Effect of Path Diversity on the loss performance of UDP Packets over the Internet Richard Haywood and Xiao-Hong Peng Electronic Engineering Aston University Birmingham, United Kingdom [email protected]

Abstract— In this paper we investigate the effect of path diversity on the packet loss performance through simulations and experiments using three overlay networks created on top of the PlanetLab testbed. The overlay network consists of a number of relay nodes that are positioned around the Internet. In each of the overlay networks UDP packets are streamed from a client to a destination for a period of 50 days, across up to six paths with a static packet dispatching algorithm. From the results obtained, we can see that increasing the number of paths could increase as well as decrease the packet loss probability, depending on the network environment and the structure of the multiple paths created. We have also investigated the correlation in the underlying network and demonstrated the performance of the metric defined.

I. I NTRODUCTION A. Background Path diversity is a means to use multiple diverse or partially diverse paths to transport packets from a source to a destination. The reason for using path diversity is to provide the possibility for fault tolerance, diverse routing and error reduction as a result of load balancing and smaller fluctuations in traffic load. Many researchers have investigated the topic of path diversity. In [1] the authors look at Sprint’s (a large US Internet Service Provider (ISP)) internal network and find that for a large percentage of Points of Presence pairs that there are multiple partially diverse paths between them. Additionally the authors look at the diversity across multiple networks and highlight that the diversity is not fully exploited. The reason that diversity is not fully exploited is because of the conventional architecture of the Internet. When packets are sent from a source to a destination, they generally go via the same set of routers. This is because the packets will travel across the same Autonomous Systems (AS) [2]. ASs are each administratively independent, with the potential for financial gain if the setup of the network is not widely known. Each AS is connected to a number of other ASs, so if the destination for the packet is not in its own network it must forward the packet onto one of the connected ASs. To decide where to forward the packet, reachability details are required for each of the connected ASs. The Boarder Gateway Protocol (BGP) provides reachability data to other ASs without the need to reveal its topology or routing policy to potential competitors[3]. To ensure that

packets get to their destination BGP is designed to converge onto a route. If BGP does not converge then there is the possibility that circular routing might exist, and thus for packets to be trapped in part of the network. The problem with converging onto a route is that this could limit the potential exploitation of diversity. The most obvious solution for this problem would be to change BGP so that it does not advertise the single best route but passes on details about more available routes, which is the technique proposed by the authors of [4]. If we were developing the Internet again then altering BGP would be one of the best ways to exploit diversity. The problem is because of the scale of the Internet replacing BGP would be very difficult to achieve. It would be possible for both BGP and the scheme proposed in [4] to run side by side, but the full benefits would only be achieved when all routers had been changed over. Another way to exploit diversity is through the use of multihoming, which utilises diversity by connecting to multiple ISPs and sending different packets onto each network. For multihoming to be as effective as possible the details of peering arrangements are needed. If both networks peer with each other, or have lots of common peers then it is unlikely that the packets will be routed through different paths towards the destination. In [5], the authors set up a multihoming network by connecting to two ISPs. When 80,000 destinations are contacted it is shown in [5] that with their selection of ISPs on average just four links were in common. If the destination ASs are ignored then just one link in common.

Fig. 1.

Illustration of Path Diversity

In this paper we use an overlay network to exploit diversity. An overlay network operates on top of the Internet and contains a number of relay nodes, as illustrated in Figure 1 with one additional path. If a relay node is used when sending

a packet then the packet will not always follow the same path as if it had been sent directly. The advantage of an overlay network is that it works at the application layer and as such does not need routing hardware to be altered. Secondly it does not require clients to physically connect to additional ISPs, which is the most likely scenario for residential users. Overlay networks have received some attention including [6]; which provides an overview of both path diversity and content delivery networks and then the authors looks at overlay networks in the wireless environment. In addition, the authors of [7] set up an overlay network and use active probing to monitor network characteristics. The system provided in [7] enables an application to indicate to lower layers to select the best network path based on the type of traffic, resulting in low loss, high Transmission Control Protocol (TCP) throughput or low latency. The problem with active probing is that it has a significant overhead and could unfortunately aid in the creation of congestion. In our work we use a static dispatching algorithm to minimise the overhead of the overlay network. In addition to the investigations into the methods of utilising diversity, some authors [8][9] have also looked at the methods for maximising the throughput. They have achieved this by altering the amount of data sent down different paths. The methods proposed in these papers are better than the active probing techniques used in [7], in terms of bandwidth efficiency. However, they still require additional feedback to the server. In some environments this is not desired as this increases the amount of processing required and can limit the number of acceptable clients [10]. In this paper we assume that a reverse channel is unavailable.

II. S IMULATIONS The causes of packet loss in packet switching networks, such as the Internet, are numerous. In this paper we classify causes of packet loss into two types; congested and noncongested losses. Congestion occurs when more packets arrive at a router than it is able to transmit, due to bandwidth or processing limitations, and the router runs out of space to buffer the packets. The cause of non-congested losses are a lot wider, ranging from router failure to bit level packet corruption.

Fig. 2.

Sample topology

In our simulations we set up a topology similar to that shown in Figure 2. The UDP stream from the server has a total data rate of 800 kbps and is split equally over the overlay paths created. To cause packet losses, we included exponential on/off background traffic down each link, in such a way that on average after every 40 milliseconds of traffic generation it is stopped for 8 seconds. The background traffic is configured so that when active it has a rate equal to the link capacity of 2Mbps. Each node is set to have enough space to buffer just one packet. To simulate burst losses we used a two state Markov process as shown in Figure 3, where packets are successfully passed from one router to the next in the good state and are dropped in the bad state. The P and Q values used in the simulation were set to P = 0.0014 and Q = 0.854.

B. Our Work This paper intends to address the concerns about the possible benefits which could arise through the use of path diversity. In our simulation work we found that when a network experiences congestion solely then the use of path diversity reduced the number of packets lost in the system. In reality packet losses can be caused by congestion and noncongestion sources. In order to investigate the combination of the two loss sources in a large scale and complicated network environment we create an overlay network using the PlanetLab testbed on the Internet. It has always been assumed that the wired Internet experiences mainly congestion losses [11] and fewer non-congestion losses of data packets. This assumption has been used in the fundamental design of many protocols. In our investigation we want to test the losses from the two different sources by using the User Datagram Protocol (UDP), because it has no error recovery methods. The remainder of this paper is arranged as follows. Section II presents the simulation results which provide the motivation for our work. Section III describes the set-up of the PlanetLab testbed. Section IV shows the results gained from the PlanetLab testbed. Section V provides conclusions on our work and discusses the potential for future work.

Fig. 3.

A two-state Markov model

To find suitable values for the number of links on each path we created a large number of two-layer topologies to simulate the Internet [12]. With the generated topologies we randomly selected a server and a client from all of the available hosts, and then from the remaining nodes we randomly selected a further 10% of the nodes to act as relays. The number of links for each path can now be calculated based on the topologies generated and the locations of the relay nodes. Using this data we were able to list the paths in terms of their length and select the shortest path as the primary one. Figure 4 shows the number of links for each path calculated as the average of over 1000 different two-layer topologies. The results in Figure 5 shows the effect of increasing the number of path when the networks either experience only congested or non-congested losses. From the results shown we can see that when the network experiences congestion then increasing the number of paths improves the loss percentage,

Group US1

19 Path Length 18

Number of Links

17

16

US2

15

14

13

12

EU

11 0

1

2

3

4

5

Path Number

Fig. 4.

Number of links on each path in the simulation

Task Source Destination Relay 1 Relay 2 Relay 3 Relay 4 Relay 5 Source Destination Relay 1 Relay 2 Relay 3 Relay 4 Relay 5 Source Destination Relay 1 Relay 2 Relay 3 Relay 4 Relay 5

0.03

Nodes righthand.eecs.harvard.edu planet2.cs.ucsb.edu planetlab2.cs.uoregon.edu planetlab2.csee.usf.edu pl2.cs.utk.edu planetlab2.unl.edu planetlab2.cse.msu.edu planet-lab1.cs.ucr.edu planet2.scs.cs.nyu.edu planetlab1.utep.edu planetlab1.cs.uchicago.edu planetlab2.cs.purdue.edu planetslug2.cse.ucsc.edu plab2.eece.ksu.edu planetlab2.sics.se planetlab1.fct.ualg.pt planetlab2.ls.fi.upm.es planetlab02.mpi-sws.mpg.de planetlab2.olsztyn.rd.tp.pl planetlab2.nrl.dcs.qmul.ac.uk planetlab1.cs.uit.no

TABLE I G ROUPED N ODES

Burst Losses Congetion Losses

0.02

40

0.015

35

0.01

30

US1 US2 EU

Path length (links)

Proportion of Packets Lost

0.025

0.005

25

20

0 0

1

2 3 Number of Extra Paths

4

5

15

Fig. 5. The percentage of occurrences of all paths being in error against the path correlation

10 0

but in the non-congested environment the opposite is true. For a more complete review the reader is referred to [13]. III. P LANET L AB S ETUP In our investigation of path diversity we set up a number of relay nodes in different ASs to create an overlay network. The reason for having relays is to force the IP packets to be routed along paths which they would not normally take to reach their ultimate destination. Using the PlanetLab nodes shown in Table I we deployed our software to create an overlay network similar to that shown in Figure 1. The path diversity software consists of three parts; a server, relay, and sink. The server generates packets at regular intervals (in our case every 20ms) which are then transmitted over the network. Each packet is sent either directly to the sink or to a specific relay node. Packets are sent to different nodes in a round robin manner with out repetition. Our dispatching algorithm differs to that of some other path diversity systems, such as in [14] where a packet is divided into N smaller packets, which are

1

Fig. 6.

2 3 Number of extra paths

4

5

Number of links on each path

then sent down N paths (one-to-one) in parallel. The amount of information transferred is the same for both dispatching algorithms. However, more data has to be transferred in the smaller packet case as each packet requires a header, and this could result in more packet losses because the likelihood of random errors is related to the number of bits transferred. Also, with each extra packet transferred there are additional requirements such as buffering and routing in the network [15]. The content of the packet is simply an identifying number with padding to make the payload up to 1000 bytes long which should not be fragmented by the network as it is smaller than the Ethernet maximum transmission unit (MTU). When the experiment start the relay node is sent details on the destination address and port number. All the packets that arrive at the relay application are then forwarded to the destination. The sink node logs the identifier numbers of the received packets and the local time of its arrival. Once the

0.1 US1 US2 EU

Proportion Lost

0.08

0.06

0.04

0.02

0 0

1

2

3

4

5

Extra Paths

Fig. 8.

Percentage of Packets Lost

that node are used in the experiment. The complete experiment was conducted over a period lasting 50 days which results in every group being tested at least 3 times with some paths being examined 11 times. IV. E XPERIMENT R ESULTS AND A NALYSIS Fig. 7. The number of links between select routers and relay nodes (shaded) used in the US1 overlay network

experiments have been completed the resulting log files are processed. To collect a range of results we completed the same experiment with three different sets of PlanetLab nodes. At most one node was selected from a PlanetLab site and each node selected would only participate in one group to minimise any interference that might have been caused between them. The nodes were grouped into geographical regions so as not to introduce large numbers of geographical common links, such as under-sea cabling. For each group of nodes, the two most geographically dispersed nodes were selected, one to be the source node and the other the sink node. The remaining nodes were assigned as relay nodes. The length of the path from the source to each relay node was computed using traceroute as well as the number of links from each relay node to the sink. The sequence of relay nodes was selected to minimise the total number of links making up one path. The path lengths in terms of the number of links are shown in Figure 6. An example of the overlay network setup is shown in Figure 7, this is the topology for the US1 group of nodes. Starting with just the primary path we streamed packets over the Internet for a period of 24 hours. We increased the number of paths used every 24 hours by adding a relay nodes until all the nodes in a group were utilised. 24 hours after all the nodes were added then the experiment would restart. As nodes can be disconnected at any time and the overlay topology needs to be kept constant, we only examine the overlay networks where all the relay nodes are present. This means that if a node became uncontactable then only the paths not requiring

The results presented here are always the mean values from the processed data with plus and minus one standard deviation plotted as the error bars. Figure 8 shows for all three groups how the use of path diversity affects the total packet loss rate. In the worst case the use of path diversity increases the average loss rate from 0.01118% to 0.02862% with one additional path. However when five paths are used between the same source and destination the loss rate falls to 0.01095%. To characterise the system we adopted a two-state Markov model, as illustrated in Figure 3, where packets are not received when in the Bad state and packets are successfully received when in the Good state. Using the data we gathered from the network we calculated the P and Q values such that P is the probability that the next packet will be lost if the current one has been received, and Q is the probability that if a packet is lost that the next packet will be successfully received. Figures 9 and 10 show how the P and Q values in Figure 3 change when path diversity is used. In general when five paths are used the Q value is improved compared to the single path case, this will result in bursts of shorter duration. In the case of the two US topologies the improvement in Q comes at the cost of P rising too, which will result in more bursts. The change in P and the Q values leads to the question of how many path could be in error at the same time. The reasons for multiple paths being in error at the same time could either be because independent links on all path simultaneously experience error at the same time or because common links are in error. The initial set up process was simply based on minimising the number of links between the client and server. The results shown in Figures 11 and 12 show that in some cases errors were experienced on multiple path at the same time, while in

0.01

1 P Value Q Value

0.9 0.8 0.4

0.007

0.3 0.005 0.004

Q Value

0.006

0.2

0.003 0.002

Cumulative Proportion of Occurences

0.008

P Value

No Extra Paths 1 Extra Path 2 Extra Paths 3 Extra Paths 4 Extra Paths 5 Extra Paths

0.5

0.009

0.7 0.6 0.5 0.4 0.3 0.2

0.1

0.001

0.1

0

0 0

1

2

3

4

0

5

0

1

Extra Paths

Fig. 9.

US1:P and Q value

Fig. 11.

0.35

0.5

2 3 4 Number of simultaneously faulty routes

1

0.3

No Extra Paths 1 Extra Path 2 Extra Paths 3 Extra Paths 4 Extra Paths 5 Extra Paths

0.95

0.4 Q Value

P Value

0.25

0.35

0.1

Cumulative Proportion of Occurences

0.45

0.15

6

US2: Number of simultaneously unavailable paths

P Value Q Value

0.2

5

0.9

0.85

0.8

0.3 0.75

0.05

0

0.25 0

1

2

3

4

5

0.7 0

Extra Paths

Fig. 10.

EU:P and Q value

other cases errors just occurred on one path. As an example, in the case of the US1 group, when three extra paths are selected, 98% of all losses occur across all four paths at the same time. This is very different to the EU group, where when three paths are used just 1.6% of all losses occurred on all paths at the same time. The change in values for P and Q will affect the number of consecutive losses and the number of consecutive successes. Figures 13 and 14 show the length of losses in the system. In the case of Figure 13, path diversity provides the benefits expected as it reduces the number of packets in a burst loss. However this is not always the case as it can be seen in Figure 14 that path diversity in fact increases the number of consecutively lost packets. If the change in the loss percentage for the system is not significant and the length of burst losses are getting shorter then the length of burst successes must be get shorter as well. Figures 15 and 16 show the number of consecutively received packet. From the results presented in Figure 16 we can see that in the single path case is approximately the average with some instances of path diversity increasing the

Fig. 12.

1

2 3 4 Number of simultaneously faulty routes

5

6

EU: Number of simultaneously unavailable paths

success duration as well as reducing the loss duration, however in Figure 15 their is an initial large drop in the number of consecutively received packets when one path is added but no much change there after. The alterations in the duration of successes and losses are very important when Forward Error Correction is used (FEC). FEC works in a group and it performs best when the errors are well distributed between groups. When using FEC an original k packets are encoded to produce n packets which are then sent across the network. If the client receives any k or more packets then the original k data packets can be recovered. The effect of receiving fewer than k packets is dependent on the encoding algorithm; some FEC encoders use a non-systematic code, in this case if fewer than k packets are received then none of the stream is decodable. The alternative to a nonsystematic code is a systematic one, this is where the original data is transmitted and the redundancy is sent separately. With a systematic code if fewer than k packets are received then a number of these packets could still be used by the application. In Figures 17 and 18 we show the results if a systematic FEC code was used with a block size of n = 30. The results

1

1 No Extra Paths 1 Extra Path 2 Extra Paths 3 Extra Paths 4 Extra Paths 5 Extra Paths

0.8 Cumulative Proportion of Occurences

Cumulative Proportion of Occurences

0.95

No Extra Paths 1 Extra Path 2 Extra Paths 3 Extra Paths 4 Extra Paths 5 Extra Paths

0.9

0.9

0.85

0.8

0.7 0.6 0.5 0.4 0.3 0.2 0.1

0.75

0 1

10

Fig. 13.

100 1000 Number of consecutivly lost packets

10000

100000

1

US2:Number of consecutively loss packets

Fig. 15.

100 1000 10000 Number of consecutivly received packets

100000

1e+06

US2:Number of consecutively received packets

1

1 No Extra Paths 1 Extra Path 2 Extra Paths 3 Extra Paths 4 Extra Paths 5 Extra Paths

No Extra Paths 1 Extra Path 2 Extra Paths 3 Extra Paths 4 Extra Paths 5 Extra Paths

0.9 0.8 Cumulative Proportion of Occurences

0.95

Cumulative Proportion of Occurences

10

0.9

0.85

0.8

0.75

0.7 0.6 0.5 0.4 0.3 0.2

0.7

0.1 0

0.65 1

10

1

10

Number of consecutivly lost packets

Fig. 14.

Fig. 16.

EU:Number of consecutively loss packets

presented in these two graphs show that FEC can work better when used in conjunction with path diversity rather than when FEC is used on just a single path. To aid in our analysis we calculated the correlation in terms of the percentage of the paths which are correlated (CN ) with other paths in our overlay network, defined by the formula:

CN = 1 −

PN

In,N n=1 Ln

N

(1)

Where N is the number of paths, Ln is the number of links on path n and In,N are the number of independent links on path n in the environment with N paths. Had the network we examined been a homogeneous network then we would have expected to clearly see a correlation between the path independence and the loss probability, however the Internet is a heterogeneous network and as shown in Figures 19 there is little correlation. Similarly in Figure 20, where the losses occur on all paths of the overlay network at the same time, shows no evidence of being correlated either.

100 Number of consecutivly received packets

1000

10000

EU:Number of consecutively received packets

V. C ONCLUSION AND DISCUSSION A. Conclusion In our work we have investigated the effect of path diversity on the packet loss properties when UDP packets are streamed across an overlay network on top of the Internet. We assume that the feedback channel is unavailable in our investigation, and that packets are dispatched towards the destination in a round robin manner. The results presented cover the scenarios where up to six paths are selected, based on the selection criterion that the N paths with the fewest number of links are selected among all the paths available in the overlay network. From the results obtained, we can see that increasing the number of paths could increase as well as decrease the packet loss probability. There are many factors that contribute to the loss properties concerned. In general, if congestion is the main source of packet losses, compared to the non-congestion related (such as random and burst) losses, path diversity can help divert traffic and reduce the loss probability as a result. However, if non-congestion related losses are dominant, path diversity may lead to more losses (this would mean that the

1

0.35

0.3

0.25 0.99 Proportion Lost

Proportion of sucessfully decoded packets

0.995

0.985

0.2

0.15

0.98 0.1 No Extra Paths 1 Extra Path 2 Extra Paths 3 Extra Paths 4 Extra Paths 5 Extra Paths

0.975

0.05

0.97 20

22

24

26

28

0 0.15

30

0.2

0.25

0.3 0.35 Path Correlation (CN)

Value of K

Fig. 17. US2:The effect of changing the value of k on the percentage of packets received

Fig. 19.

0.4

0.45

0.5

Percentage of loss against the path correlation

0.004 1 0.0035

Proportion of time all paths are in error

Proportion of sucessfully decoded packets

0.998

0.996

0.994

0.992

22

24

26

28

0.002

0.0015

0.001

0 0.15

0.988 20

0.0025

0.0005

No Extra Paths 1 Extra Path 2 Extra Paths 3 Extra Paths 4 Extra Paths 5 Extra Paths

0.99

0.003

30

0.2

0.25

0.3 0.35 Path Correlation (CN)

0.4

0.45

0.5

Value of K

Fig. 18. EU:The effect of changing the value of k on the percentage of packets received

benefits of path diversity in a congested environment are being negated due to the increase in non-congested losses). Our path ordering was based on minimising the number of links. There are a number of other ways which could have been chosen, including ordering based on minimising the number of shared links. The number of shared links can be characterised by path correlation, defined in this paper. Through the use of other path selection schemes might effect the proportions of congested and non-congested loss. B. Discussion There are a number of different ways that this work can be taken forward. One of the most common uses of UDP is in video streaming, so it would be interesting to see what effect these conditions would have on a real video stream. Going one step further it would also be interesting to see the effect on video encoded at a lower rate combined with FEC to see if this improves the end users experience.

Fig. 20. The percentage of occurrences of all paths being in error against the path correlation

Another way would be the investigation of different path selection algorithms. In this paper we used a simple path selection algorithm which selects the path to keep the number of links to a minimum. Other algorithms will be considered and compared in order to reduce the effect of non-congestion related losses, and consequently bring about benefits from applying path diversity. The correlation metric, defined in this paper, can be further validated through experiments on both simulators and on a real world testbed. ACKNOWLEDGEMENT The authors would like to thank the EPSRC and Xyratex for their financial support and guidance without which this would not have been possible. R EFERENCES [1] R. Teixeira, K. Marzullo, S. Savage, and G. M. Voelker, “Characterizing and measuring path diversity of internet topologies,” in SIGMETRICS ’03: Proceedings of the 2003 ACM SIGMETRICS international conference on Measurement and modeling of computer systems. New York, NY, USA: ACM, 2003, pp. 304–305.

[2] T. G. Griffin, F. B. Shepherd, and G. Wilfong, “The stable paths problem and interdomain routing,” IEEE/ACM Transactions on Networking, vol. 10, pp. 232–243, 2002. [3] Network Working Group, Y. Rekhter, and T. Li, “RFC 1771: A border gateway protocol 4 (bgp-4),” March 1995, status: STANDARD. [4] W. Xu and J. Rexford, “Miro: multi-path interdomain routing,” in SIGCOMM ’06: Proceedings of the 2006 conference on Applications, technologies, architectures, and protocols for computer communications. New York, NY, USA: ACM, 2006, pp. 171–182. [5] J. Han, D. Watson, and F. Jahanian, “An experimental study of internet path diversity,” IEEE Transactions on Dependable and Secure Computing, vol. 3, pp. 273–288, October–December 2006. [6] J. G. Apostolopoulos and M. D. Trott, “Path diversity for enhanced media streaming,” IEEE Communications, vol. 42, pp. 80–87, Aug. 2004. [7] D. G. Andersen, H. Balakrishnan, F. Kaashoek, and R. Morris, “The Case for Resilient Overlay Networks,” in 8th Workshop on Hot Topics in Operating Systems, Elmau/Oberbayern, Germany, May 2001. [8] T. Nguyen and S.-C. Cheung, “multimedia streaming using multiple tcp connections,” in IPCCC. IEEE IPCCC 2005, April 2005. [9] A. Begen, Y. Altunbasak, and O. Ergun, “Multi-path selection for multiple description encoded video streaming,” in IEEE International Conference on Communication, vol. 3, 2003, pp. 1583–1589. [10] S. Wu and S. Banerjee, “An area-based feedback implosion control mechanism withdeterministic timeouts,” in IEEE Global Telecommunications Conference, 2001, pp. 1693–1697. [11] S. H. Low, F. Paganini, and J. C. Doyle, “Internet congestion control,” IEEE Control Systems Magazine, pp. 28 – 43, February 2002. [12] A. Medina, A. Lakhina, I. Matta, and J. Byers, “BRITE: Universal topology generation from a user’s perspective,” Boston University, Tech. Rep. 2001-003, 1 2001. [13] R. Haywood and X.-H. Peng, “On packet loss performance under varying network conditions with path diversity,” Aston University, Tech. Rep., 2008. [14] J. Balam and J. D. Gibson, “Path diversity and multiple descriptions with rate dependent packet losses,” Information Theory and Applications Workshop, February 2006. [15] D. D. Clark, “RFC 817: Modularity and efficiency in protocol implementation,” July 1982.

Effect of Path Diversity on the loss performance of UDP ...

number of relay nodes that are positioned around the Internet. ... Service Provider (ISP)) internal network and find that for a ... networks in the wireless environment. ..... [10] S. Wu and S. Banerjee, “An area-based feedback implosion control.

227KB Sizes 0 Downloads 166 Views

Recommend Documents

The Effect of Team Production on Individual Performance
and individual exploitation for overtime-work and long-hour reductions. Variation .... the late work shift, and from 10:00 pm until 6:30 am in the night work shift.

Effect of Schizotypy on Cognitive Performance and Its ...
spatial n-back tests; and cognitive performance speed, which reflects a reduction of the ..... scription of a functional polymorphism and its potential application to.

Effect of Cost and differentiation strategies on firm performance ...
Effect of Cost and differentiation strategies on firm performance article.pdf. Effect of Cost and differentiation strategies on firm performance article.pdf. Open.

The Effect of Crossflow on Vortex Rings
University of Minnesota, Minneapolis, MN, 55414, USA. DNS is performed to study passive scalar mixing in vortex rings in the presence, and ... crossflow x y z wall. Square wave excitation. Figure 1. A Schematic of the problem along with the time hist

The effect of mathematics anxiety on the processing of numerical ...
The effect of mathematics anxiety on the processing of numerical magnitude.pdf. The effect of mathematics anxiety on the processing of numerical magnitude.pdf.

Studying the Effect of Delay Diversity on a DS-CDMA ...
We compare the RAKE error rate performance under situations of no handoff ... DS-CDMA downlink, delay diversity, macro-cellular diversity, micro-cellular diversity, ..... of the total downlink power, and it defines the coverage region of the cell.

The effect of mathematics anxiety on the processing of numerical ...
The effect of mathematics anxiety on the processing of numerical magnitude.pdf. The effect of mathematics anxiety on the processing of numerical magnitude.pdf.

studying the effect of delay diversity on a ds-cdma ...
However, the micro-cellular diversity, as opposed to macro-cellular diversity, .... 3.6 Comparison of BER performance with correlated/independent channel ...... has 30% of the total downlink power, and it defines the coverage region of the cell.

On the Effect of Bias Estimation on Coverage Accuracy in ...
Jan 18, 2017 - The pivotal work was done by Hall (1992b), and has been relied upon since. ... error optimal bandwidths and a fully data-driven direct plug-in.

On the Effect of Bias Estimation on Coverage Accuracy in ...
Jan 18, 2017 - degree local polynomial regression, we show that, as with point estimation, coverage error adapts .... collected in a lengthy online supplement.