1: %file packetg_2c.tex V.8/7/02
2: %\documentstyle[preprint,aps,psfig,epsfig]{revtex}
3: \documentstyle[aps,prl,multicol,epsf]{revtex}
4: %\tighten
5:
6: \begin{document}
7: \def\be{\begin{equation}}
8: \def\ee{\end{equation}}
9: \def\bc{\begin{center}}
10: \def\ec{\end{center}}
11: \def\bea{\begin{eqnarray}}
12: \def\eea{\end{eqnarray}}
13: \draft
14: %\twocolumn[\hsize\textwidth\columnwidth\hsize\csname
15: %@twocolumnfalse\endcsname
16: \title{Packet Transport on Scale Free Networks}
17: \author{Bosiljka Tadi\'{c}$^1$ and G.J.Rodgers$^2$}
18:
19: \address{$^1$Jo\u{z}ef Stefan Institute, P.O. Box 3000, 1001 Ljubljana,
20: Slovenia}
21: \address{$^2$Department of Mathematical Sciences, Brunel University,}
22: \address{Uxbridge, Middlesex UB8 3PH, U.K.}
23: \maketitle \thispagestyle{empty}
24: \begin{abstract}
25: We introduce a model of information packet transport
26: on networks in which the packets are posted by a given rate and
27: move in parallel according to a local search algorithm.
28: By performing a number of simulations
29: we investigate the major kinetic properties of the transport as a function
30: of the network geometry, the packet input rate and the buffer size.
31: We find long-range correlations in the power spectra of arriving packet
32: density and the network's activity bursts. The packet transit time
33: distribution shows a power-law dependence with average transit time
34: increasing with network size. This implies dynamic queueing
35: on the network, in which many interacting queues are mutually driven by
36: temporally correlated packet streams.
37:
38: {\it Keywords:} Internet, traffic, information packet
39: \end{abstract}
40: \pacs{PACS numbers: 02.50.cw, 05.40.-a, 89.75Hc.}
41: %]
42: %\narrowtext
43: %\vskip2pc]
44: \begin{multicols}{2}
45:
46: \section{Introduction}
47:
48: The Internet has become a central feature of all our lives.
49: This has lead to interest in
50: information packet transport on massive, heterogeneous, random,
51: network geometries.
52:
53: Many empirical studies of packet transport on the Internet have been
54: carried out \cite{csabai,t1,chong,erramilli}. The principal
55: conclusion of these measurements is that the aggregate
56: packet streams are fractal, obeying long-range correlated in time.
57: Of particular interest are
58: studies of packet density at a particular link (or on a node) and
59: ping time statistics \cite{t1,chong}, in which the round trip time,
60: or the time it
61: takes a packet to travel to a destination and back to its source,
62: are measured.
63: Analysis of the power spectrum of the packet density and
64: round trip times allows one to distinguish two regimes with {\it free flow}
65: and {\it jammed} traffic, respectively, depending on the traffic intensity.
66: In \cite{erramilli} it was shown that the power-law
67: behaviour of the distribution of packet inter-arrival times
68: has a significant effect on packet queuing performance, and
69: consequently on the overall packet transport.
70:
71:
72: These dynamical transport and queuing processes are taking
73: place on the Internet, the network made up of
74: routers and computers as vertices and cables as edges.
75: Recent studies of the geometry of the Internet,
76: \cite{fal,gov,Vazquez}, indicate that the
77: degree distribution has a power-law, scale free behaviour
78: \cite{ba,DMS,BT,KRR}.
79:
80: Another topological property, the betweenness, or the total number of
81: shortest paths going through each node \cite{load,newman},
82: was also found to have a power-law
83: distribution with an exponent $\approx 2.2$ on a scale free graph.
84: In terms of transport processes,
85: this corresponds to the distribution of the number of packets
86: transferred through a node (in a long-time limit),
87: in a system in which the packet transport is
88: dominated by non-interacting packets that always
89: take the shortest route between source and destination.
90:
91: In reality, the information packet transport is much more complex for
92: the following reasons: (1) no global navigation is technically available;
93: (2) many packets are being transported simultaneously, hindering each others
94: motion, depending on the search algorithm and network structure.
95: In particular, the kinetics of many interacting packets leads to a
96: qualitatively new feature, which is manifested in the {\it queueing}
97: \cite{book} of
98: packets at individual nodes. In the network we have the problem of many
99: interacting queues.
100: Thus a close relationship seems to exist between the kinetic properties of
101: the packet transport, such as the distribution of round trip times, or
102: the number of packets on a node, and the network on
103: which the kinetics takes place. The interplay between network structure
104: and packet kinetics becomes particularly important when
105: one imagines the transport in a network with buffers, which restrict the
106: length of packet queue at each node, or with reduced ability of nodes
107: to handle the packets.
108: In recent studies of traffic on idealized geometries---on a linear chain
109: \cite{micro} and on Cayley trees \cite{Arenas,Guimera} and one- and
110: two-dimensional lattices \cite{Guimera} a sharp transition from a free to
111: congested traffic was found.
112: The smallest buffer along the chain causes jamming in the linear model
113: \cite{micro}, while transport through the node at the top of hierarchy is
114: crucial for the onset of congestion on the hierarchical lattice
115: \cite{Arenas,Guimera}.
116: In a network with scale-free structure of links the multiplicity of
117: potential paths among pairs of nodes may alter the role of buffers in
118: the crossover to jammed traffic. It is expected that
119: the local geometry of the hub nodes, and their buffer sizes,
120: is of critical importance, while the buffer sizes of the
121: majority of other nodes are largely irrelevant.
122: The relative importance of nodes,
123: of course, depends on the search algorithm being used.
124: For practical reasons,
125: search algorithms can only be applied to small sections of the network.
126: An attempt was reported recently \cite{Albert} to find optimal topology of
127: the network for a given local algorithm in the presence of congestion.
128:
129: In this paper we introduce a simple model of packet transport on
130: networks in an attempt to understand which of the elements present in a
131: real system are necessary for the system to display the observed empirical
132: results. In particular, we study the effect of the
133: geometry of the network on the major
134: properties of the packet transport. We develop a model for simultaneous
135: transport of packets
136: and implement this model on scale free and randomly grown networks.
137: Packets are posted by a given input rate $R$ and moved according to
138: a local algorithm which uses up to next-neighbour search
139: and the local geometry. In this way the packet queue at each individual
140: node is dynamically formed by packets moving from neighbouring
141: nodes. We measure a number of kinetic quantities that allow us to
142: characterize the nature of the packet transport process on the two networks.
143:
144: The paper is organized as follows: In Section II we introduce the model by
145: first presenting the algorithm of graph growth and then the search algorithm
146: for packet kinetics on the graphs. In Section III we define precisely the
147: quantities that we determine in the simulations and present the results
148: for the packet time series and power spectra of these quantities.
149: In Section IV we study the distribution of transit times of packets,
150: the queue lengths and the networks' output rate.
151: In Section V we present a short summary of the results and
152: discuss the main conclusions.
153:
154: %***********************************************************
155: \section{The Model}
156:
157: Our model is developed in two separate stages.
158: Firstly the network is grown and then
159: we simulate the motion of packets on the network.
160:
161: {\bf Network Structure}
162:
163: We consider two different tree networks, one of which is scale-free and
164: the second is a grown random network.
165:
166: The {\bf scale free} network (SFN) structure \cite{ba} is grown as follows.
167: At each time step one new node is added to the network and is linked to
168: a node of degree $k$ with a rate \cite{BT}
169: \begin{equation}
170: p(k,t)= \frac{k+\alpha}{D(t)+\alpha N(t)}.
171: \end{equation}
172: Here $D(t)$ is the total degree of the network at time $t$ and
173: $N(t)$ is the total number of nodes in the network, and $\alpha >0$
174: is a tunable parameter. For networks grown in this way, the
175: degree distribution is power-law \cite{ba,BT,DMS,KRR}. More precisely,
176: the number
177: of nodes with degree $k$, $P(k)$, behaves as $P(k) \sim k^{-\tau}$ with
178: $\tau =2+\alpha$.
179: Choosing $\alpha=0.2$ means that this network has
180: the same in-degree distribution \cite{fal} as that observed on the Internet.
181:
182: The {\bf random network} (RGN) is grown by adding one new node at each
183: time step
184: and connecting it to a node of degree $k$ with rate $1/N(t)$, independent of
185: $k$. In this case the degree distribution $P(k)$ behaves as $P(k)\sim 2^{-k}$.
186:
187: {\bf Sending Packets}
188:
189: At each time step, with probability $R$, a new information
190: packet is initiated.
191: This is done by randomly selecting a source node and a
192: destination node for that packet.
193:
194:
195: {\bf Packet Transport}
196:
197: At each time step, each node is investigated in turn, and if it has
198: a packet on it then the top packet on the node attempts to move.
199: This is done by searching the nearest and next nearest neighbours
200: of the node for the destination node of that packet. If found, the packet is
201: moved to or towards the its destination node.
202: If the destination node isn't
203: found, the packet moves to a randomly selected neighbour.
204: Each node has a buffer size of integer $B$, the maximum
205: number of packets that can be on a node at any one time.
206: If a packet attempts to move to a full node, it is unable to stay there
207: and moves back to the node it came from.
208: We assume that every node has the same buffer size.
209: Similarly, in order to fully investigate the
210: effect of the geometry of the graph, we assume that all the links have
211: the same capacity. When the packet reaches its destination node, it is
212: removed from the network.
213:
214: In terms of queuing theory \cite{book},
215: our nodes can be thought of as single server
216: queues, with buffer size $B$, and a deterministic service time distribution.
217: The service discipline is last in first out (LIFO). For single server
218: queues with this discipline some
219: approximations have been developed \cite{abate} for the waiting time
220: distribution.
221: Finally the inter-arrival time distribution is not put in by hand, but
222: arises naturally as a function of
223: the network geometry and the kinetics of the packets at neighbouring nodes
224: on the network.
225: The waiting time distribution has two different contributions: First,
226: a node receives packets with a Poisson distribution
227: as a result of packets being initiated on that node; and, second,
228: the node receives packets from neighbouring nodes
229: with a time series that turns out to have a long range, power-law
230: temporal correlation.
231:
232:
233: Thus we have packet transport rules which yield paths for the packet
234: which are close to, but longer than, the minimum path between source and
235: target node. We also have networks which are trees containing no loops.
236: We do not believe that the loops present in real networks \cite{fal,BT,KRR}
237: play an important role in the observed behavior
238: \cite{csabai,t1,chong,erramilli}.
239:
240:
241:
242:
243: \section{Traffic Time Series and Power Spectra}
244:
245: We have performed a number of simulations of the kinetics of the packet
246: transport on both the scale free network and the random grown network of size
247: $N=10^3$ nodes.
248: We considered a number of different quantities that characterise the
249: kinetics and allow comparison to be made
250: with real data from the Internet \cite{csabai,t1}. These were
251:
252: (a) Activity, $a(t)$,
253: the total number of nodes with a packet on at time $t$.
254: This gives a measure of the fraction of the network that is busy.
255:
256: (b) Total Load, $n(t)$, the total number of packets
257: in the network at time $t$.
258:
259: (c) Load (or queue length), $q(t)$, the number of packets
260: on a node at time $t$.
261:
262: (d) Load on the active nodes at time $t$, $l_a(t)= \sum _{active}q(t)$.
263: This quantifies the load carried by the active nodes at a particular time step.
264:
265: (e) Density, $\rho(t)$,
266: the total number of packets arriving at the hub node at time $t$.
267:
268: (f) Transit time, $T_{tr}$, the time taken by a packet to reach
269: its destination.
270:
271: The characteristics of these quantities change as a function of the geometry
272: of the network, the buffer size $B$ and the input rate $R$. In order to
273: examine these changes, we calculated their power-spectra. In general,
274: the power spectrum of a time series $X(t)$ is defined by
275: \begin{equation}
276: S_X(f)=|\sum_{t=1}^{\infty}X(t)e^{ift}|^2.
277: \end{equation}
278: For a time series with no
279: temporal correlation, the plot of $S_X(f)$ against
280: $f$ is independent of $f$.
281: If the time series $X(t)$ has an auto-correlation function
282: $r_X(k)\sim k^{-\phi}$ then $S_X(f)\sim f^{-\phi}$.
283:
284: The output rate, or average number of packets leaving the
285: system per time step, is, along with the transit time $T_{tr}$, a
286: measure of the efficiency of the system. It is defined by
287: \begin{equation}
288: \mu = R-\frac{n(t)}{t}.
289: \end{equation}
290: In general the value of the ratio $\mu /R$ allows one to determine
291: \cite{Whitt} whether the queues in the system are getting longer
292: with time. As it will be shown below, in our model, the queues
293: are growing with time for all values of the parameters.
294:
295:
296:
297:
298: A comparison of the packet traffic on the scale-free network and the
299: randomly grown network in our simulations is
300: demonstrated by time series for load carried by active nodes
301: and by number of nodes that are active at given time shown in Fig.\ 1.
302: With identical driving conditions,
303: the random network tends to distribute the
304: activity over many nodes simultaneously, whereas in the scale-free
305: network the activity is localised to the hub, which carries most of
306: the packets, and a few other high degree nodes in the centre of the network.
307:
308:
309:
310: In Figure 2 we compare the power spectra of the time series for activity,
311: load at active nodes, density at the hub, and number of packets in the network
312: for the two network geometries operating in the identical driving conditions
313: of low input rate $R=0.01$ and buffer size $B=100$ for network size
314: $N=1000$ nodes.
315: For these driving conditions the long-range correlations in the time series
316: are present in certain range of frequencies in both networks. The slopes are
317: generally closer to -1 in the case of SFN ($1/f$-noise), whereas in the RGN
318: we measure the slope close to -2, except for the density where it is close
319: to -0.5, and a whiter spectrum (c.f. Fig.\ 2).
320: These features, however, change with increased input rate and/or decreased
321: buffer size, suggesting that the character of the transport
322: depends on the relative ratio of these parameters.
323:
324: In particular, when the input rate is increased by a factor of four and the
325: buffer size is unchanged ($B=100$, $R=0.04$) the power spectrum of the
326: density at the hub in SFN behaves as $1/f^2$ (top line in Fig.\ 3).
327: In the case of RGN the slope becomes approximately -1 and the range
328: of frequencies where the correlation occurs is reduced.
329: Changes in the spectrum of density indicates that the character of packet
330: transport at the hub changed from $1/f$ noise in the free-flow regime,
331: to simple diffusive transport.
332:
333: In Figure 3 we also show
334: the power spectrum of the density measured at a node away from the hub
335: in the free-flow regime for both network geometries. It is interesting
336: to note that, in the SFN, in contrast to the density at hub, the spectrum
337: inside the network for low traffic intensity is a white noise
338: (bottom curve in Fig.\ 3). In the
339: RGN, however, the correlations similar to the ones at the hub are measured,
340: once again supporting the conclusion that the RGN tends to distribute
341: the activity over the entire graph.
342:
343:
344:
345:
346:
347: \section{ Packet Transit Times and Queueing Properties}
348:
349:
350: To further investigate how the network performs under dense packet traffic
351: we study transit times for individual packets. Within our algorithm we
352: mark the first $10^3$ packets with an additional time label and monitor the
353: time that they spend on the network
354: from posting until arrival at their respective destinations. Packets are
355: still posted probabilistically with rate $R$. In the inset to Fig.\ 4 an
356: example of the time series of transit times against posting time
357: of the packets is shown. It is noticeable that
358: the duration of the journey for each packet
359: is different, both because the difference in distances that they make on
360: the graph and because of the time that they spend waiting in queues at
361: particular nodes.
362: The distribution of transit times is given in main part of Fig.\ 4 both
363: for SFN and RGN for low packet density ($R=0.01$, $B=100$).
364:
365: For the range of values of the transit times $T_{tr}$ we find
366: power-law distributions, but with different slopes, indicating the
367: difference in the efficiency of the two network structures.
368: Namely, $\tau_T=1.25 \pm 0.02$, for SFN, and $\tau_T=0.53\pm 0.02$, for RGN.
369: In addition, the transit times larger than approximately $10^3$ time steps
370: for SFN, and $10^4$ in the case of RGN, contribute to tail of the
371: distribution.
372: Among these are packets that become buried deep in queues at certain
373: nodes in the network. Thus, it is the dynamics of queueing, in addition to
374: the network's structure, that determines the network's transport efficiency.
375: Next we study some properties of queues in both networks.
376:
377: In Figure 5 we show the probability distributions of the queue lengths
378: (loads) measured at individual nodes throughout the network {\it after}
379: each time step is
380: completed. Apart from the cut-off which is determined by the maximum
381: buffer size (here $10^3$), there is a larger
382: probability to find a queue of given length $q$ in the RGN compared to the SFN,
383: in agreement with generally lower efficiency of the randomly grown network.
384: The relative appearance
385: of large queue lengths in the RGN decays with the power-law exponent
386: $\tau _q=1.4 \pm 0.02$, and with $\tau_q = 0.48 \pm 0.02$ in the SFN.
387: In both cases there is a part of the network in which nodes carry a small
388: load, with a similar power-law decay.
389:
390:
391:
392: Compared to standard single-queue theories \cite{book}, here we have
393: many interacting queues in which packets hop from one queue to the other along
394: the edges of the graph, as they are directed by the search algorithm.
395: Thus the input rate of packets at each node is different and related to the
396: local connectivity of that node and number of paths that the actual search
397: algorithm selects to pass through that node at a single time step.
398: (As already mentioned in the introduction, we keep the output rate of one
399: packet per time step equal at each node, in order to single out the effects
400: of networks structure more clearly.)
401: Therefore, in the SFN, the kinetics of the queues at the nodes with high
402: connectivity are of primary importance, whereas queues at the
403: majority of other nodes in the network rarely exceed a few packets.
404: This is due to the topology of SFN, in which cluster of nodes linked to
405: the hub increases faster than any other subgraph,
406: and also due to the search algorithm we are using, which is suitable for that
407: topology. In the RGN the same search algorithm is much
408: less effective, because the dominant cluster grows only logarithmically
409: in time making the edges more evenly distributed throughout the network.
410:
411: Thus, the input rate at the hub varies in time and it is precisely given by
412: the density $\rho (t)$ that we determined in Section\ III. In the SFN
413: $\rho (t)$ was found to exhibit long-range correlations of $1/f$ type
414: for low intensity of traffic (cf. Fig.\ 2), that changes to a short-range
415: $1/f^2$ correlations when traffic intensity increases. In the RGN within
416: the same conditions, for comparison, the spectrum is much closer to white
417: noise. Next we investigate how the queueing dynamics at the hub affects
418: performance of the network.
419:
420: We determine the network's output rate $\mu$, defined as the number of
421: packets arriving to their destination per time step. Therefore, the
422: difference of the input and output rate $R-\mu = n(t)/t$ determines the
423: workload \cite{Whitt} on the level of the entire network.
424: In main part of Fig.\ 6 we show the workload
425: in SFN for two input rates $R$ and fixed buffer size $B=100$.
426: For large input rate
427: (top curve) $R-\mu$ saturates for large times $t$ at a finite average value,
428: indicating that total load in the network $n(t) =(R-\mu)t$ increases
429: {\it linearly} in time. In this case the jamming that first occurs at
430: the hub seems to spread gradually throughout network.
431: For comparison, in the case of low input rate (bottom curve) the
432: workload $R-\mu$, after a sharp initial increase, decays
433: approximately as $\sim t^{-1/2}$, indicating that total load $n(t)$
434: increases {\it sub-linearly} in this flow regime.
435: Taking the average asymptotic value of $\mu$, in the inset to Fig.\ 6 we
436: show the ratio of the network's output and input rates $\mu /R$
437: for the SFN and the RGN, which, once again shows the systematically better
438: performance of the scale-free structure for a range of input rates.
439:
440:
441: \section{Conclusions and Discussion}
442:
443: In this paper we have introduced a new model of packet transport on
444: networks and investigated its major properties on
445: both randomly grown and scale free networks.
446: The model incorporates in a natural way three basic elements---network
447: topology, search algorithm, and queueing dynamics---which determine
448: overall efficiency of the transport. By fixing the input rate and using the
449: near and next near neighbour search algorithm,
450: which is fairly efficient on structured graphs, in this work we
451: focused on the effects of networks' topology on the queueing dynamics.
452: For this purpose we also keep link capacities at a minimal one packet per
453: time step.
454:
455:
456: We found long-range temporal correlations for many of the important
457: quantities within the system, in particular for the density of arriving
458: packet streams, with
459: non-universal exponents that were dependent on the traffic intensity and
460: the network on which the transport was taking place.
461: The power-law dependences of the transit time distribution with
462: the exponent $\tau_T < 2$ implies that the average time that a packet
463: spends on the network increases with network size, causing queueing
464: processes on the network. We find the power-law behavior
465: of the queue length distribution reflects the networks' structural
466: efficiency in the transport.
467: The character of the queueing dynamics in our model is dominated by
468: queues at highly connected nodes, which are driven by fractal packet
469: streams from neighbouring nodes. Measured by the single-queue criteria,
470: the queues in our model are predetermined to increase with time.
471: However, measured on the level of entire network, we find that the networks'
472: workload may increase sub-linearly for low traffic intensity, and linearly
473: when traffic approaches the congested regime.
474: At the same time the observed long-range correlations
475: and power-law dependences of the distributions are different in two
476: traffic regimes, a transition seems to take place along a line $R_c(B)$.
477: A detailed analysis of this transition was not included in this work.
478:
479: Finally we found that the scale free network was much more
480: efficient, in terms of network output to
481: input ratio, than the randomly grown network. Allied to
482: this was the observation that transport on the scale free network was
483: very dependent on the queuing at a few highly connected nodes, and many of
484: the low-connectivity nodes had very low activity. In contrast,
485: the randomly grown network distributed the activity much more
486: evenly over the graph.
487:
488: This work is an initial study, in future we intend to investigate a
489: system of this type with (i) different search algorithms and (ii) different
490: queuing disciplines, with a view to
491: investigating their effect on packet transport.
492:
493:
494:
495: \section{Acknowledgements}
496: We would like to thank NATO for partial financial support through PEST grant
497: PST.CLG.978404. Work of B.T. was supported by the Ministry
498: of Education, Science and Sports of the Republic of Slovenia.
499:
500:
501: %***********************************************************
502: \begin{thebibliography}{99}
503: \bibitem{csabai}I. Csabai, J. Phys. A {\bf 27} 417 (1994).
504:
505: \bibitem{t1} M. Takayasu, H. Takayasu and T. Sato, Physica A {\bf 233} 824
506: (1996).
507: \bibitem{chong}K. B. Chong and Y. Choo, physics/0206012.
508: \bibitem{erramilli} A. Erramilli, O. Narayan and W. Willinger,
509: IEEE/ACM Trans. Networking {\bf 4} 209 (1996).
510:
511:
512: \bibitem{fal} M. Faloutsos, P. Faloutsos and C. Faloutsos, Proc. ACM SIGCOMM,
513: Comput. Commun. Rev. {\bf 29} 251 (1999).
514: \bibitem{gov} R. Govindan and H. Tangmunarunkit, Proc. IEEE Infocom 2000,
515: Tel Aviv, Israel (2000).
516: \bibitem{Vazquez} A. Vazquez, R. Pastor-Satorras, and A. Vespignani,
517: cond-mat/0112400.
518:
519: \bibitem{ba} R. Albert and A.-L. Barabasi, Rev. Mod. Phys. {\bf 74} 47 (2002).
520: \bibitem{DMS}S.N. Dorogovtsev, J.F.F. Mendes, and A.N. Samukhin,
521: Phys. Rev. Lett. {\bf 85}, 4633 (2000).
522: \bibitem{BT} B. Tadi\'{c}, Physica A {\bf 293} 273 (2001).
523: \bibitem{KRR} P.L. Krapivsky, G. J. Rodgers, and S. Redner,
524: Phys. Rev. Lett. {\bf 86}, 5401 (2001).
525:
526: \bibitem{load} K.-I. Goh, B. Kahng and D. Kim, Phys. Rev. Lett. {\bf 87}
527: 278701 (2001).
528: \bibitem{newman}M. E. J. Newman, Phys. Rev. E {\bf 64} 016131 (2001).
529:
530: \bibitem{book} G. Gross and C.M. Harris, {\it Fundamentals of Queuing Theory},
531: (Wiley, New York 1998).
532:
533:
534: \bibitem{micro} T. Huisinga, R. Barlovic, W. Knospe, A. Schadschneider and M.
535: Schreckenberg, Physica A {\bf 294} 249 (2001).
536:
537: \bibitem{Arenas}A. Arenas, A. Diaz-Guilera, and R. Guimera, Phys. Rev. Lett.
538: {\bf 86}, 3196 (2001).
539: \bibitem{Guimera}R. Guimera, A. Arenas, A. Diaz-Guilera, and F. Giralt,
540: cond-mat/0206077.
541:
542: \bibitem{Albert}R. Guimera, A. Arenas, A. Diaz-Guilera, F. Vega-Redondo, and
543: A. Carbales, cond-mat/0206410.
544:
545: \bibitem{abate} J. Abate and W. Whitt, Opns. Res. Letters {\bf 20} 199 (1997).
546: \bibitem{Whitt} W. Whitt, {\it Stochastic-Process Limits}, Ch. V ,
547: (Springer, New York 2002).
548:
549: \end{thebibliography}
550: \end{multicols}
551: %************************************************************
552: \newpage \twocolumn
553: \narrowtext
554: \begin{figure}
555: \epsfxsize=80mm\epsffile[46 72 514 486]{pack_fig1.ps}
556: \vskip 0.8true cm
557: \caption{\label{fig1} Average load at active nodes $\ell _a (t)$
558: vs. time $t$ (top panel) and number of active nodes in the network $a(t)$ at
559: time $t$ (lower panel) in jamming regime $R=0.04$, $B=100$ for scale-free
560: graph (curve a) and for randomly-grown graph (curve b) with the same
561: driving conditions.
562: }
563: \end{figure}
564:
565: \begin{figure}
566: \epsfxsize=80mm\epsffile[43 71 473 615]{pack_fig2.ps}
567: %\vskip 0.8true cm
568: \caption{\label{fig2} (bottom to top) Power spectrum of activity,
569: load on active nodes, density at hub, and total load for scale-free
570: graph (left column) and in randomly grown graph (right column) for
571: driving conditions corresponding to the free flow regime ($R=0.01$, $B=100$).
572: Power-law fits are shown by solid lines. Copies of the fits for SFN are
573: also shown in the corresponding panels for RGN as dotted lines to
574: display the differences in slopes and in the correlation ranges. }
575: \end{figure}
576:
577:
578: \begin{figure}
579: \epsfxsize=80mm\epsffile[63 96 505 546]{pack_fig3.ps}
580: \vskip 0.8true cm
581: \caption{\label{fig3}Power spectrum of the density measured at the hub for
582: $R=0.04$ and $B=100$ in (a) SFN and (b) RGN, and at far node in
583: free-flow regime ($B=1000$) in (c) RGN and (d) SFN. Curve (c)
584: has been shifted shifted to display it more clearly and data
585: has been log-binned. Fitted slopes are (top to bottom) -2, -1, -0.59,
586: and 0.03.}
587: \end{figure}
588:
589:
590: \begin{figure}
591: \epsfxsize=80mm\epsffile[70 70 508 488]{pack_fig4.ps}
592: \vskip 0.8true cm
593: \caption{\label{fig4}Probability distribution of transit times of
594: marked packets in the conditions of free flow ($R=0.01$,$B=100$).
595: for RGN and SFN. Log-binning ratio 1.1 . Inset: A set of transit
596: times of 1000 marked packets against posting time, for the same
597: conditions as in the main figure.
598: }
599: \end{figure}
600:
601:
602: \begin{figure}
603: \epsfxsize=80mm\epsffile[44 71 508 488]{pack_fig5.ps}
604: \vskip 0.8true cm
605: \caption{\label{fig5}Probability distribution of the queue lengths
606: counted after each time step was completed for RGN and SFN and the
607: respective fit lines.
608: Driving rate $R=0.12$ and buffer size $B=1000$. Log-binning ratio 1.1 .
609: }
610: \end{figure}
611:
612:
613:
614: \begin{figure}
615: \epsfxsize=80mm\epsffile[43 71 507 488]{pack_fig6.ps}
616: \vskip 0.8true cm
617: \caption{\label{fig6}Time dependence of the workload $n(t)/t$ in SFN for two
618: input rates $R=0.08$ (top) and $R=0.01$ (bottom) and buffer size $B=100$.
619: Fits: $R-\mu = 0.023 - 0.5t^{-1}$ (solid line) and $R-\mu = 0.07t^{-0.5}$
620: (dotted line).
621: Inset: Average output rate of the network $\mu /R$ normalized to input
622: rate $R$, vs $R$ for SFN (bullets) and RGN (squares).}
623: \end{figure}
624:
625:
626:
627:
628:
629:
630: %************************************************************
631: %\end{multicols}
632: \end{document}
633: %%*****************************************************************************#!/bin/csh
634: -f
635: # this uuencoded Z-compressed .tar file created by csh script uufiles
636: # for more information, see e.g. http://xxx.lanl.gov/faq/uufaq.html
637: # if you are on a unix machine this file will unpack itself:
638: # strip off any mail header and call resulting file, e.g., packfigures.uu
639: # (uudecode ignores these header lines and starts at begin line below)
640: # then say csh packfigures.uu
641: # or explicitly execute the commands (generally more secure):
642: # uudecode packfigures.uu ; uncompress packfigures.tar.Z ;
643: # tar -xvf packfigures.tar
644: # on some non-unix (e.g. VAX/VMS), first use an editor to change the
645: # filename in "begin" line below to packfigures.tar_Z , then execute
646: # uudecode packfigures.uu
647: # compress -d packfigures.tar_Z
648: # tar -xvf packfigures.tar
649: #
650: uudecode $0
651: chmod 644 packfigures.tar.Z
652: zcat packfigures.tar.Z | tar -xvf -
653: rm $0 packfigures.tar.Z
654: exit
655:
656: