged da (u=) g) aggreged da (u=) In his analysis we are using packes ransmied o compare power consumpion: ypical WSN radio receivers (eg CC242) use he same amoun of power in receive mode wheher acually receiving or We also assume conenion is avoided by he MAC layer in pracice, reducing he number of ransmissions will reduce MAC-layer conenion [9] and hus increase he power efficiency n i %-age Repors Delivered Aggregion Figure 2 Compleeness vs aggregion policy (d=3,f=6,=3,p=7,n=258) Average SD Noe h he variabiliy, shown by he sandard deviion (SD) for aggreged da (u=) is similar o h for non-aggreged da Noe also, h as he maximum ransmi limi increases (from o ) he average number of repors delivered approaches % The deail in Table shows h he energy used increases slighly as u increases, bu is significanly less han for non-aggreged da The energy per repor is also significanly lower Da aggregion has he maximum effec for a very deep spanning-ree, and minimum effec for a very shallow one Fig 3 shows he effec on he compleeness of varying he heigh and fan-ou of he spanning-ree The oal number of reporing nodes is kep approximely 256 as a balanced spanning ree is used i is possible o keep his number exacly consan, and he repors received is normalized in he range () Table Compleeness for differen policies (d=3,f=6,=3,p=7,n=258) Repors Avg Repors SD Toal-x No agg u= u= u= u= u= u= Wih he seleced parameers (p=7, ransmilimi=3), he non-aggreged da case performs very badly for a very deep spanning ree, and performs as well as he u= aggreged da case for a very shallow ree The u= policy provides good compleeness; he u= policy provides near % compleeness In Auhorized licensed use limied o: The Library NUI Maynooh Downloaded on March 3, 29 2:9 from IEEE Xplore Resricions apply

4 Fig 4 he energy cos of each policy is shown (packes ransmied) 9 Repors Delivered Transmissions per repor delivered 2 d=5,w=3 d=7,w=2 d=256,w= d=3,w=6 d=4,w=4 d=,w=256 d=2,w=5 S/T params Figure 3 Effec of ree widh and deph %-age Repors Delivered p 7 Figure 5 Compleeness for varying p The significanly reduced energy usage, as measured by he number of packes ransmissions, allows he ransmission limi o be increased in he aggreged da case wihou a severe power cos (see Figure 6) This resuls in improved reliabiliy (compleeness), and reduced variabiliy The resuls of using differen policies o modify he ransmission limi for aggreged da are shown below d=3,w=6 d=2,w=5 d=,w=256 d=256,w= d=7,w=2 d=5,w=3 d=4,w=4 Figure 4 Energy efficiency of differen policies (for clariy, he policy axis is invered wr Fig 3) Noe h in all cases, excep for he single-hop spanning ree, he non-aggreged cos is significanly greer han he aggreged da cos As he policy gradually increases he ransmi limi, he energy cos rises slighly for improvemens in compleeness Fig 5 shows he effec of varying he probabiliy of packe loss beween nodes p on he compleeness When he packe loss re is small, hen he compleeness rio becomes independen of he policy Bu for larger packe loss res (smaller values of p), ypically associed wih wireless communicion, he qualiy of he response is significanly improved using aggregion wih increased ransmission limis Packes Transmied per Repor Received p Figure 6 Energy efficiency for varying p (e: invered p axis wr Fig 3) The resuls show h increasing he ransmission limi for aggreged da leads o increased reliabiliy, while losing lile of he energy gains The analysis shows h near % compleeness can be achieved using da aggregion and a weighed ransmission limi For a wide range of nework configurions and values of p ij, he policy u= seems o be he bes balance beween energy use and reliabiliy Auhorized licensed use limied o: The Library NUI Maynooh Downloaded on March 3, 29 2:9 from IEEE Xplore Resricions apply

5 5 Large numbers of reransmissions Large numbers of reransmissions lead o: increased lency of responses, and 2 increased ime o respond o a node failure These problems associed wih his can be addressed in a number of ways: o limiing he maximum value of reransmissions, o acceping he increased lency as he cos of o high-compleeness, or running spanning-ree mainenance in parallel wih re-ransmissions (using a hreshold less han he ransmi limi o iniie S/T repair) In pracice, here is lile poin in increasing he reransmi limi oo far: for example, if p=7, a ransmi limi of 8 will provide a packe success re wih reransmissions (q) of 9999% Fig 7 shows he effec of limiing he maximum number of ransmissions o parially address his issue A maximum ransmi limi of 8 was found o be effecive for a 25 node nework wih p=7% - his resul is valid excep for very deep neworks (>32 hops) The ransmi limi was se o a maximum of 64 (he larges value acually ained was 28) 9 8 rees were creed from he nework configurion by selecing links wih he required link qualiies ( p ) Duplice da is avoided by using he sandard Alernive Bi Proocol (ABP) This requires a minimum amoun of da o be kep per neighbour An alernive policy using 3-way handshaking was also examined: he cos of an exra packe (an ack-ack) per da packe, he memory requiremen is reduced o keeping deails on a single neighbour conversion any one ime The resuls are shown here, bu hey provide slighly lower efficiency and compleeness resuls in exchange for increased scalabiliy The simulion resuls also include da sreaming, where he same da value is sen muliple imes o he sink, as used for example in [] We include resuls wih & wihou local acknowledgemens and reransmissions Space does allow presenion of he full simulion resuls, so represenive resuls are shown in figures 9-2 Two differen spanning rees, buil offline, are used: one is a deep spanning ree (max deph=37, roo children=); he oher is a broad spanning ree, represenive of hose produced by algorihms such as [d] (max deph=8, roo children=3) In creing he spanning rees, ranges of link srenghs & noise levels were seleced o ry and mainain an average receive probabiliy of c 7%, based on prior measuremens of packe recepion res The base ransmission limi is se o 3 (allowing for 2 reransmissions) In he sreaming da cases, he number of da values o send is se o 3 Fig 8 shows he compleion res for he wo spanning rees used Compleeness Maximum Transmi Limi Figure 7 Compleeness vs policy for differen ransmi limi maxima 6 Simulion resuls The resuls achieved here were verified using he TinyOS2/TOSSIM simulor (See TinyOS 2 disribuion for deails, hp://wwwinyosne) wih a 225-node nework and he sandard 5-5-mediummica2-gridx nework configurion file provided The deails are crucial o hese resuls, as spanning %-age delivered none u= u=2 u= u= u= u= sr srack Deep Broad Figure 8 Compleion res for differen rees Noe h neiher ree is as exreme as hose shown for Numerical Analysis (Fig 3), so he resuls are closer o each oher Fig 9 shows he energy efficiency Noe he inclusion of non-aggreged, sreaming da The high number of duplices delivered reduces i efficiency significanly The aggreged policies all have much higher efficiencies (for an average p=7) Tree Auhorized licensed use limied o: The Library NUI Maynooh Downloaded on March 3, 29 2:9 from IEEE Xplore Resricions apply

6 Packes Transmied per Repor none u= u=2 u= u= u= u= sr srack Deep Broad Figure 9 Energy efficiency for differen rees (p=7) Figs and show he effec of varying he link qualiy (p) on he compleion re and energy efficiency of differen aggregion policies Larger uvalue policies achieve near % compleion for weaker links Tree Packes Transmied per Repor none u= u=2 Type u= u= u= u= sr sr-ack link qualiy 98 Figure effec of p on he energy efficiency (broad spanning ree, link qualiy axis invered) Fig 2 shows he impac of assering a maximum on he ransmi limi (u) for he broad spanning ree %-age delivered none u= u=2 u= u= u= u= sr sr-ack link qualiy p 5 Figure Effec of p on he compleion re (broad spanning ree) %-age delivered none u= u=2 u= u= u= u= max ransmi limi Figure 2 Effec of a maximum ransmi limi Aggreged da performs much beer from an energy viewpoin Noe h he u= policy loses highly aggreged packes low link qualiy values (Fig ) Sreaming da wihou acknowledgemens leads o a low compleion re a large cos Sreaming wih acknowledgemens provides a higher compleion re, bu a cos due o duplicion of repors o he roo In summary, as shown in boh he numerical analysis, and he simulion resuls, da aggregion can significanly reduce he power used in reporing collecive, nework-wide sus o he roo of a spanning ree (especially for nodes nearer he roo) The reliabiliy, as measured by he compleeness, is also significanly improved by using da aggregion and an increased reransmission limi for aggreged da packes Da sreaming (wih local acknowledgemens) can increase he compleion re, bu a significan energy cos, and resuling in he delivery of many duplices o he roo ( shown here) Increasing he ransmission limi for nonaggreged da improves he compleion re somewh, bu again a very significan energy cos The improvemens provided by aggregion are greer for deeper spanning rees Auhorized licensed use limied o: The Library NUI Maynooh Downloaded on March 3, 29 2:9 from IEEE Xplore Resricions apply

7 2 Conclusion The compleeness of nework-wide, managemen responses in wireless sensor neworks can be significanly improved by aggreging all he responses, and using reransmissions wih an aggregion-proporional ransmission limi This resuls in a high probabiliy of near-% compleeness of response, while sill realizing significan power savings when compared o he nonaggreged case Sreaming he same da value muliple imes wihou aggregion is paricularly inefficien, as he source sends duplice da values even afer successful recepion he roo Increasing he ransmission limi can lead o increased delay in reacing o node failures This can be limied by seing a maximum value for he ransmi limi Effecive values for he limi depend on he recepion success res for each link in he spanning ree; his emphasizes he imporance of using link qualiy merics when building he ree [] Addiion benefis can be achieved by using da aggregion and aggregion-dependen ransmi limis The reducion in da raffic means h he nework managemen operion will have a reduced impac on normal WSN da collecion aciviies Also, hough explored here, he reducion in packes provides an opporuniy o reduce he ime required o complee he collecion of nework-wide sus da Fuure work is o explore dynamic spanning ree generion policies and deermine heir impac on he compleeness of he aggreged response We also plan o implemen he da aggregion approaches examined here in a real-world sensor nework, and valide he compleeness and energy resuls 3 References [] S Brown and CJ Sreenan, A New Model for Upding Sofware in Wireless Sensor Neworks, IEEE Nework, IEEE, Nov-Dec 26, pp [2] C-C Han, R Kumar, R Shea, M Shrivasava, Sensor nework sofware upde managemen: a survey, Inernional Journal of Nework Managemen, Wiley, July/Aug 25, pp [3] J Heidemann, F Silva, C Inanagonwiw, R Govindan, D Esrin, and D Ganesan, Building efficien wireless sensor neworks wih low-level naming, in Proc 8 h ACM Symposium on Opering Sysems Principles, ACM, 2, pp [4] WR Heinzelman, J Kulik, and H Balakrishnanm, Adapive Proocols for Informion Disseminion in Wireless Sensor Neworks, in Proc 5h annual ACM/IEEE Inl Conf on Mobile compuing and neworking, ACM, 999, pp [5] WR Heinzelman, A Chandrakasan, and H Balakrishnan, Energy-Efficien Communicion Proocol forwireless Microsensor Neworks, in Proc 33rd Annual Hawaii Inl Conf on Sysem Sciences, IEEE, 2, pp vol 2 [6] T Pham, EJ Kim, and M Moh, On Da Aggregion Qualiy and Energy Efficiency of Wireless Sensor Neworks Proocols Exended Summary, in Proc s Inl Conf on Broadband Neworks, IEEE, 24, pp [7] C Florens, and R McEliece, Packes disribuion algorihms for sensor neworks, in Proc 22 nd Annual Join Conf of he IEEE Compuer and Communicions Socieies, IEEE, 23, pp vol 2 [8] C Inanagonwiw, D Esrin, R Govindan and J Heidemann, Impac of Nework Densiy on Da Aggregion in Wireless Sensor Neworks, in Proc Proceedings of he 22 nd Inl Conf on Disribued Compuing Sysems, IEEE, 22, pp [9] C Inanagonwiw, R Govindan and D Esrin, Direced diffusion: scalable and robus communicion paradigm for sensor neworks, in Proc 6h Annual Inl Conf on Mobile Compuing and Neworking, ACM, 2, pp [] G Tolle and D Coller, Design of an Applicion- Cooperive Managemen Sysem for Wireless Sensor Neworks, in Proc 2 nd European Conf On Wireless Sensor Neworks, IEEE, 25, pp 2-32 [] H Hassanein and J Luo, Reliable Energy Aware Rouing in Wireless Sensor Neworks, in Proc 2 nd IEEE Workshop in Dependabiliy and Securiy in Sensor Neworks and Sysems, IEEE, 26, pp [2] S Madden, MJ Franklin, and MJ Hellersein, TAG: a Tiny AGgregion Service for Ad-Hoc Sensor Neworks, in Proc 5 h Symposium on Opering Sysems Design and Implemenion, ACM SIGOPS Opering Sysems Review, ACM, Dec 22 (36/SI), pp 3-46 [3] N Shrivasava, C Buragohain, D Agrawal and S Suri, Medians and Beyond: New Aggregion Techniques for Sensor Neworks, in Proc 2nd Inl Conf on Embedded neworked sensor sysems, ACM, 24, pp [4] C-Y Wan, AT Campbell, L Krishnamurhy, PSFQ: A Reliable Transpor Proocol for Wireless Sensor Neworks, in Proc s Inl workshop on Wireless sensor neworks and applicions, ACM, 22, pp - [5] JP Benson, U Roedig, A Barrosso and SJ Sreenan, On he Effecs of Aggregion on Reliabiliy in Sensor Neworks, in Proc 65 h Vehicular Technology Conference, IEEE, 27, pp Auhorized licensed use limied o: The Library NUI Maynooh Downloaded on March 3, 29 2:9 from IEEE Xplore Resricions apply