Transcript PPT - EECS

VPNs and
Network Management
EECS 489 Computer Networks
http://www.eecs.umich.edu/courses/eecs489/w07
Z. Morley Mao
Monday, April 2, 2007
Acknowledgement: Some slides taken from Kurose&Ross
VPN Taxonomy
VPN
Network
End-to-end
Provider-based
Provider-based
Customer-based
Customer-based
L3
L2
ATM
Frame Relay
LAN
What is a VPN?
 Making a shared network look like a private
network
 Why do this?

Private networks have all kinds of advantages
• (we’ll get to that)

But building a private network is expensive
• (cheaper to have shared resources rather than
dedicated)
History of VPNs
 Originally a telephone network concept

Separated offices could have a phone system
that looked like one internal phone system
 Benefits?
 Fewer
digits to dial
 Could have different tariffs
• Company didn’t have to pay for individual long distance
calls
 Came
with own blocking probabilities, etc.
• Service guarantees better (or worse) than public
phone service
Original data VPNs
 Lots of different network technologies in those
days



Decnet, Appletalk, SNA, XNS, IPX, …
None of these were meant to scale to global proportions
Virtually always used in corporate settings
 Providers offer virtual circuits between customer
sites


Frame Relay or ATM
A lot cheaper than dedicated leased lines
 Customer runs whatever network technology over
these
 These still exist (but being replaced by IP VPNs)
Advantages of original data
VPNs
 Repeat: a lot cheaper than dedicated
leased lines
Corporate users had no other choice
 This was the whole business behind frame-relay
and ATM services

 Fine-grained bandwidth tariffs
 Bandwidth guarantees
 Service Level Agreements (SLA)
 “Multi-protocol”
How has the world changed?
 Everything is IP now

Some old stuff still around, but most data
networks are just IP
 So, why do we still care about VPNs???
IP VPN benefits
 IP not really global (private addresses)

VPN makes separated IP sites look like one
private IP network
 Security
 Bandwidth guarantees across ISP
 QoS, SLAs
 Simplified network operation
 ISP
can do the routing for you
End-to-end VPNs
 Solves problem of how to connect remote
hosts to a firewalled network
Security and private addresses benefits only
 Not simplicity or QoS benefits

End-to-end VPNs
 Solves problem of how to connect remote
hosts to a firewalled network
FW/
VPN
Internet
IPsec
Tunnels
Remote
Host
Remote
Host
Site (private
network)
Site
Host
Site
Host
Provider-based end-to-end VPNs
 Used for instance when enterprise pays for
employee access, wants it to go through
enterprise network
I know Cisco did this
 But never used that much

• Business model didn’t take off

Used even less now
• In part because VPN client comes with windows OS?
 The tunneling technology commonly used
for roaming dialup though
Customer-based Network VPNs
Site
Site
CE
CE
Internet
CE
CE
Site
Site
Customer buys own equipment, configures IPsec
tunnels over the global internet, manages addressing
and routing. ISP plays no role.
Customer-based Network VPNs
 Great for enterprises that have the
resources and skills to do it

Large companies
 More control, better security model
 Doesn’t
require trust in ISP ability and
intentions
 Can use different ISPs at different sites
 But not all enterprises have this skill
Provider-based Network VPNs
Site
Site
CE
CE
PE
PE
ISP
PE
PE
CE
CE Site
Site
Provider manages all the complexity of the VPN.
Customer simply connects to the provider equipment.
Model for customer
 Attach to ISP router (PE) as though it was
one of your routers
 Run routing algorithm with it

OSPF, RIP, BGP
 PE will advertise prefixes from other sites
of same customer
Various PPVPN issues
 Provider provisioned VPNs
 Tunnel type?
 IPsec (more secure, more expensive)
 GRE etc.
 How to discover which customer is at which
PE?

Don’t want PEs without given customer to
participate in routing for that customer
 How to distinguish overlapping private
address spaces
Chapter 9: Network Management
Chapter goals:
 introduction to network management
 motivation
 major components
 Internet network management framework
 MIB: management information base
 SMI: data definition language
 SNMP: protocol for network management
 security and administration
 presentation services: ASN.1
Chapter 9 outline
 What is network management?
 Internet-standard management framework
 Structure of Management Information: SMI
 Management Information Base: MIB
 SNMP Protocol Operations and Transport Mappings
 Security and Administration
 ASN.1
What is network management?
 autonomous systems (aka “network”): 100s or 1000s
of interacting hardware/software components
 other complex systems requiring monitoring, control:
 jet airplane
 nuclear power plant
 others?
"Network management includes the deployment, integration
and coordination of the hardware, software, and human
elements to monitor, test, poll, configure, analyze, evaluate,
and control the network and element resources to meet the
real-time, operational performance, and Quality of Service
requirements at a reasonable cost."
Infrastructure for network management
definitions:
managing entity
managing
entity
agent data
data
network
management
managed devices contain
managed device managed objects whose
data is gathered into a
Management Information
agent data
Base (MIB)
managed device
protocol
agent data
agent data
managed device
managed device
Network Management standards
OSI CMIP
 Common Management
Information Protocol
 designed 1980’s: the
unifying net
management standard
 too slowly
standardized
SNMP: Simple Network
Management Protocol
 Internet roots (SGMP)
 started simple
 deployed, adopted rapidly
 growth: size, complexity
 currently: SNMP V3
 de facto network
management standard
Chapter 9 outline
 What is network management?
 Internet-standard management framework
 Structure of Management Information: SMI
 Management Information Base: MIB
 SNMP Protocol Operations and Transport Mappings
 Security and Administration
 ASN.1
SNMP overview: 4 key parts
 Management information base (MIB):

distributed information store of network
management data
 Structure of Management Information (SMI):
 data definition language for MIB objects
 SNMP protocol
 convey manager<->managed object info, commands
 security, administration capabilities
 major addition in SNMPv3
SMI: data definition language
Purpose: syntax, semantics of
management data welldefined, unambiguous
 base data types:
 straightforward, boring
 OBJECT-TYPE
 data type, status,
semantics of managed
object
 MODULE-IDENTITY
 groups related objects
into MIB module
Basic Data Types
INTEGER
Integer32
Unsigned32
OCTET STRING
OBJECT IDENTIFIED
IPaddress
Counter32
Counter64
Guage32
Time Ticks
Opaque
SNMP MIB
MIB module specified via SMI
MODULE-IDENTITY
(100 standardized MIBs, more vendor-specific)
MODULE
OBJECT TYPE:
OBJECT TYPE:OBJECT TYPE:
objects specified via SMI
OBJECT-TYPE construct
SMI: Object, module examples
OBJECT-TYPE: ipInDelivers
ipInDelivers OBJECT TYPE
SYNTAX
Counter32
MAX-ACCESS read-only
STATUS current
DESCRIPTION
“The total number of input
datagrams successfully
delivered to IP userprotocols (including ICMP)”
::= { ip 9}
MODULE-IDENTITY: ipMIB
ipMIB MODULE-IDENTITY
LAST-UPDATED “941101000Z”
ORGANZATION “IETF SNPv2
Working Group”
CONTACT-INFO
“ Keith McCloghrie
……”
DESCRIPTION
“The MIB module for managing IP
and ICMP implementations, but
excluding their management of
IP routes.”
REVISION “019331000Z”
………
::= {mib-2 48}
MIB example: UDP module
Object ID
Name
Type
Comments
1.3.6.1.2.1.7.1
UDPInDatagrams Counter32 total # datagrams delivered
at this node
1.3.6.1.2.1.7.2
UDPNoPorts
Counter32 # underliverable datagrams
no app at portl
1.3.6.1.2.1.7.3
UDInErrors
Counter32 # undeliverable datagrams
all other reasons
1.3.6.1.2.1.7.4
1.3.6.1.2.1.7.5
UDPOutDatagrams Counter32 # datagrams sent
udpTable
SEQUENCE one entry for each port
in use by app, gives port #
and IP address
SNMP Naming
question: how to name every possible standard object
(protocol, data, more..) in every possible network
standard??
answer: ISO Object Identifier tree:
 hierarchical naming of all objects
 each branchpoint has name, number
1.3.6.1.2.1.7.1
ISO
ISO-ident. Org.
US DoD
Internet
udpInDatagrams
UDP
MIB2
management
OSI
Object
Identifier
Tree
Check out www.alvestrand.no/harald/objectid/top.html
SNMP protocol
Two ways to convey MIB info, commands:
managing
managing
entity
entity
request
response
agent data
Managed device
request/response mode
trap msg
agent data
Managed device
trap mode
SNMP protocol: message types
Message type
GetRequest
GetNextRequest
GetBulkRequest
InformRequest
SetRequest
Response
Trap
Function
Mgr-to-agent: “get me data”
(instance,next in list, block)
Mgr-to-Mgr: here’s MIB value
Mgr-to-agent: set MIB value
Agent-to-mgr: value, response to
Request
Agent-to-mgr: inform manager
of exceptional event
SNMP protocol: message formats
SNMP security and administration
 encryption: DES-encrypt SNMP message
 authentication: compute, send MIC(m,k):
compute hash (MIC) over message (m),
secret shared key (k)
 protection against playback: use nonce
 view-based access control
SNMP entity maintains database of access
rights, policies for various users
 database itself accessible as managed object!

Chapter 9 outline
 What is network management?
 Internet-standard management framework
 Structure of Management Information: SMI
 Management Information Base: MIB
 SNMP Protocol Operations and Transport Mappings
 Security and Administration
 The presentation problem: ASN.1
The presentation problem
Q: does perfect memory-to-memory copy
solve “the communication problem”?
A: not always!
struct {
char code;
int x;
} test;
test.x = 256;
test.code=‘a’
test.code
test.x
a
00000001
00000011
host 1 format
test.code
test.x
a
00000011
00000001
host 2 format
problem: different data format, storage conventions
A real-life presentation problem:
grandma
2004 teenager
aging 60’s
hippie
Presentation problem: potential solutions
1. Sender learns receiver’s format. Sender translates
into receiver’s format. Sender sends.
2. Sender sends. Receiver learns sender’s format.
Receiver translate into receiver-local format
3. Sender translates host-independent format. Sends.
Receiver translates to receiver-local format.
Solving the presentation problem
1. Translate local-host format to host-independent format
2. Transmit data in host-independent format
3. Translate host-independent format to remote-host
format
grandma
aging 60’s
hippie
2004 teenager
ASN.1: Abstract Syntax Notation 1
 ISO standard X.680
used extensively in Internet
 like eating vegetables, knowing this “good for you”!

 defined data types, object constructors
 like SMI
 BER: Basic Encoding Rules
 specify how ASN.1-defined data objects to be
transmitted
 each transmitted object has Type, Length, Value
(TLV) encoding
TLV Encoding
Idea: transmitted data is self-identifying
T: data type, one of ASN.1-defined types
 L: length of data in bytes
 V: value of data, encoded according to ASN.1
standard

Tag Value
1
2
3
4
5
6
9
Type
Boolean
Integer
Bitstring
Octet string
Null
Object Identifier
Real
TLV
encoding:
example
Value, 259
Length, 2 bytes
Type=2, integer
Value, 5 octets (chars)
Length, 5 bytes
Type=4, octet string
Network Management: summary
 network management
 extremely
important: 80% of network “cost”
 ASN.1 for data description
 SNMP protocol as a tool for conveying
information
 Network management: more art than science
 what to measure/monitor
 how to respond to failures?
 alarm correlation/filtering?
Examples of network management
 Challenges
 Reacting quickly to alleviate congestion
 Avoiding over-reacting and causing oscillations
 Limiting bandwidth & CPU overhead on routers
 Load-sensitive routing
 Routers adapt to link load in a distributed
fashion
 At the packet level, or on “group of packets”
 Traffic engineering
 Centralized
computation of routing parameters
 Network-wide measurements of offered traffic
Do IP Networks Manage
Themselves?
 TCP congestion control
 Senders react to
congestion
 Decrease sending rate
 But… the TCP sessions
receive lower throughput
 IP routing protocols
 Routers react to
failures
 Compute new paths
 But… the new paths
may be congested
2
2
4
2
1
1
4
1
3
1
5
4
3
2
1
3
1
5
4
3
Do IP Networks Manage
Themselves?
 In some sense, yes:
 TCP senders send less traffic during congestion
 Routing protocols adapt to topology changes
 But, does the network run efficiently?
 Congested link when idle paths exist?
 High-delay path when a low-delay path exists?
2
2
1
4
4
1
1
3
3
2
2
1
1
5
5
4
3
4
3
1
Adapting the Routing to the
Traffic
 Goal: modify the routes to steer traffic
through the network in most effective way
 Approach #1: load-sensitive protocols
Distribute traffic & performance
measurements
 Routers compute paths based on load

 Approach #2: adaptive management system
 Collect measurements of traffic and topology
 Management system optimizes the parameters
 Debates still today about the right answer
Load-Sensitive Routing Protocols
 Advantages
Efficient use of network resources
 Satisfying the performance needs of end users
 Self-managing network takes care of itself

 Disadvantages
 Higher overhead on the routers
 Long alternate paths consume extra resources
 Instability from out-of-date feedback
information
Packet-Based Load-Sensitive
Routing
 Packet-based routing

Forward packets based on forwarding table
 Load-sensitive
 Compute table entries based on load or delay
 Questions
What link metrics to use?
 How frequently to update the metrics?
 How to propagate the metrics?
 How to compute the paths based on metrics?

Original ARPANET Algorithm
(1969)
 Routing algorithm
 Shortest-path
routing based on link metrics
 Instantaneous queue length plus a constant
 Distributed shortest-path algorithm (BellmanFord)
2
3
2
1
1
3
5
1
20
congested link
Performance of ARPANET
Algorithm
 Light load
 Delay dominated by transmission & propagation
 So, link metrics don’t fluctuate much
 Medium load
 Queuing delay is no longer negligible
 Moderate traffic shifts to avoid congestion
 Heavy load
 Very high metrics on congested links
 Busy links look bad to all of the routers
 All routers avoid the busy links
 Routers may send packets on longer paths
Problem: Out-of-Date Information
Lincoln Tunnel
NJ
NYC
Holland Tunnel
“Backup at Lincoln” on radio triggers congestion at Holland
 Routers make decisions based on old information
 Propagation delay in flooding link metrics
 Thresholds applied to limit number of updates
 Old information leads to bad decisions
 All routers avoid the congested links
 … leading to congestion on other links
 … and the whole things repeats
Problem: Frequent Updates
 Update messages
 Link keeps track of its metric (e.g., queuing delay)
 Link transmits updates when the metric changes
 Frequency of updates
 Frequent changes to the metric lead to frequent updates
 Significantly increases the overhead of the protocol
 Oscillation makes the problem worse
 Oscillation leads to wild swings in the link metrics
 Forcing very frequent update messages
 … that add to the load on the links in the network
Second ARPANET Algorithm
(1979)
 Link-state protocol


Old: Distributed path computation leads to loops
New: Better to flood metrics and have each router
compute the shortest paths
 Averaging of the link metric over time

Old: Instantaneous delay fluctuates a lot

New: Averaging reduces the fluctuations
 Reduce frequency of updates

Old: Sending updates on each change is too much

New: Send updates if change passes a threshold
Problem of Long Alternate Paths
 Picking alternate paths
Long path chosen by one router consumes
resource that other packets could have used
 Leads other routers to pick other alternate
paths

 Solution: limit path length
 Bound the value of the link metric
 “This link is busy enough to go two extra hops”
 Extreme case
 Limit path selection to the shortest paths
 Pick least-loaded shortest path in the network
Load-Sensitive Routing
 Timescales
 What timescale of routing decisions?
 What timescale of feedback about link loads?
 Load-sensitive routing at packet level
 Routers receive feedback on load and delay
 Routers re-compute their forwarding tables
 Fundamental problems with oscillation
 Load-sensitive routing for groups of packets
 Routers receive feedback on load and delay
 Router compute a path for the next flow or circuit
 Less oscillation, as long as circuits last for a while
Reducing Effects of Out-of-Date
Info
 Send link metrics more often

But, leads to higher overhead

But, propagation delay is a fundamental limit
 Make the traffic last longer

Route on groups of packets, rather than packets

Fewer routing decisions, and more accurate feedback
 Groups of packets

Telephone network: phone call (3-minutes long)

Internet: TCP connection (10-packets long)

Internet: all traffic between a pair of hosts, or routers,
…
Traffic Engineering as a
Network-Management Problem:
Case Study
Using Traditional Routing
Protocols
 Routers flood information to learn topology

Determine “next hop” to reach other routers…

Compute shortest paths based on link weights
 Link weights configured by network
operator
2
3
2
1
1
1
3
5
4
3
Approaches for Setting the Link
Weights
 Conventional static heuristics

Proportional to physical distance
• Cross-country links have higher weights
• Minimizes end-to-end propagation delay

Inversely proportional to link capacity
• Smaller weights for higher-bandwidth links
• Attracts more traffic to links with more capacity
 Tune the weights based on the offered
traffic
Network-wide optimization of the link weights
 Directly minimize metrics like max link
utilization

Example of Tuning the Link
Weights
 Problem: congestion along the pink path
 Second or third link on the path is overloaded
 Solution: move some traffic to the bottom path
 E.g., by increasing the weight of the second link
2
3
2
1
1
31
3
5
4
3
Measure, Model, and Control
Network-wide
“what if” model
Offered
Topology/
traffic
Configuration
measure
Changes to
the network
control
Operational network
Traffic Engineering Problem
 Topology
 Connectivity and capacity of routers and links
 Traffic matrix
 Offered load between points in the network
 Link weights
 Configurable parameters for routing protocol
 Performance objective
 Balanced load, low latency, service level
agreements …
 Question: Given the topology and traffic
matrix, which link weights should be used?
Key Ingredients of the Approach
 Instrumentation
Topology: monitoring of the routing protocols
 Traffic matrix: fine-grained traffic
measurement

 Network-wide models
 Representations of topology and traffic
 “What-if” models of shortest-path routing
 Network optimization
 Efficient algorithms to find good configurations
 Operational experience to identify key
constraints
Formalizing the Optimization
Problem
 Input: graph G(R,L)
R is the set of routers
 L is the set of unidirectional links
i
 cl is the capacity of link l

 Input: traffic matrix

Mi,j is load from router i to j
 Output: setting of the link weights
 wl is weight on unidirectional link l
 Pi,j,l is fraction of traffic from i to j traversing
link l
j
Multiple Shortest Paths: Even Splitting
0.25
0.25
0.5
1.0
0.25
0.5
1.0
0.25
0.5
0.5
Values of Pi,j,l
Defining the Objective Function
 Computing the link utilization
ul = Si,j Mi,j Pi,j,l
 Utilization: ul/cl

Link load:
 Objective functions


min (maxl(ul/cl))
min(Sl f(ul/cl))
f(x)
x
1
Complexity of the Optimization Problem
 Computationally intractable problem
 No
efficient algorithm to find the link
weights
 Even for simple objective functions
 What are the implications?
 Must
resort to searching through weight
settings
Optimization Based on Local
Search
 Start with an initial setting of the link
weights

E.g., same integer weight on every link

E.g., weights inversely proportional to capacity

E.g., existing weights in the operational network
 Compute the objective function

Compute the all-pairs shortest paths to get Pi,j,l

Apply the traffic matrix Mi,j to get link loads ul

Evaluate the objective function from the ul/cl
 Generate a new setting of the link weights repeat
Making the Search Efficient
 Avoid repeating the same weight setting
 Keep track of past values of the weight setting
 … or keep a small signature of past values
 Do not evaluate setting if signatures match
 Avoid computing shortest paths from
scratch
 Explore
settings that changes just one weight
 Apply fast incremental shortest-path
algorithms
 Limit number of unique link-weight values
 Don’t explore 216 possible values for each
weight
 Stop early, before exploring all settings
Incorporating Operational Realities
 Minimize number of changes to the
network

Changing just 1 or 2 link weights is often
enough
 Tolerate failure of network equipment
Weights usually remain good after failure
 … or can be fixed by changing 1-2 weights

 Limit effects of measurement accuracy
 Good weights remain good, despite noise
 Limit frequency of changes to the weights

Joint optimization for day & night traffic
matrices
Application to AT&T’s Backbone
 Performance of the optimized weights

Search finds a good solution within a few minutes

Much better than link capacity or physical distance

Competitive with multi-commodity flow solution
 How AT&T changes the link weights

Maintenance every night from midnight to 6am

Predict effects of removing link(s) from network

Reoptimize the link weights to avoid congestion

Configure new weights before disabling equipment
Example from AT&T’s Operations
Center
 Amtrak repairing/moving part of train track
 Need to move some of the fiber optic cables
 Or, heightened risk of the cables being cut
 Amtrak notifies AT&T the timework will be done
 AT&T engineers model the effects
 Determine which IP links go over affected fiber
 Pretend the network no longer has these links
 Evaluate the new shortest paths and traffic flow
 Identify whether link loads will be too high
Example Continued
 If load will be too high
Reoptimize the weights on the remaining links
 Schedule time for new weights to be configured
 Roll back to old weights when Amtrak is done

 Same process applied to other cases
 Assessing the network’s risk to possible
failures
 Planning for maintenance of existing equipment
 Adapting link weights to installation of new
links
 Adapting link weights in response to traffic
shifts
What About Interdomain Routing?
 Border Gateway Protocol
Announcements carry very limited information
 E.g., AS path, but nothing about delay, loss, etc.

 Challenging to make load-sensitive protocol
 Hard to agree upon a common metric
 Hard to scale to such a large network
 Hard to prevent ASes from gaming the system
 Instead, individual ASes act alone
 Change routing policies based on link load
 E.g., moving some traffic to another provider
Interdomain Traffic Engineering
 Predict effects of changes to import policies

Inputs: routing, traffic, and configuration data

Outputs: flow of traffic through the network
Topology
Externally
learned
routes
BGP policy
configuration
BGP routing
model
Offered
traffic
Flow of traffic through the network
Outbound Traffic: Pick a BGP Route
 Easier to control than inbound traffic
 IP routing is destination based
 Sender determines where the packets go
 Control only by selecting the next hop
 Border router can pick the next-hop AS
 Cannot control selection of the entire path
Provider 1
“(1, 3, 4)”
Provider 2
“(2, 7, 8, 4)”
Outbound Traffic: Shortest AS Path
 No import policy on border router
 Pick route with shortest AS path
 Arbitrary tie break (e.g., smallest
router-id)
 Performance?
 Shortest AS path is not necessarily best
 Could have high delays or congestion
 Load balancing?
 Could lead to uneven split in traffic
 E.g., one provider with shorter paths
 E.g., too many ties with skewed tie-break
Outbound Traffic: Load Balancing
 Selectively use each provider
Assign local-pref across destination prefixes
 Change the local-pref assignments over time

 Useful inputs to load balancing
 End-to-end path performance data
• E.g., active measurements along each path

Outbound traffic statistics per destination
prefix
• E.g., packet monitors or router-level support
Link capacity to each provider
 Billing model of each provider

Balancing Load, Performance, and Cost
 Balance traffic based on link capacity
 Measure outbound traffic per prefix
 Select provider per prefix for even load splitting
 But, might lead to poor performance and high bill
 Balance traffic based on performance
 Select provider with best performance per prefix
 But, might lead to congestion and a high bill
 Balance traffic based on financial cost
 Select provider per prefix over time to minimize the
total financial cost
 But, might lead to bad performance
Conclusions
 Adapting routing to the traffic
 To
alleviate congestion
 To minimize propagation delay
 To be robust to future failures
 Two main approaches
 Load-sensitive
routing protocol
 Optimization of configurable parameters