Transcript Hitesh Ballani

Application layer multicast
routing
Key:
What Is Multicast?
Unicast transfer
Broadcast transfer
Multicast transfer
• Unicast
– One-to-one
– Destination – unique receiver
host address
• Broadcast
– One-to-all
– Destination – address of
network
• Multicast
– One-to-many
– Multicast group must be
identified
– Destination – address of group
Multicast application examples
• Financial services
– Delivery of news, stock quotes, financial indices,
etc
• Remote conferencing/e-learning
– Streaming audio and video to many participants
(clients, students)
– Interactive communication between participants
• Data distribution
IP Multicast
Gatech
Stanford
CMU
Berkeley
Routers with multicast support
•Highly efficient bandwidth usage
Key Architectural Decision: Add support for multicast in IP layer
So what is the big issue …
15 years since proposal, but no wide area IP multicast
deployment
• Scalability (with number of groups)
-- Routers maintain per-group state
-- Require every group to dynamically obtain a globally unique
address from the multicast address space
• IP Multicast: best-effort multi-point delivery service
-- Providing higher level features such as reliability, congestion
control, flow control, and security has shown to be more difficult
than in the unicast case
•
Other issues
-- Change in infrastructure
-- DOS attacks
-- Network management, billing etc.
-- works across space, not across time
Can we achieve efficient multi-point delivery without IP-layer support?
Application layer multicast
Gatech
Stanford
Stan1
Stan2
CMU
Berk1
Berkeley
Overlay Tree
Berk2
Stan1
Gatech
Stan2
CMU
Berk1
Berk2
Pros and Cons
• Scalability
– Routers do not maintain per-group state
– End systems do, but they participate in very few groups
– No need for globally consistent naming, allows application specific
naming
• Easier to deploy
– No infrastructural support
• Potentially simplifies support for higher level functionality
– Leverage computation and storage of end systems
– Leverage solutions for unicast congestion, error and flow control
• Efficiency concerns
– redundant traffic on physical links
– increase in latency due to end-systems
Open questions!
What are the performance implications of using an overlay?
How do end systems with limited topological information
cooperate to construct good overlay structures?
What metric should the optimization be based on?
End-to-end overlay vs. proxy-based overlay?
Does high bandwidth data dissemination require special attention?
A lot of work has been done to tackle these questions, for eg.
Narada, Yoid, Scattercast, ALMI, Overcast, ROMA, NICE, Bullet,
Selectcast, Scribe .…
Overcast : Motivation
• Offering bandwidth-intensive content
on demand
- primarily video content
- necessary to maintain full fidelity
• Long-running content availability for
multiple clients
- data distribution system for
businesses
• Bandwidth bottlenecks develop as
multiple requests are made
Key:
Line thickness indicates
desired bandwidth usage
by sending entity
What is Overcast?
• An application level multicasting system
• Provides scalable and reliable single-source multicast
• Goals
 Overlay structured to maximize bandwidth
 Utilize network topology efficiently
- limit repeated usage of physical links
 No change to existing routers
- easy deployment
• Draws upon work in content distribution, caching, and
server replication
Key insight and contribution’s
• Add storage to the network fabric for reliability
and scalability
- use disk-space to time shift multicast distribution
- trade-off between disk-space and bandwidth
• Contribution’s
 a simple protocol for forming efficient and scalable
distribution trees that dynamically adapt to changes
 a protocol for maintaining global status at the root of
the changing distribution tree
System structure
The overlay comprises of :
• A central source (may be replicated for fault tolerance)
• A no of overcast nodes (standard PCs with lot’s of
storage)
- organized into a distribution tree rooted at the source
- bandwidth efficient trees
• Final Consumers – members of the multicast group
- allows unmodified HTTP clients to join
The business model involves a content provider who installs these
nodes (proxies) in the fabric and the overlay acts as a data
distribution system for businesses
Bandwidth Efficient Overlay Trees
1
10 Mb/s
R
2
“…three ways of organizing the root and the nodes into a distribution tree.”
1
R
R
2
1
2
R
2
1
How Does Overcast Build
Bandwidth Efficient Trees?
• Goal – maximize bandwidth to root for all nodes
• Places a new node as far away from root as possible
without sacrificing bandwidth
• Additional details
– Bandwidth: measured by timing a 10KB download
– Hystersis: nodes with bandwidth within 10% of each other
considered equal
– in case of a tie, choose the closest parent as determined by
traceroute
• Won’t the results of this algorithm change over time?
The node addition algorithm
R
R
5
10
1
10
3
8
1
2
7
5
2
Physical network substrate
3
Overcast distribution tree
But ….
•
What is the 10 Kb download measuring?
 TCP begins with a ‘slow start’
 Download is over by the time it switches to AIMD
 Does not give a measure of long-term TCP throughput
•
Where is the bottleneck bandwidth?
 Nodes on the edges -- access links (leads to linear trees)
 Nodes in the core
- fat pipes (a 10 Kbps download does not give the
bandwidth)
- bottleneck due to congestion, which can vary on a short
time scale
 May have been more relevant at the time of the work ….
 The impact of the child nodes on the bandwidth ….
Dynamic Topology
• A node periodically reevaluates its position in the tree
by measuring the bandwidth been itself and
– its parent (baseline)
– its grandparent
– all its siblings
• Node can relocate to become
– child of a sibling
– sibling of a parent
• Inherently tolerant of non-root failures
– if the parent dies, a node moves up the ancestry tree
Interactions Between Node Adding
And Dynamic Topology
R
10
R
20
1
R
2
2
1
15
2
Physical network substrate
1
Overcast distribution tree Overcast distribution tree
Round 1
Round 2
But ….
• What happens when bandwidth keeps fluctuating?
 10% hysterisis gap is ineffective given the way the
bandwidth is measured
 Smart choice of bandwidth’s for the network edges in the
generated graph.
 No evaluation in case of bandwidth flux
 Even if the system does converge, the convergence time is
limited by the probes to siblings and ancestors
State tracking – the Up/Down protocol
• An efficient mechanism is needed to exchange
information between nodes
– must scale sublinearly in terms of network usage
– may scale linearly in terms of storage
• Assumes information either
– changes slowly (E.g., Health status of nodes)
– can be combined efficiently from multiple children into a
single description (E.g., Totals from subtrees)
• Each node maintains state about all nodes in its
subtree
Management of information with the
Up/Down protocol
• Each node periodically contacts its parent
• Parents assume a child (and all descendants) has died
if the child fails to contact within some interval
• During contact, a node reports to its parent
–
–
–
–
Death certificates
Birth certificates
Extra information
Information propagated from children
• Sequence numbers used to prevent race conditions
Scaling sublinearly in
terms of network usage
1
• A node (and descendants)
relocates under a sibling
• The sibling must learn about all
the node’s descendants
– Birth certificates
• The sibling passes this
information to the (original
node’s) parent
1.1
1.2.1
1.2
1.2.2
1.2.2.1
• The parent recognizes no
changes and halts further
propagation
1.3
1.2.3
No change
observed.
Propagation
halted.
Birth
certificates
for 1.2.2,
1.2.2.1
The client side – how to join a multicast
group
• Clients join a multicast group through a typical HTTP GET
request
• Root determines where to connect the client to the
multicast tree using
– Pathname of URL (multicast group being joined)
– Status of Overcast nodes
– Location of client*
• Root selects “best” server and redirects the client to that
server
Key:
Client joins
Content query
(multicast join)
Query redirect
Content delivery
R1
R2
1
3
2
4
R3
5
6
Evaluation
• Based on simulations with GT-ITM
– Five 600-node graphs
• 3 transit domains (backbone)
• 8 stub networks per domain
• 25 nodes per stub
– Assigned Bandwidth
• 45Mbps, 1.5Mbps, 100Mbps
• T3, T1, Fast Ethernet
– One node supports 20 clients
• (MPEG-1 video)
Bandwidth utilization
• Backbone
– Adds transit nodes
first
• Random
– All nodes chosen
randomly
• Fraction = Overcast bandwidth/Optimal bandwidth
• At full participation – distribution trees are different
Discussion (cont.)
•
Is a tree structure effective for high bandwidth data dissemination?
 Kostic et. al. differ!
 Bullet achieves this by distributing data in a disjoint manner to strategic
points - the nodes organize as a mesh
•
For good performance, the content provider will have to manually choose
strategically placed nodes in the core which are anyway connected by big fat
pipes – so what is the relevance of the provided automation?
•
What are the reliability semantics?
- what if a node is down while a transmission completes – how does the log
help, does the node receive the file, does the root know of this?
- when is data removed from the storage?
•
Not suitable for live streams
- although the authors were never aiming for this!
Discussion (cont.)
•
What about NAT’s
- how can any overcast node be inside a NAT?
•
In case of a tie in bandwidth measure, why not use latency instead of hop
count – leads to a shortest widest kind of selection
•
Where does the root get the location of client from, when doing server
selection?
•
Simply coupled TCP connections not the best way to realize the bandwidth
potential of the topology
- May be a problem in case of heterogeneous receivers
- ROMA uses loosely coupled connections with fast forward error correction
•
Seems more like a contribution to smarter content distribution systems than to
application level multicasting
 Bit Torrent does the same thing!
Enabling Conferencing Applications
on the Internet using an
Overlay Multicast Architecture
Past Work
• Self-organizing protocols
– Yoid (ACIRI), Narada (CMU), Scattercast (Berkeley), Overcast
(CISCO), Bayeux (Berkeley), …
– Construct overlay trees in distributed fashion
– Self-improve with more network information
• Performance results showed promise, but…
– Evaluation conducted in simulation
– Did not consider impact of network dynamics on overlay
performance
Focus of This Paper
• Can End System Multicast support real-world
applications on the Internet?
– Study in context of conferencing applications
– Show performance acceptable even in a dynamic and
heterogeneous Internet environment
• First detailed Internet evaluation to show the
feasibility of End System Multicast
Enhancements of Overlay Design
• Two new issues addressed
– Dynamically adapt to changes in network conditions
– Optimize overlays for multiple metrics
• Latency and bandwidth
• Study in the context of the Narada protocol (Sigmetrics
2000)
– Techniques presented apply to all self-organizing protocols
Optimize Overlays for Dual Metrics
Source rate
2 Mbps
60ms, 2Mbps
Receiver X
Source
30ms, 1Mbps
• Prioritize bandwidth over latency
• Break tie with shorter latency
Adapt to Dynamic Metrics
• Adapt overlay trees to changes in network condition
– Monitor bandwidth and latency of overlay links
• Link measurements can be noisy
– Aggressive adaptation may cause overlay instability
persistent:
react
bandwidth
transient:
do not react
raw estimate
smoothed estimate
discretized estimate
time
• Capture the long term performance of a link
– Exponential smoothing, Metric discretization
Evaluation overview
Compare performance of their scheme with
-- Benchmark (IP Multicast)
-- Other overlay schemes that consider fewer network metrics
Overlay Scheme
Choice of Metrics
Bandwidth
Bandwidth-Latency
Bandwidth-Only
Latency-Only
Random
Latency
Three Scenarios Considered
Primary Set
1.2 Mbps
Primary Set
2.4 Mbps
Extended Set
2.4 Mbps
(lower)  “stress” to overlay schemes  (higher)
• Does ESM work in different scenarios?
• How do different schemes perform under various
scenarios?
BW, Primary Set, 1.2 Mbps
Internet pathology
Naïve scheme performs poorly even in a less “stressful” scenario
RTT results show similar trend
Scenarios Considered
Primary Set
1.2 Mbps
Primary Set
2.4 Mbps
Extended Set
2.4 Mbps
(lower)  “stress” to overlay schemes  (higher)
• Does an overlay approach continue to work under a
more “stressful” scenario?
• Is it sufficient to consider just a single metric?
– Bandwidth-Only, Latency-Only
BW, Extended Set, 2.4 Mbps
no strong correlation between
latency and bandwidth
Optimizing only for latency has poor bandwidth performance
RTT, Extended Set, 2.4Mbps
Bandwidth-Only cannot avoid
poor latency links or long path length
Optimizing only for bandwidth has poor latency performance
Discussion
• Does a source rate of 2.4 Mbps represent a realistic setting!
• Issues not addressed
 Scalability
- how to lower network costs for large sized groups
- how to make the tree building algorithm more scalable
 Extremely dynamic environments
- how to achieve shorter time scale adaptation
- trade-off between overlay stability and the detection time
• Optimizing for 2 dynamic metrics
- Optimizing for both bandwidth and latency may be tricky; the
discretization and hysterisis combined with the probe latency may
affect the granularity of adjustment
Discussion (cont)
• Is end-system multicast suitable for real time applications?
 Does provide distribution across time
 But for applications with timing constraints, it may not be
such a good idea!
 These are some of the most well connected machined on the
Internet!
• Is the internet a good place for such measurements and
comparisons?
 The authors do a decent job of making the analysis fair and
objective
On a general note ….
• What is the impact of such overlays on the other traffic?
• The usefulness of the overlay concept
 Will they be a success iff nobody uses them?
 Interaction between multiple independent overlays
• What is the motivation for their deployment?
 Users come in and go out !
 Cheating by end hosts
• All the ALM’s provide a best effort service!
 are they comparable?
 which one provides the best best-effort service