Presentation is loading. Please wait.

Presentation is loading. Please wait.

Robert C. Chalmers and Kevin C. Almeroth

Similar presentations


Presentation on theme: "Robert C. Chalmers and Kevin C. Almeroth"— Presentation transcript:

1 Modeling the Branching Characteristics and Efficiency Gains in Global Multicast Trees
Robert C. Chalmers and Kevin C. Almeroth UC Santa Barbara, Computer Science Infocom 2001 my name title work done in conjunction with my advisor

2 web site - http://www.nmsl.cs.ucsb.edu/mwalk/
Goals goal: to characterize multicast efficiency efficiency in relation to duplicate unicast streams goal: to accurately model tree topologies do inter-domain trees share common properties? requirement: a clear understanding of the shape of multicast trees apparent from title - two goals (name them) when talking about efficiency, focused on bandwidth efficiency in relation to unicast mcast is more efficient since packet duplication occurs when paths diverge when looking at tree topologies, mainly interested in finding common properties that accurately describe the majority of inter-domain multicast trees to satisfy either goal we must have a clear understanding of shape and its impact on measures such as efficiency focus on where branching occurs within the tree => to get a better feeling for how shape impact efficiency, we’ll take a closer look at a few key factors of shape 11/29/00 web site -

3 web site - http://www.nmsl.cs.ucsb.edu/mwalk/
Key Factors of Shape height increase in shared links breadth near bottom - sharing near top - duplication number of receivers more likely to share name the factors 11/29/00 web site -

4 web site - http://www.nmsl.cs.ucsb.edu/mwalk/
Key Factors of Shape height increase in shared links breadth near bottom - sharing near top - duplication number of receivers more likely to share height is important when looking at efficiency taller trees are more efficient since more links are shared for unicast - higher number of duplicate packets over those links 11/29/00 web site -

5 web site - http://www.nmsl.cs.ucsb.edu/mwalk/
Key Factors of Shape height increase in shared links breadth near bottom - sharing near top - duplication number of receivers more likely to share where breadth occurs in the tree is important late branching improves the efficiency of tall trees since more receivers are taking advantage of the shared links early branching reduces efficiency since it involves packet duplication 11/29/00 web site -

6 web site - http://www.nmsl.cs.ucsb.edu/mwalk/
Key Factors of Shape height increase in shared links breadth near bottom - sharing near top - duplication number of receivers more likely to share the total number of receivers is a key factor as more receivers join they are more likely to share existing portions of the tree 11/29/00 web site -

7 web site - http://www.nmsl.cs.ucsb.edu/mwalk/
Key Factors of Shape height increase in shared links breadth near bottom - sharing near top - duplication number of receivers more likely to share The bottom line is where does branching occur within the tree and what is its affect on sharing => so how do we go about measuring efficiency? branching where does it occur how does it affect sharing 11/29/00 web site -

8 web site - http://www.nmsl.cs.ucsb.edu/mwalk/
Measuring Efficiency compare total mcast and ucast hops  = 1 - Lm / Lu % gain using mcast Lm = 18, Lu = 38  = / 38 = .52 52% reduction in links traversed (3) (3) (7) (1) (1) (1) (1) (1) (1) (7) (1) (1) (5) (1) in previous work we introduced a simple metric compares total number of multicast hops to the total number of unicast hops in a distribution tree read formula delta represents the percentage gain in bandwidth utilization achieved by using multicast rather than unicast in other words, what percentage of duplicate packets are removed from the network take this tree for example there are 18 links in the tree, therefore L_m = 18 but if you count the number of hops consumed by duplicate unicast packets (shown in red next to each link) you find L_u = 38 calculating delta tells us that in this case multicast provides a 52% reduction in the number of links traversed => but measuring efficiency only goes so far, could we possibly estimate the efficiency of a tree if we new some properties of the tree. (1) (1) (1) (1) 11/29/00 web site -

9 Estimating Efficiency
expands on work to price mcast (Chuang-98) Lm / Lu = Nk k  for a range of generated networks assuming Lu = Lu / N  = 1 - N where  = k - 1  = -0.2 expanding on work done by Chuang and Sirbu to price multicast, we have developed an estimate for multicast efficiency which is a function of the number of group receivers. they presented a power-law (read formula) where L_m = total number of multicast hops L_u (bar) = average number of hops between any two nodes in the net N = total number of receivers k = economies of scale factor for a number of generated topologies they found k be approximately 0.8 now, if we make a simplifying assumption that L_u (bar) = the average distance between the source and each receiver we derive an estimate for delta (read formula) where N is the number of receivers and epsilon is the efficiency factor which should be approx -0.2 => so, this provides us with a power-law characterization of mcast efficiency which depends only on the number of receivers 11/29/00 web site -

10 Estimating Efficiency (cont’d)
20-40 receivers  60-70% improved efficiency 80% savings for 150 users, 90% for 1,000 the characterization is interesting it tells us how efficient a inter-domain multicast trees of varying sizes should be we see that within a small number of receivers, mcast out-performs ucast by a wide margin. on this graph we see a plot for three values of epsilon the value of -0.2 as derived from Chuang and Sirbu and values and which is the range we observed for real multicast groups but, jumping ahead just a little => how could we be sure that this estimate was representative of real multicast groups C and S used mainly generated network topologies and randomly distributed receiver distributions 11/29/00 web site -

11 web site - http://www.nmsl.cs.ucsb.edu/mwalk/
Efficiency Analysis we collected RTCP and mtrace data for a number of multicast sessions reconstructed the sessions off-line to measure efficiency as well as collect topology info such as degree distributions we already knew that the characterization held for generated tress the idea was to validate the efficiency characterization for real groups and then loosen both temporal and spatial constraints to ensure that the characterization held throughout the space between => first step was to determine whether receiver duration had an effect on multicast efficiency 11/29/00 web site -

12 web site - http://www.nmsl.cs.ucsb.edu/mwalk/
Receiver Duration 1-minute timeout early clustering 20-minute timeout infinite timeout starting with a small 1-minute timeout while maintaining the original inter-arrival times for each receiver we calculated the metric and plotted it against the estimate as we see, the data fits quite well an interesting effect can be seen between 5-10 receivers where efficiency is higher than expected due to receiver clustering a number of receivers join the group from a single sub-net then we extended the timeout to 5, 20 minutes and eventually to the length of the session receivers still joined according to the original data, but once in the group they did not leave data fits quite well duration does not effect efficiency => next step was to randomize the original arrival times 11/29/00 web site -

13 web site - http://www.nmsl.cs.ucsb.edu/mwalk/
Inter-Arrival Times random receiver activity by ignoring duration and randomizing inter-arrival times, we ignore the time domain entirely the data fits the estimate very well an interesting effect of randomization is a reduction in the magnitude of the efficiency factor from 0.34 to 0.30 due to smoothing over of temporal clustering which normally improves efficiency => next step, stretch the spatial domain || drops from 0.34 to 0.30 11/29/00 web site -

14 Receiver Distribution
synthesized group distribution To randomize receiver distribution we generated a synthesized dataset with almost 2000 receivers random IPs were chosen and traces were made through the Internet using a local source faithfully represented the inter-domain multicast topology data fits extremely well this means that it doesn’t matter which source/receivers you choose, efficiency is similar || still lower than for real groups 11/29/00 web site -

15 web site - http://www.nmsl.cs.ucsb.edu/mwalk/
Efficiency and Shape do consistent efficiencies imply similar shapes? global mcast trees share common properties long paths from source to backbone branching in backbone only occurs at a limited number of peering points long paths from backbone to receiver the range of tree shapes are constrained by the underlying network connectivity efficiencies may be similar, but does that imply the shapes of all these trees are similar intuitively speaking they should be for inter-domain multicast trees, most receivers are necessarily in other domains than the source so, for the source to reach the receivers it must traverse its own domain, the backbone then the domains of each receiver this should result in tall trees with most branching occurring at a limited number of peering points within the backbone in general, we’ve found that the range of possible tree shapes are constrained by the underlying network connectivity => so, in a way we’ve come full circle. It’s now a question of branching and where the network will allow it to occur within the tree. 11/29/00 web site -

16 Out-degree Frequencies
most nodes have a very low out-degree result is chains of relay nodes (Pansiot-98) degree frequencies in Internet routers follow a skewed distribution (Faloutsos-99) so, if we talk about branching, then we are really looking at out-degree what we see is that the majority of nodes (almost 80%) of a multicast tree have only a single out-going link this results in long chains of relay nodes which simply forward the packet along a single path generally trees are more likely to be tall than wide similarly, routers in the Internet have been shown to have a skewed degree distribution again, the underlying network dictates where branching can occur => so, our intuition tells us that the majority of the branching should occur at a few well-defined places in the back-bone and at the leaf routers connecting directly to receivers 11/29/00 web site -

17 Average Degree average degree grows with N
and this is indeed what we witness in general, the total average degree throughout the tree grows with the number of receivers, N but, if we separate internal links from those directly connecting receivers we see that the rise in average degree is mainly due to receivers clustering around leaf routers the body of the tree stays fairly consistent and has an average degree very near 1.5 the backbone has a limited number of candidate links that can be grafted into the tree as the number of receivers grow, those links become saturated new growth occurs mainly at leaf routers which is good for multicast efficiency internal degree tapers off near 1.5 11/29/00 web site -

18 web site - http://www.nmsl.cs.ucsb.edu/mwalk/
Conclusions accurately characterized mcast efficiency over a range of receiver dynamics and distributions useful in effort to deploy multicast identified common properties of mcast trees useful to improve tree generation techniques shape is constrained by network connectivity for more information achieved our two goals accurately characterized multicast efficiency identified common properties of inter-domain multicast trees we’ve shown that our efficiency estimate performs well over a wide range of receiver dynamics and distributions this should be very useful in the push to deploy multicast as a predictor of future efficiency gains we’ve identified a number of properties that could be used to improve tree generation techniques and have validated the use of randomly placed sources and receivers throughout a realistic network topology most importantly, we have shown that the shape of real multicast trees are constrained by the underlying network topology if you are interested in more info, please visit our website any questions 11/29/00 web site -

19 Addendum

20 web site - http://www.nmsl.cs.ucsb.edu/mwalk/
Unicast vs. Multicast ratio of ucast to mcast path lengths for 1198 receivers in the SYNTH-1 dataset majority of mcast paths are times the ucast path length 11/29/00 web site -


Download ppt "Robert C. Chalmers and Kevin C. Almeroth"

Similar presentations


Ads by Google