The EndtoEnd Effects of Internet Path selection

1
The End-to-End Effects
of Internet Path selection

• ECE697A Oct. 2002
• Prof. Lixin Gao
• Presenter Dongha Lee

2
Outline

• Introduction
• Methodology and measurements
• Methodology
• Datasets
• Results
• Robustness
• Evaluation
• Conclusion

3
Introduction.

• How good is Internet routing from a users
  perspective, and why?
• Current
• There are so much diversity in Internet.(bandwith,
  propagation delay, congestion)
• How much of this diversity should be attributed
  to differences in load, differences in capacity,
  or differences in the routing infrastructure.
• Goal
• to understand the degree to which end-to-end
  performance is being determined by the current
  state of Internet routing.
• to understand which mechanisms are responsible.
• Terms
• path and route
• path selection

4
Introduction Terms

• path the complete set of hops traversed
  between two hosts

hop
hop
hop
hop
hop
5
Introduction traceroute example

• from CerfNet Route Server (AS1838) to
  www.yahoo.com
• route-servertraceroute www.yahoo.com
• Translating "www.yahoo.com"...domain server
  (12.129.192.148) OK
• Type escape sequence to abort.
• Tracing the route to www.yahoo.akadns.net
  (66.218.71.80)
• 1 mdf1-gsr12-1-gig-1-1.lax1.attens.net
  (12.129.192.237) AS 17233 0 msec 0 msec 0 msec
• 2 mdf1-gsr12-1-gig-1-1.lax1.attens.net
  (12.129.192.237) AS 17233 0 msec 0 msec 0 msec
• 3 gar3-p320.la2ca.ip.att.net (12.122.255.249)
  AS 7018 4 msec 0 msec 0 msec
• 4 gbr6-p90.la2ca.ip.att.net (12.123.28.197) AS
  7018 0 msec 0 msec 0 msec
• 5 tbr1-p013601.la2ca.ip.att.net (12.122.11.141)
  AS 7018 4 msec 0 msec 4 msec
• 6 ggr1-p340.la2ca.ip.att.net (12.122.11.226)
  AS 7018 4 msec 0 msec 0 msec 56 msec
• 7 so-4-0-1.core1.LosAngeles1.Level3.net
  (209.0.227.45) AS 3356 56 msec 56 msec
• 8 so-4-0-0.mp1.LosAngeles1.Level3.net
  (209.247.10.193) AS 3356 56 msec 56 msec 56
  msec
• 9 so-3-0-0.mp2.SanJose1.Level3.net
  (64.159.1.130) AS 3356 56 msec 56 msec 56 msec
• 10 so-3-0-0-0.ipcolo3.SanJose1.Level3.net
  (64.159.2.41) AS 3356 56 msec 56 msec
• 11 unknown.Level3.net (64.152.69.30) AS 3356
  56 msec 56 msec 56 msec
• 12 www.yahoo.akadns.net (66.218.71.80) AS
  26101 56 msec 56 msec 60 msec

Total 12 hops
6
Introduction Terms

• route the data structures exchanged between
  routers to describe connectivity.

This is Route!
Router A
Router B
7
Introduction Terms

• path selection the combined set of route
  selection decisions made at all the routers in a
  paths.

Path selection
8
Introduction - Examples
waltz.kotel.co.kr
Path B
AS226
AS3847
Route
AS3561
Path A
AS3559
AS2150
Total Path selection 2
rsdist.ra.net
http//www.merit.edu/ipma/npd/npd.as.html
9
Methodology

• Question Is there an alternate path to our
  destination over which we would obtain better
  performance?
• Default path
• it is easy to directly measure the default path
  between two hosts using traceroute or ping and so
  on.
• Alternate path
• ?

10
Methodology
Each host public traceroute server
Host B
Host A
Host C
11
Methodology Potential alternate path

• Potential alternate path
• - We identify alternate paths by constructing a
  weighted graph in which each host is represented
  by a vertex and each path is represented by a
  corresponding edge.
• - And for each pair of hosts, A and B, we remove
  the edge connecting them and perform a
  shortest-path computation between A and B using
  the remaining edges

• Shortest path is then the path with lowest total
  weight (sum of weights of all edges)
• Shortest path not necessarily fewest edges (or
  hops)

12
Methodology
Each host public traceroute server
BA 150ms
Host B
Host A
AB 200ms
AC 75ms
BC 60ms
Best alternate path
CA 65ms
ACCB 125ms
CB 50ms
Host C
13
Characteristics of the dataset

• Whats the Percent of paths covered?
• The number of distinct paths measured divided by
  the number of potential alternate paths that
  could have been measured.
• i.e.

14
Results the mean
round-trip time

• CDF of the difference between the mean round-trip
  time recorded on each path, and the best mean
  round-trip time derived for an alternate path.
• (Recorded Timeon each path
• - Derived Time for an alternate path)

15
Results the mean
round-trip time

• CDF of the difference between the mean round-trip
  time recorded on each path, and the best mean
  round-trip time derived for an alternate path.

• (Red) value below zero are those for which the
  best alternate path was worse than the default
  path.
• (Green) better than the default path.

16
Results the mean loss rate
ratio
17
Results bandwidth about N2,
N2-NA

• Why measure the band-width?
• While the previous graph suggest that there are
  alternate paths with better performance
  characteristics, they do not indicate the amount
  of available bandwidth on these paths.

18
Characteristics of the dataset

• Using the dataset N2! Since N2 measures
  round-trip time and loss rate observed within a
  TCP session.

19
Results bandwidth about N2,
N2-NA

• How can we measure the bandwidth?
• TCP performance is inversely related to latency
  and loss rate.
• Then compute the resulting TCP bandwidth
  according to the TCP model of Mathis et al.

20
Results bandwidth about N2,
N2-NA

• Results
• 70 to 80 percent of the paths have alternates
  with improved bandwidth

21
Robustness

• Several biases in the methodology that might skew
  the results.
• So, need to evaluate the robustness our basic
  finding with respect to four of these factor.
• The use of the mean instead of the median
• Random variation among measurement samples
• Time-of-day dependence
• Long-term averaging of path samples

22
Robustness mean vs
median

• The mean value may be affected if the underlying
  distribution is highly skewed, so the median is
  usually considered a superior statistic.
• Result the difference is negligible.

23
Robustness variation in the
datasets

• All of our measurements demonstrate large ranges
  and consequently, it is possible that the
  difference between the means can be attributed
  largely to random variation in the data.
• Source of random variation
• Upgrades to the network infrastructures during
  the traces
• Path changes (i.e. due to routing policy changes
  or to route flaps)
• congestion

24
Robustness variation in the
datasets
Most paths have relatively tight error bounds
The percentage of paths for which a better
alternate path can be found at the 95 confidence
level represents those paths whose improvement
cannot be well explained simply by variation
25
Robustness time of day
effects

• The Internet as a whole is likely to be more
  congested during peak working hours and less
  congested at night or on weekends
• Results
• The overall effect occurs regardless of the time
  of day.
• Alternate paths seem to do better during times
  known to have heavier load.

26
Robustness long-term averaging of
data

• In order to gauge the effect of this long-term
  averaging, use the UW4-A, for which we measured
  all paths con-currently.
• The results shows that we are slightly more
  likely to be able to find good alternate paths on
  a fine-grained timescale than on a long-term
  timescale.

27
Evaluation

• The 1st hypothesis
• Best alternate paths are caused by avoiding parts
  of the Internet with particularly poor quality
  (e.g. congested exchange point) or by using
  connectivity to parts of the Internet with
  exceptionally good quality.
• The 2nd hypothesis
• Best alternate paths result primarily from
  avoiding congestion, rather than by minimizing
  propagation delay.

28
Evaluation Host AS popularity in alternate
paths

• If it were the case that only a handful of nodes
  were somehow causing the existence of the most of
  the superior alternate paths, then we should be
  able to remove those hosts and see a dramatic
  shift of the CDF curve for the remainder of the
  dataset.

29
Evaluation Host AS popularity in alternate
paths

• Results
• Top ten hosts are not the source of a
  disproportionate number of the best alternate
  paths.
• Conclusion.
• The prevalence of alternate paths with best
  round-trip times cannot be attributed to a small
  number of hosts.

30
Evaluation Host AS popularity in alternate
paths

• The number of times each host appears as an
  inter-mediate host in some superior alternate
  path (not necessarily the very best alternate),
  and weighted by the degree to which the alternate
  path was better than the corresponding default
  path.
• So, we cannot attribute the existence of superior
  alternate paths to a small number of hosts.

31
Evaluation Host AS popularity in alternate
paths

• The effect of ASes in the center of the network,
  rather than individual end hosts.
• Each point represents a single AS in the dataset.
• Results
• The availability of alternate path is not being
  inflated by a small number of either good of poor
  ASes.

This AS appeared same times in both paths
32
Evaluation Congestion vs propagation
delay

• Although we cannot directly measure propagation
  delay, so we can estimate it from our data by
  taking the tenth-percentile of the measured
  round-tip time.
• If congestion is a major source of routing
  inefficiency and of avoiding congested links is a
  major reason for the existence of best alternate
  paths, then two hypothesis should be hold.

33
Conclusion

• For a large number of paths in the Internet,
  there are alternate paths that exhibit superior
  quality as measured by round-trip, loss rate, and
  bandwidth.

34
Question?

• Any question?
• Thanks