service <strong>of</strong> Carrier 1 using a special port number), and Other(all other port numbers). <strong>The</strong>y contribute to 78.8%, 15.1%,0.2%, and 5.9% <strong>of</strong> the total traffic volume, respectively.We use a threshold <strong>of</strong> 10 sec <strong>of</strong> idle time to decide thata connection has terminated. Changing this value does notqualitatively affect the simulation results.For simplicity, we assume that there are four applications,each involving one traffic type, running on smartphones. ForWeb, Email, and Sync applications, a transfer is defined asconsecutive connections <strong>of</strong> the same traffic type whose interconnectiontime (the interval between the last packet <strong>of</strong> oneconnection and the first packet <strong>of</strong> the next connection) isless than 1 sec. Note that a transfer may consist <strong>of</strong> multipleconnections and connections may overlap (e.g., concurrentconnections supported by smartphone browsers). Atthe end <strong>of</strong> each transfer, each application independently callsTerminateTail with probability <strong>of</strong> z, which quantifies theapplicability <strong>of</strong> TOP, to perform binary predictions (whetherITT is greater than TT, see §VI-B) with accuracy <strong>of</strong> w.In other words, the probabilities <strong>of</strong> a correct prediction, anincorrect prediction, and no prediction (TerminateTail isnot invoked) are zw, z(1−w), and 1−z, respectively. Sincebinary predictions are performed, each application uses an ITT<strong>of</strong> 60 sec if the ITT is predicted to be greater than TT. Varyingthis from 30 sec to infinity, or using the exact prediction value<strong>of</strong> ITT changes the results in Figure 7 by no more than 0.01.We assume that the “Other” application is unaware <strong>of</strong> TOP.Figure 7 plots the impact <strong>of</strong> TOP on ∆E, ∆S, ∆D T , and∆D by varying w and z for Carrier 1. In each plot, thez = 0 curve is a horizontal line at y = 0 correspondingto the comparison baseline i.e., the default case where TOPor fast dormancy is not used. Figure 7 clearly shows that,increasing z, the applicability <strong>of</strong> TOP, brings more savingsat the cost <strong>of</strong> increasing the state promotion delay. On theother hand, increasing w, the prediction accuracy, not onlybenefits resource savings but also reduces the state promotionoverhead. Under the case where z = 0.8 and w = 90%,TOP saves the overall radio energy E, the DCH tail timeD T , and the total DCH time D by 17.4%, 55.5%, and 11.7%,respectively, with the state promotion delay S increasing by14.8%. <strong>The</strong> results for Carrier 2 show similar trends. Underthe condition <strong>of</strong> z = 0.8 and w = 90%, TOP can save E,D T , and D by 14.9%, 60.1%, and 14.3%, respectively withS increasing by 9.0%.We compare TOP with other schemes for saving the tailtime. In each plot <strong>of</strong> Figure 8, the X axis is the state promotiondelay ∆S, and the Y axis corresponds to saved resources(∆E, ∆D T , or ∆D) for Carrier 2. A more downward orleftward curve indicates a better saving scheme since given afixed ∆S, we prefer a more negative value <strong>of</strong> ∆E, ∆D T , or∆D indicating higher resource savings. Each plot <strong>of</strong> Figure 8contains four curves. <strong>The</strong> “TOP” curve corresponds to usingTOP with w = 80% and z being varied from 0.5 to 1.0.<strong>The</strong> “FD” (fast dormancy) curve is generated using the sameparameters, but in the “FD” scheme, applications use fastdormancy without being scheduled by TOP. In other words,an application (Web, Email, or Sync) sends a T messagewhenever its predicted ITT is greater than TT. <strong>The</strong> “timer”curve corresponds to a strategy <strong>of</strong> proportionally decreasingα and β timers that affect all sessions in the trace.<strong>The</strong> “TE” curve denotes employing TailEnder [6] to save energyand radio resources. As described in §I, for delay-tolerantapplications, their data transfers can be delayed and batchedto reduce the tail time. TailEnder is a scheduling algorithmthat schedules transfers to minimize the energy consumptionwhile meeting user-specified deadlines by delaying transfersand transmitting them together. <strong>The</strong> TailEnder algorithm wasimplemented in our simulator using the default parametersdescribed in [6]. We apply TailEnder on all Email and Synctransfers and vary the deadline (the maximally tolerated delay)from 0 to 5 minutes. A longer deadline can potentially savemore resources but a user has to wait for longer time.We discuss the results in Figure 8. TOP outperforms fastdormancy(FD), whose curve lies on the right <strong>of</strong> the “TOP”curve. To achieve the same savings in D, E, and D T , thestate promotion delay <strong>of</strong> TOP is always less than that <strong>of</strong>FD by 10% <strong>of</strong> the overall promotion delay in the defaultscheme. Further, reducing inactivity timers incurs additionalstate promotions, overwhelming the savings <strong>of</strong> D and E. <strong>The</strong>fundamental reason for this is the static nature <strong>of</strong> the inactivitytimer paradigm where all packets experience the same timeoutperiod. We also notice that TailEnder can reduce the overallstate promotion delay (as indicated by the negative∆S values)due to its batching strategy. However, its applicability is verylimited, yielding much less savings, and it incurs additionalwaiting time for users. <strong>The</strong> comparison results for Carrier 1is qualitatively similar, implying that invoking fast dormancywith a reasonable prediction accuracy (around 80%) surpassesthe traditional approach <strong>of</strong> tuning inactivity timers in balancingthe trade<strong>of</strong>f, and TOP’s coordination algorithm effectivelyreduces the state promotion overhead caused by concurrentnetwork activities.B. Evaluation using locally collected tracesWe perform case studies <strong>of</strong> two applications (Pandorastreaming and Web browsing) using traces locally collectedfrom an Android G2 phone using Carrier 2’s UMTS network.We investigate each application separately without injectingconcurrent traffic, then apply the coordination algorithm onthe aggregated traffic <strong>of</strong> both applications.1) Pandora radio streaming: Pandora [1] is an Internetradio application. We collected a 30-min trace using Tcpdumpby logging onto one author’s Pandora account, selecting a predefinedradio station, then listening to seven tracks (songs).By analyzing the trace, we found that the Pandora trafficconsists <strong>of</strong> two components: the audio/control traffic and theadvertisement traffic. Before a track is over, the content <strong>of</strong>the next track is transferred in one burst utilizing the maximalbandwidth. <strong>The</strong>n at the exact moment <strong>of</strong> switching to the nexttrack, a small traffic burst <strong>of</strong> control messages is generated.<strong>The</strong> second component is periodical advertisement traffic froman Amazon EC2 server for every one minute. Each such burstcan trigger an IDLE→DCH promotion.
the <strong>Cichlid</strong> <strong>Fishes</strong> <strong>of</strong> <strong>Lake</strong> Nyma.bars, if present, very faint ; interorbital width 3 to 44 inlength <strong>of</strong> head, diameter <strong>of</strong> eye 34 to 44 (in specimens <strong>of</strong>70 to 165 mm.) ; D. XVI-XVIII 9-11 ; A. I11 7-9 ;caudal truncate .............. 4. williamsi.2. Tooth-band <strong>of</strong> lower jaw transverse ; body uniformly brownish,or with traces <strong>of</strong> vertical bars and <strong>of</strong> two longitudinalbands; interorbital width 2$ to 33 in length <strong>of</strong> head,diameter <strong>of</strong> eye 34 to 4 (in specimens <strong>of</strong> 90 to 140 mm.) ;D. XVII-XVIII 9-10 ; A. I11 7-9 . 6. Iucerna.B. Dcpth <strong>of</strong> body 3 to 33 in the length ; tooth-band <strong>of</strong> lower jawrounded.I. Colour uniformly dark ; interorbital width 4# in length <strong>of</strong>head ; D. XVIII 9 ; A. I11 8 .... 6. fuscus.2. A longitudinal stripe along middle <strong>of</strong> side, another above upperlateral line, black on gold in female, gold on dark brownishin male ; felnale with a black submarginal band on dorsalfin, male with dorsal pale; interorbital width 34 to 4 inlength <strong>of</strong> head ; D. XVIIJXIX 8-9. 7. auratus.11. Snout shorter than or as long as diameter <strong>of</strong> eye.A. Upper pr<strong>of</strong>ile gently sloping, occipital crest well developed ;3 or 4 (rarely 5) series <strong>of</strong> scales on cheek, 5 (rarely 6) from origin<strong>of</strong> dorsal to lateral line ; D. XV-XVIII 9-10.8. novemfa&atus.B. Pr<strong>of</strong>ile <strong>of</strong> snout steeply descending, occipital crest weak.3 to 6 series <strong>of</strong> scales on the cheek.1. 6 to 8 scales from origin <strong>of</strong> dorsal to lateral line; D. XVI-XVIII 8-10 ; A. I11 7-8.a. Width <strong>of</strong> tooth-band <strong>of</strong> lower jaw leas than) length <strong>of</strong> head ;. . 9. microstma.tooth-band slightly roundedb. Width <strong>of</strong> tooth-band <strong>of</strong> lower jaw ) length <strong>of</strong> head or(uaually) more ; tooth-band transverse.10. tropheops.Diameter <strong>of</strong> eye 3 to 39 in length <strong>of</strong> head, inter-orbital width 23 to 34 (in fishes <strong>of</strong> 80 to[tropheops.120 mrn.) ............................ 10 a. tropheopsDiameter <strong>of</strong> eye 3 to 3i in length <strong>of</strong> head, interorbitalwidth 3 to 38 (in fishes <strong>of</strong> 76 to112 mm.) ............................ 10 b. tropheops gracilior.2. 6 to 7 scales from origin <strong>of</strong> dorsal to lateral line ; D. XVII-XIX 8-10 ; A. I11 8-9 ; diameter <strong>of</strong> eye 24 to 34 in length<strong>of</strong> head, interorbital width 24 to 3f (in fishes <strong>of</strong> 76 to115 mm.) ...................... 11. macrophthalmus.1. Pseudotropheus eleqans, sp. n.A singlc specimen, 110 mm. in total length, from Deep Bay(coll. Christy).This species resembles P. livingstonii in the relatively narrowand rounded band <strong>of</strong> teeth and in the cmarginate caudal.<strong>The</strong> lnrgcr eye and somewhat narrower przcorbital distinguisllit from 11. livinqstonii ; also it has weaker jaws, the lowercontained three times in the length <strong>of</strong> head. <strong>The</strong> dentigerousarea <strong>of</strong> the lower pharyngeal is subtriangular instcad <strong>of</strong>almost heart-shaped, as it is in other species <strong>of</strong> Pseudotropheus.