Transcript Grids, Beowulfs, and Proteins, oh my!

Stealth Multicast - A New Paradigm
for Incremental Multicast
Deployment
Dr. Aaron Striegel
Notre Dame Systems & Software Laboratory
Dept. of Computer Science & Engineering
University of Notre Dame
October 20, 2003
Striegel, Guest Lecture - UIUC ([email protected])
1
Talk Overview
• Information Dissemination
• Motivation
• Stealth Multicast
– Basic Architecture
– Example: MYDEKI
– Preliminary Results
• Future Research
– Inter-domain Peering
– Network stack enhancement
Striegel, Guest Lecture - UIUC ([email protected])
2
Information Dissemination
• Distribute rich content in a timely fashion to users
– Problem: Internet evolved as point-to-point
• Inefficient but currently manageable via unicasts
• Two main approaches
– Active involvement - Multicast
• Close temporal proximity
• Applications, network can participate
– Community participation -> network efficiency
– Passive involvement - Caching
• Multiply-accessed data over time
• No required participation of apps/network
– Exploit existing characteristics of network
» HTTP Caching
» Packet-level caching
Striegel, Guest Lecture - UIUC ([email protected])
3
Caching vs. Multicast
• Caching
– Cannot handle rapidly changing data
• Data w/close temporal proximity
– Easy deployment
• No global participation
• Multicast
– Deployment problems
• Global participation
– Addressed by ALM
» Delay-sensitive traffic, rich user base
• Economic incentive
– Bandwidth glut, ISP benefit
– Can handle close temporal data
• Group-oriented activities - synchrony is an issue
Striegel, Guest Lecture - UIUC ([email protected])
4
Current State
• Caching: Yes
Multicast: ??
– Several recent studies [2000, 2003]
• Lackluster adoption 150 groups (1999) -> 250 (2001)
• Most groups are single source (SSM)
– Why have *, G, CBTs, etc.?
» Harder form of multicast anyway
– Key lesson from caching
• Incremental deployment is key
– Big-bang theory is impossible
– Transition as easy as possible (FUD inertia)
• Immediate benefit
– Large benefit with minimal investment
• Directable economic benefit
– Avoid “If you build it, they will come…”
Striegel, Guest Lecture - UIUC ([email protected])
5
Motivation
• Research premise
– Transparent bandwidth conservation technique
– Change the paradigm of multicast
• Incremental deployment
– Zero dependence on external forces
• Immediate benefit
– Exploit the redundancy in the network
• Economics
– Offer a significant and quantifiable benefit
• Stealth Multicast
– Dynamically convert packets to/from multicast
– Target
• Small to medium group-oriented apps
• Delay-sensitive apps
5-500 users
– On-line gaming, video streaming
– Improve ALM-based apps
Striegel, Guest Lecture - UIUC ([email protected])
6
Stealth Multicast
• Two governing principles
– Externally transparent
• Zero modification - application (server/client)
• Zero modification - external Internet
• Seamless operation
– Negligible QoS impact?
• Should not noticeably impact QoS
• What are noticeable QoS changes?
– Depends upon application
» Large buffer - streaming video
» On-line game - zero buffer
– Informal definition
» Additional delay should not make the application
unusable versus separate unicasts
Striegel, Guest Lecture - UIUC ([email protected])
7
Stealth Multicast Model
Conversion
Other Domains
Servers
ISP Domain
Clients
Company LAN
(Content Provider)
Unicast
Multicast
Striegel, Guest Lecture - UIUC ([email protected])
Unicast
8
Multicast Detection
VGDM - Virtual Group Detection Manager
Virtual
Group
Mgr
Checksum
Calculation
Disp
Tree
Construction
H
Rules
State
Management
Filter
Edge Router
Incoming Traffic
(Unicast only)
Striegel, Guest Lecture - UIUC ([email protected])
Outgoing Traffic
(Unicast+Multicast)
9
Further Examination
• Benefits
– Dynamically convert
– Zero modification
– Multicast transport via
virtual groups
– Exact billing
Benefit
Delay
• Drawbacks
–
–
–
–
Non-zero queuing delay
Aggregation effects
Imperfect virtual groups
Not a universal solution
Striegel, Guest Lecture - UIUC ([email protected])
Virtual Group Delay
Minimum gain
Maximum delay
10
Multicast Transition Options
Approach
Pure
Multicast
Application
Assisted
Customer
Aware
Stealth
Mode
Separate
Unicast
Application
Multicast
Stealth
Unicast
Unicast
Unicast
Internet
Multicast
Unicast
Unicast
Unicast
Unicast
ISP
Multicast
Stealth
Stealth
Stealth
Unicast
None
ISP
Company
ISP Edge
None
Perfect
Perfect
Good
OK
None
Customer,
ISP
Customer,
ISP
Customer,
ISP
ISP
None
App change
Deployment
App change
Deployment
Deployment
Deployment
None
VGDM
Detection
Accuracy
Benefits
Costs
Striegel, Guest Lecture - UIUC ([email protected])
11
Talk Overview
• Information Dissemination
• Motivation
• Stealth Multicast
– Basic Architecture
– Example: MYDEKI
– Preliminary Results
• Future Research
– Inter-domain Peering
– Network stack enhancement
Striegel, Guest Lecture - UIUC ([email protected])
12
MYDEKI Architecture
• Implementation of stealth multicast
– Multicast and You Don’t Even Know It
– Three variations
• ISP
• Local
• Application-assisted
– Defines key issues of stealth multicast
• Triggers - Virtual group release
• Transport - Dynamic groups
• State management - Where to place state
Striegel, Guest Lecture - UIUC ([email protected])
13
Virtual Group Triggers
• Trigger
– Dilemma: Gain for waiting
– When to release the virtual group
• MHT - Maximum Hold Time
• TSW - Time Search Width
• PSW - Packet Search Width
Target packet
MHT
TSW
PSW
Striegel, Guest Lecture - UIUC ([email protected])
time
14
Triggers - Continued
• Triggers/thresholds
– MinGS - Minimum group size
– MaxGS - Maximum group size
– MVG - Maximum virtual groups
New group
Multicast
Unicast
VG 0
MaxGS
MinGS
..
MVG
VG N
Striegel, Guest Lecture - UIUC ([email protected])
15
System Balance
• VGDM Limit
– MVG - Maximum Virtual Groups
• Hard limit - should be avoided
• No multicast benefit - overloaded
– Inputs
• Filter effectiveness
– Eliminate non-candidate traffic
• Triggers - dispatch
– PSW, TSW, MHT
– MaxGS
• Tradeoff
– Capacity, QoS vs. efficiency
Striegel, Guest Lecture - UIUC ([email protected])
16
Multicast Transport
• Issue
– Members (clients) not known a priori
– Dynamically construct tree
• Approaches
– Exhaustive tree construction
• All variations, all egress points
• Problem - failure and tree reconstruction
– Broadcast/hold
• Costly - queuing at edge
• Depends upon client distribution
– Encapsulation-based
• Embed tree inside the packet
• Use DSMCast - DiffServ Multicast
Striegel, Guest Lecture - UIUC ([email protected])
17
Why Encapsulation?
• Benefits
– Dynamically construct trees
• Dynamic routing - QoS
• Heterogeneous QoS
– Stateless core
• Transport only - hardware
– Simplified billing
• Know tree, know cost
• Tradeoff
– Deployment - upgrade core
– Additional bandwidth
• Save 16x vs. 18x vs. None
• Excellent starting point for evaluation
Striegel, Guest Lecture - UIUC ([email protected])
18
State Management
• Issue
– Unique portions of packet
• Compress multiple packets for different destinations into
a single packet
– Dest port, dest IP
– Who is responsible for exporting?
• Egress A vs. Egress B vs. Egress C
• Approaches
– Include in packet
• Similar to encapsulation-based approach
– Distributed knowledge
• Egress points share knowledge
Striegel, Guest Lecture - UIUC ([email protected])
19
Why Packet-based?
• Benefits
– Simple
• No timing/state issues
• Good first step
– No problem with dynamics
• Imprecise matching algorithms
• Weaknesses
– Bandwidth cost
– Limitations on clients per group
• MTU violations
Striegel, Guest Lecture - UIUC ([email protected])
20
Example - State Management
IP
DSM
State UDP
Edge
5
5
10
15
3
Dest IP
129.186.5.4
129.186.6.20
143.24.6.4
102.67.8.25
50.6.8.20
Striegel, Guest Lecture - UIUC ([email protected])
Payload
Dest Port
7894
2004
5003
5796
6901
21
Application-Assisted Method
• Virtual group detection
– Imprecise nature - best guess
• Application-assist
– Application knows about VGDM
– Application sends 1 packet w/state to VGDM
– VGDM constructs tree
• Benefits
– No change to the client - deployment
– Precise group construction
• Issues
– Billing
– Requires change to server app
Striegel, Guest Lecture - UIUC ([email protected])
22
Other Issues - TCP, IPSec, IPv6
• TCP
– Limited benefit
• Why?
– Extra state
– Retransmit of lost packets
– Potential benefit
• Web serving - initial request
– Assume no cookies
• CNN on 9/11
• IPSec / VPNs
• IPv6
Striegel, Guest Lecture - UIUC ([email protected])
23
Simulation Studies
• Simulation setup
– Ns-2 simulator
• Freely available simulator
– GenMCast module for ns-2, GRIM - simulation management
• Network setup
– Medium-sized multicast groups
• On-line gaming apps - 8 to 64 clients
• UDP traffic - 40 server apps
– Compare various approaches
• Based on VGDM location + others
– Local, Stealth, None, App-Assist, ALM
• Evaluation metrics
– Bandwidth savings
– End-user QoS
Striegel, Guest Lecture - UIUC ([email protected])
24
Effect of Clients - Link BW
No savings, unicast
Higher up-link cost
Striegel, Guest Lecture - UIUC ([email protected])
25
Effect of clients - Domain BW
Increasing savings
vs. unicast
Trades B/W for
client B/W
Striegel, Guest Lecture - UIUC ([email protected])
26
Effect of Clients - QoS (Delay)
Limited impact
of queuing
Striegel, Guest Lecture - UIUC ([email protected])
27
Other Results
• Other aspect
– Out of order delivery
• VGDM Overload
• Traffic aggregation
– OS/app effect
• Spacing between packets
Striegel, Guest Lecture - UIUC ([email protected])
28
Talk Overview
• Information Dissemination
• Motivation
• Stealth Multicast
– Basic Architecture
– Example: MYDEKI
– Preliminary Results
• Future Research
– Inter-domain Peering
– Network stack enhancement
Striegel, Guest Lecture - UIUC ([email protected])
29
Immediate Research Areas
• Practical transport
– Encapsulation - interesting but academic
– Compare various approaches
• Fixed grouping w/hierarchy
– How to find the group that maps to the egress points
– Combination of groups
• Broadcast w/hold
– Impact of egress point sparsity
• ALM
– Apply ALM on a domain-wise level
– Fixed vs. dynamic groups
Striegel, Guest Lecture - UIUC ([email protected])
30
Future Research Areas
• Inter-domain peering
– Transparent bandwidth conservation
• Packet caching and stealth multicast
• Edge routers of domains exchange info
– Stealth multicast
• Avoid conversion to/from multicast/unicast
• Construct tree for new domain
– Packet caching
• Share cache in other domains
– Issues
• Billing, QoS
• Resource management
• Protocol / security
Striegel, Guest Lecture - UIUC ([email protected])
31
Future Research Areas
• Network stack modification
– Present: Minimize overhead
• Avoid extra IP/TCP or IP/UDP headers
– Premise
• Can we increase redundancy but increase overall system
performance?
– Enhance network stack
• Add End of Data marker - TCP
• Modify sendmail / Apache to use
– Issues
• Benefit to the network
• Downstream impact -> net system impact
Striegel, Guest Lecture - UIUC ([email protected])
32
Conclusions
• Stealth multicast
– New paradigm for multicast
– Offers several key benefits
• Solves multicast deployment issue
– Zero modification outside of the domain
• Inherent resource management
• Offers directable economic benefit
– Interesting research problems
• Transport, state management
• Inter-domain peering, stack optimization
Striegel, Guest Lecture - UIUC ([email protected])
33
Questions?
[email protected]
http://www.cse.nd.edu/~striegel
GenMCast Package (ns-2)
http://www.cse.nd.edu/~striegel/GenMCast
Striegel, Guest Lecture - UIUC ([email protected])
34
Why Multicast?
• Premise
– Distribute same information to
large number of users
– Simple solution
• Separate unicasts
– Highly inefficient
• N users -> N transmissions
– Example:
• ND Football Game - Streaming audio
– 5,000 listeners - 56k modem - 4kB/s
– Total BW: 20 MB/s
• Videoconference - audio + video
– 50 participants - 250 kB/sec
– Total BW: 12.5 MB/s
Striegel, Guest Lecture - UIUC ([email protected])
35
Multicast
• Operation
– Reduces packet transmission
to an efficient tree
– Relies on network state for
replication
• Benefits
– Reduced bandwidth
• N receivers << N bandwidth
– Bottleneck relief
• Relief close to source
– Simplifies sender management
• Send to group vs. individuals
Striegel, Guest Lecture - UIUC ([email protected])
36
Obstacles/Issues
• Why is multicast being adopted so slowly?
– Chicken or the egg?
• Development/deployment/demand circle
– Global efficiency
• Efficiency - adoption - global scale
– Metcalfe’s Law - Network benefit
• Additional infrastructure
– Killer app
• Does not offer a fundamentally new service
• No killer app - 10x, 100x improvement
– Depends upon deployment to become a killer app
– ISP Economics
Striegel, Guest Lecture - UIUC ([email protected])
37
Multicast - ISP Economics
• Questionable benefit
– ISPs sell bandwidth
– Multicast reduces bandwidth consumption
– ISP perspective
• Pay significantly, sell less
• Billing
– How to charge for multicast
– How to get billing info
• Distributed state of multicast
• Other concerns
– Internet growth myth - investment inertia
– Problem can be solved by more bandwidth
Striegel, Guest Lecture - UIUC ([email protected])
38
DSMCast Operation
C
A
B
D
E
Node
A
B
Replicate
B
D,E
Striegel, Guest Lecture - UIUC ([email protected])
39
Other Issues
• IPSec
– Problem
• Can’t see UDP header - encrypted
– Match checksum - OK
• IPv6
– Two issues
• Security headers
• Address size
– Force egress state
Striegel, Guest Lecture - UIUC ([email protected])
40
Other Research Projects
• Grid computing - GRIM
– Grid Resource Interface & Manager
– Managing simulations
• Script generation, stat management, grid operation
• Scalable datacenters
– Scale datacenter to 10-40 Gb/s+
– Manage train wreck of devices
• FW, LB, IDS, etc.
• DiffServ Multicast
– Heterogeneous receivers
– Resource management
Striegel, Guest Lecture - UIUC ([email protected])
41