comparison of available bandwidth estimation techniques ... - CiteSeerX

The 17th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’06)

COMPARISON OF AVAILABLE BANDWIDTH ESTIMATION TECHNIQUES IN PACKET-SWITCHED MOBILE NETWORKS ´ Carlos Ubeda Castellanos† , Dimas López Villa† , Oumer M. Teyeb† , Jan Elling† , Jeroen Wigard‡ ‡ Center for Teleinfrastrucktur, Cellular Systems Division Nokia Networks, Aalborg R&D Aalborg University, Niels Jernes Vej 12 Niels Jernes Vej 10 9220 Aalborg East, Denmark 9220 Aalborg East, Denmark {cubeda, dimas, oumer, je}@kom.aau.dk [email protected]

†

A BSTRACT The relative contribution of the transport network towards the per-user capacity in mobile telecommunication systems is becoming very important due to the ever increasing air-interface data rates. Thus, resource management procedures such as admission, load and handover control can make use of information regarding the available bandwidth in the transport network, as it could end up being the bottleneck rather than the air interface. This paper provides a comparative study of three well known available bandwidth estimation techniques, i.e. TOPP, SLoPS and pathChirp, taking into account the statistical conditions of the available bandwidth and assessing the variability of their estimations. Simulation-based studies on a mobile transport network show that pathChirp outperforms TOPP and SLoPS, both in terms of accuracy and efficiency. I.

ited. Fig. 1 illustrates a usage scenario where AP can be used by employing agents in different network components within different systems. The source agents could be located on the BSC of GERAN in a GPRS system, on the RNC of UTRAN in UMTS and on the BS of E-UTRAN, whereas the destination agents could be at the service provider. The goal of this paper is to propose a method from existing ABwE techniques so that it can be applied to current and emerging mobile communication networks. The paper is organized as follows: §II. presents the state of the art in ABwE. §III. summarizes the previous work on comparison of ABwE techniques. §IV. studies the viability of comparing the different methods under the same statistical conditions. §V. shows the simulation results illustrating the performance of the different methods. Finally, §VI. summarizes the main conclusions from this study and gives some pointers regarding future work.

I NTRODUCTION

Understanding the dynamic properties of the end-to-end (E2E) available bandwidth (ABw) is beneficial for the proper resource management in existing and emerging mobile communication systems. The increasing trend in the wireless interface data rates means that the requested data rate for a certain service might not be guaranteed, not only because of the air interface bandwidth limitation, but also due to a limitation in the transport network’s ABw. Available bandwidth estimation (ABwE) is a very challenging task due to the heterogeneity of the current systems and the different traffic characteristics of different data flows. One possible way of ABwE is the deployment of specialized software on every router, which continuously reports the router’s load. However, this is impractical, not only because it is costly to upgrade existing routers, but also because of the overload situation that might occur due to the huge amount of reporting traffic that ensues. Also, it is difficult to obtain an E2E measure since operators are not usually willing to share information regarding the loads in their routers and links. An alternative is to use E2E software that runs on the end hosts. This is usually called active probing (AP). However, this approach means an inference of the ABw, not a direct measuring, which entails several hindrances. The ABw is a time varying metric that exhibits variability depending on the observed time-scale [1], and traffic prioritization could affect the methods’ performance. Despite all these disadvantages, AP is more practical than direct ABw measurements when there is no previous knowledge of the network, and the resources are limc 1-4244-0330-8/06/$20.00°2006 IEEE

II.

S TATE OF THE A RT IN AVAILABLE BANDWIDTH E STIMATION

ABwE techniques can be classified into Direct or Iterative Probing depending on whether they sample the ABw, or they iteratively check if the input rate is larger than the ABw. This section gives a brief description of the most common methods, and the reader is referred to [1] for a more complete survey. A.

Direct Probing

Direct Probing (DP) techniques give a value of the ABw by estimating the cross-traffic rate. A perfect example of these techniques is Delphi [2], which assumes a complex multi-fractal model to characterize the cross-traffic. The main advantage is the real-time adaptation of this model to the current traffic.

Agent SGSN BTS

GERAN

Operator Service Network

BSC

UE

GGSN

Agent Node B UTRAN

RNC

UE

E-UTRAN

IP backbone

3G GGSN

Internet Services

Agent BS

3G SGSN

SN

UE

Figure 1: Location of the agent in the transport network.


Nevertheless, it requires previous knowledge of the capacity of the tight-link1 . B. Iterative Probing Iterative Probing (ItP) techniques are based on self-induced congestion, which consists of sending streams of packets whose input rate iteratively increases. The lowest input rate overloading the network is taken to be the ABw. ItP techniques are able to do the estimation without any knowledge about the tight-link capacity. Train of Packet Pairs (TOPP) [3], implemented in a tool called DietTopp, sends streams of packet pairs, uniformly increasing their input rates each iteration. The rate is changed by modifying the input gap of each pair. The ABw is estimated as the maximum input rate not larger than the measured rate at the destination. Self-Loading Periodic Streams (SLoPS) [4], implemented in a tool called pathload, sends streams of equally spaced packets. Instead of changing the input rate linearly as TOPP, it performs a binary search. The rate varies by modifying the packet size. It takes into account the variability of the ABw by giving a range of variation, rather than a single value. It requires feedback from the destination to set the next input rate. pathChirp [5] sends streams of exponentially spaced packets called chirps, so the instantaneous input rate changes. Only one iteration is needed to get an ABwE, since it probes the network with different input rates in each stream. III.

very specific scenarios. Thirdly, most of the method parameters used in the simulations are not clearly stated. Fourthly, most of the studies compare different methods under different time-scales and using different number of samples of the ABw, which can vary the statistical properties from one estimation to another [1]. Finally, the variability of estimations has not been studied. This paper provides a comparative study of different ABwE techniques in terms of accuracy, efficiency and variability, attempting to normalize time-scale and number of samples of the ABw. IV.

Techniques that require previous knowledge of the tight-link capacity are not studied in this paper, since it is not possible to determine such capacity without assuming that the tight-link and the narrow-link are the same, leading to errors [1]. Therefore, the study is focused on ItP techniques, specifically TOPP, SLoPS and pathChirp. First of all, it is necessary to define some structural concepts. As Fig. 2 points out, a set of K probing packets forms a stream. A stream is sent M times, which constitutes a fleet, and the results are averaged in order to improve the accuracy. A non-intrusiveness gap TN I is left between streams in order to reduce the average probing rate and to prevent a stream from interfering with the next one. This process is iterated I times. The time between the first and last packets of the whole procedure is defined as the probing time. 1 τ

R ELATED W ORK

There has been a lot of research in this area over the last two decades. Many techniques have been proposed, and some performance and comparative studies have been done. In [6], performance comparison of pathChirp and pathload in a singlehop network environment is given among others. In their experiments all the tools showed an accuracy within 30%. In [7], results of a series of ABwE experiments conducted on a high-speed testbed are presented. ABwE tools including pathChirp, pathload and others based on packet pairs are evaluated. The results show that packet pair techniques perform worse than pathChirp and pathload. Comparison of pathChirp, pathload and TOPP in single and multi-hop real network is given in [5]. PathChirp was shown to perform better in terms of both accuracy and efficiency. In [8], the difference between ABw measurements in wired and wireless networks is discussed. DietTopp is compared in performance with pathload in a wired testbed. The tool is also evaluated in a wireless environment. It is shown that packet size is critical for ABw measurements. There is some controversy on the validity of the aforementioned comparisons between the different ABwE techniques. First, some of the previous work gives very low accuracy results that are used to propose a method as the best choice. Secondly, some conclusions are drawn from very few results under 1 The tight-link is the link with the minimum ABw of a path, while the narrow-link is the one with the minimum capacity.

I MPLEMENTATION OF I T P T ECHNIQUES

ΤNI

2 τ

M τ

1) 1

2

K

2)

I)

Figure 2: Scheduling structure. The statistical conditions of the ABw process are fixed by the time-scale τ and the number of samples N [1]. The timescale is the time a stream interacts with the cross-traffic in a certain moment, i.e. the stream duration (see Fig. 2). Every two consecutive packets (and every packet pair in TOPP) of a stream, an input rate is set. A sample of the ABw is obtained as a result of checking whether such input rate is larger than the ABw or not. So, the number of samples is related to the number of packets per stream. A.

Viability Study of Statistical Comparison

An initial mathematical analysis of ItP techniques has been conducted to see if it is possible to set the parameters of the three suggested techniques in such a way that they are working under the same statistical conditions to measure a similar ABw range [Rmin , Rmax ]. Fig. 3 represents such study for Rmin = 1M bps. In 3(a), the time-scale as a function of the number of samples is shown for the three methods, using their optimum packet sizes [4, 5, 9]. There is an extra plot of


TOPP for the minimum packet size to avoid link-layer effects (P = 200B) [4]. In 3(b) , the maximum input rate2 of SLOPS and pathChirp is compared. Statistical Comparison 100 (a) τ (ms)

80 60 40 20 0

1

3

5

7

9

decision increases TOPP probing time. As a trade-off between the probing time and the network overload, TN I is fixed at least as large as the time-scale, which reduces the average probing rate by at least 50%. Specifically, TN I = 40ms. Table 1 summarizes the main parameters for the simulations, while the rest of the method parameters are fixed according to the original proposals for SLoPS and TOPP, and pathChirp is configured for Internet traffic. Note that the number of iterations in SLoPS cannot be initially determined (?) and depends on the resolution, set at 0.25M bps.

11 13 15 17 19 21 23 25 27 29 31 33 35 N

Rmax (Mbps)

50 TOPP P=200B TOPP P=800B SLoPS pathChirp P=1000B

40 30 20

(b)

10 0

1

3

5

7

9

Parameter

TOPP

SLoPS

pathChirp

N I P (Bytes) R(M bps)

4 25 500 [1.0-9.0]

25 ? [200-1500] [1.0-7.5]

12 1 1000 [1.1-7.9]

Table 1: Configuration parameters.

11 13 15 17 19 21 23 25 27 29 31 33 35 N

Figure 3: Statistical comparison. From Fig. 3(a), it can be seen that SLoPS and TOPP can not be compared for a significant number of samples. TOPP is only comparable with pathChirp, but only if a probing packet size that is very sensitive to cross-traffic is used [9]. Fig. 3(b) shows that pathChirp and SLoPS are not comparable since they measure very different ranges for the same time-scale and number of samples. From the results, it can be seen that it is not possible to study the performance of TOPP, SLoPS and pathChirp under the same statistical conditions to measure a certain ABw range, even modifying their recommended parameters. Therefore, it is necessary to choose between the time-scale or the number of samples as the common statistical parameter. The time-scale is directly related to probing time, the latter being a determining factor for the field of application of this study. In the following simulations, τ is fixed for all three methods, while N is chosen for each method to be as large as possible to measure a similar ABw range without modifying excessively their optimum parameters. B. Adjustment of Parameters In order not to modify τ and N during the probing time, the maximum measurable range in SLoPS is fixed at Rmax = 7.5Rmin . Therefore, the desired ABw range is set at R ∈ [1.0 − 7.5]M bps. Considering that the variance of the ABw is considerably reduced for time-scales larger than 10ms [1], τ is set at 40ms, which will also not lead to a long probing time. SLoPS sets M at 12, and TOPP sends only one stream per fleet, whereas in pathChirp, an optimum M value is not determined. In order to keep the same conditions during the simulations for the three methods, it has been decided to use M = 10, although this 2 TOPP

is not shown since its Rmax does not depend on N and τ .

V.

S IMULATION R ESULTS

In order to study the performance of the suggested methods under different cross-traffic models in single and multi-hop paths, several simulations have been conducted3 . A.

Simulation Scenario

Fig. 4 represents the simulation network model, composed of a probing source and a probing destination, i.e. the ABwE agents, joined by an H-hop path. It is based on one-hop persistent cross-traffic routing (i.e. the packets from a given crosstraffic source only travel through one hop of the path.) so as to control the ABw in each hop. The methods are evaluated under different kinds of crosstraffic: Constant Bit Rate (CBR) (which is the fluid traffic model initially assumed by all three techniques), Packet Size Distribution CBR (PSD-CBR) (based on CBR, but making use of a random packet size distribution obtained from a study of Internet traffic characteristics [10])4 , and Poisson traffic model. 3 All the simulation results presented in this paper are generated using NS-2. 4 PSD-CBR sends equally spaced packets and the average cross-traffic rate is obtained from the expected packet-size.

XD1

XD2

XDH-1

XDH X-Traffic Destinations

Probing Source 1

2

3

H

H+1

Probing Destination X-Traffic Sources XS 1

XS 2

XS 3

XS H

Figure 4: Network topology for the simulations.


1) Packet Size Dependency Fig. 5 shows the cross-traffic packet size dependency under CBR traffic in a single-hop scenario. The figure is obtained averaging the relative errors calculated within the whole interval for each packet size. TOPP is strongly dependent on the packet size. Specifically, it performs the best until the cross-traffic packet size is larger than the probing packet size, 500 Bytes in this case. Such dependency is easily explained by the fact that it uses packet pairs, which are very sensitive to cross traffic [9]. This constraint can be mitigated increasing the probing packet size. However, this means, for example, increasing the time-scale not to reduce the number of samples, as Fig. 3 shows. Note that the packet size is not a key factor in SLoPS and pathChirp performance, keeping an average relative error around 10%. 2) Cross-traffic Pattern Dependency Fig. 6 compares TOPP, SLoPS and pathChirp under different kinds of cross-traffic. In Poisson and CBR traffic models it has been chosen Px = 500B because it is a representative average cross-traffic size [11]. TOPP is quite accurate for CBR traffic if Px ≤ P , as explained previously. However, TOPP does not work for PSDCBR and Poisson models, which points out that it is sensitive not only to cross-traffic packet size distributions, but also to random inter-arrivals. PathChirp performs the best in single-hop scenario regardless of the kind of cross-traffic. This behavior could be due to 5A

higher range R ∈ [8 − 60]M bps is also studied, but the results are not included since they are very similar to the low range results.

Average relative error (%)

CBR packet size dependency 90 80 70 60 50 40 30 20 10 0 40

ABwE (Mbps) ABwE (Mbps)

B. Results

Performance comparison under different cross−traffic models

ABwE (Mbps)

Measurements of the ABw are taken in steps of 0.5Mbps from 1Mbps to 7.5Mbps5 for different cross-traffic packet sizes Px ∈ {40, 100, 200, ..., 1500} bytes, except in PSD-CBR. The capacity of the tight-link is set at 10Mbps, whereas the latency is fixed at 10ms. The joining links have a latency of 10ms and 1Gbps of capacity. In multi-hop simulations (H = 5), the same configuration as in single-hop is used , but the ABw of the non-tight-links is fixed at 9Mbps to make sure there is only one tight-link. Each estimation is repeated 25 times in order to study the variance of the methods. Considering that SLoPS gives an interval of variation of the ABw, the center of such interval is used as the estimated value to compare SLoPS with the other two methods.

7.5 7 6.5 6 5.5 5 4.5 4 3.5 3 2.5 2 1.5 1

7.5 7 6.5 6 5.5 5 4.5 4 3.5 3 2.5 2 1.5 1

TOPP SLoPS pathChirp

(a) 1

1.5

2

2.5

3

3.5

4 4.5 ABw (Mbps)

5

5.5

6

6.5

7

1

1.5

2

2.5

3

3.5

4 4.5 ABw (Mbps)

5

5.5

6

6.5

7

1

1.5

2

2.5

3

3.5

4 4.5 ABw (Mbps)

5

5.5

6

6.5

7

7.5

(b) 7.5

(c) 7.5

Figure 6: Performance comparison in single-hop scenario under CBR (a), PSD-CBR (b) and Poisson (c) traffic. The x-axis represents the offered ABw, while the y-axis represents the estimated ABw. the fact that pathChirp makes use of the excursions analysis, which takes into account the burstiness of traffic [5]. However, it works slightly worse in multi-hop (see Fig. 7), since it is more likely that the exponential spacing of the chirp is modified by the other hops of the path. As pathChirp probes the network with each rate once per stream, the spacing variation is decisive. As Fig. 7 shows, SLoPS is less sensitive to multi-hop paths as it repeatedly measures the same rate during the stream, which makes it more robust against variations of the stream structure. Both pathChirp and SLoPS tend to underestimate the ABw except for values very close to the lowest bound of the measurable range. 3)

TOPP

7.5 7 6.5 6 5.5 5 4.5 4 3.5 3 2.5 2 1.5 1

Study of Variance

SLoPS pathChirp

186

332

478 624 770 916 1062 Cross−traffic packet size (Bytes)

1208

Figure 5: CBR packet size dependency.

1354

1500

Fig. 8 shows the variance of the methods under PSD-CBR traffic for a single-hop scenario. TOPP experiences a great variability, which is most of the times comparable to the estimation. Due to the random nature of the packet size distribution used in PSD-CBR traffic, TOPP interacts with different packet sizes in each estimation. Therefore, the aforementioned TOPP packet size dependency is also the source of its variability. On


ABwE (Mbps)

PSD−CBR traffic model simulation 7.5 7 6.5 6 5.5 5 4.5 4 3.5 3 2.5 2 1.5 1

TOPP SLoPS pathChirp

1

1.5

2

2.5

3

3.5

4 4.5 ABw (Mbps)

5

5.5

6

6.5

7

7.5

Figure 7: Performance under PSD-CBR in multi-hop. the other hand, the variance of SLoPS and pathChirp is quite small in comparison, the latter showing a better behavior at the end of the interval. 4) Efficiency Table 2 summarizes the efficiency of the different methods based on their level of interference with the network load and their speed. PathChirp is the most efficient method since its load and probing time is the smallest. Nevertheless, TOPP is the less intrusive method due to its long probing time resulting in a low average rate. Parameter

TOPP

SLoPS

pathChirp

Load (KB) Probing time (s) Average rate (Mbps)

977 19.96 0.40

1079 3.96 2.23

127 0.76 1.37

TOPP pathChirp

1.5

2

2.5

[7] Alok Shriram, Margaret Murray, Young Hyun, Nevil Brownlee, Andre Broido, Marina Fomenkov, and Kimberly C. Claffy. Comparison of public end-to-end bandwidth estimation tools on high-speed links. In Constantinos Dovrolis, editor, PAM, volume 3431 of Lecture Notes in Computer Science, pages 306–320. Springer, 2005. ISBN 3-540-25520-6. [8] Andreas Johnsson, Bob Melander, and Mats Björkman. Bandwidth measurement in wireless networks. In Mediterranean Ad Hoc Networking Workshop, Porquerolles, France, June 2005. [9] Attila Pasztor and Darryl Veitch. The packet size dependence of packet pair like methods. IEEE/IFIP International Workshop on Quality of Service (IWQoS), 2002.

[11] CISCO White Papers. Performance measurements of advanced queuing techniques in the CISCO IOS, 2000.

SLoPS

2

σ (Mbps)

6

1

[3] Bob Melander, Mats Björkman, and Per Gunningberg. A new end-to-end probing and analysis method for estimating bandwidth bottlenecks. IEEE Global Internet Symposium, November 2000.

[10] Sean McCreary and Kimberly C. Claffy. Trends in wide area IP traffic patterns - a view from Ames Internet exchange. ITC Specialist Seminar, 2000.

Study of variance under PSD−CBR traffic

0

[2] Vinay J. Ribeiro, Mark Coates, Rudolf H. Riedi, Shriram Sarvotham, Brent Hendricks, and Richard Baraniuk. Multifractal cross-traffic estimation. In Proc. of ITC Specialist Seminar on IP Traffic Measurement, September 2000.

[6] Diane Kiwior, James Kingston, and Aaron Spratt. Pathmon, a methodology for determining available bandwidth over an unknown network. IEEE Sarnoff Symposium, April 2004.

C ONCLUSIONS

8

2

[1] Manish Jain and Constantinos Dovrolis. Ten fallacies and pitfalls on end-to-end available bandwidth estimation. In Alfio Lombardo and James F. Kurose, editors, Internet Measurement Conference, pages 272– 277. ACM, 2004. ISBN 1-58113-821-0.

[5] Vinay J. Ribeiro, Rudolf H. Riedi, Richard G. Baraniuk, Jiri Navratil, and Les Cottrell. pathChirp: Efficient available bandwidth estimation for network paths. In Passive and Active Measurement Workshop, April 2003.

The main contribution of this paper is a comparative study of ABwE techniques taking into account the statistical conditions of the ABw. Even though the performance of none of the investigated methods is found to be outstanding, pathChirp excels as the best tool in terms of both accuracy and efficiency. It shows no packet size dependency and an acceptable behavior in multi-hop environments under different cross-traffic models. The performance of SLoPS is not that bad compared with pathChirp, but it needs quite a longer time to give an estimation. TOPP is found to be very sensitive to the cross-traffic

4

R EFERENCES

[4] Manish Jain and Constantinos Dovrolis. End-to-end available bandwidth: measurement methodology, dynamics, and relation with TCP throughput. IEEE/ACM Trans. Netw., 11(4):537–549, 2003.

Table 2: Efficiency parameters.

VI.

packet size and is also very slow. We are modifying the algorithm employed by pathChirp to improve its performance. The enhancements we are undergoing include using profiles other than the exponential structure described in [5], optimizing the different parameters of the method, and using iteratively one estimation as input for the next one. We would also like to evaluate pathChirp in a network under traffic prioritization such as Differentiated Services (DiffServ), and identify the possibilities of using RTT values instead of one way delay measurements, which requires only a source agent, since the service provider might not have control over the E2E path.

3

3.5

4 4.5 ABw (Mbps)

5

5.5

6

6.5

7

7.5

Figure 8: Study of variance under PSD-CBR traffic in singlehop.