Guaranteed Service Level Agreements across Multiple ISP Networks

1
Guaranteed Service Level Agreements across
Multiple ISP Networks

• Panita Pongpaibool
• Hyong S. Kim
• Carnegie Mellon University

2
Outline

• Motivation
• SLA guarantee across multiple ISPs
• Challenges
• Proposed solutions
• Performance evaluation
• Cost modeling
• Conclusion

3
Motivation

• Service level agreements (SLAs) service
  contract between ISPs and customer
• SLA in single ISP is common
• IP backbone guaranteed delay, loss rate
• But a connection is likely to traverse multiple
  ISPs
• Each connection requires end-to-end SLA
  (per-customer demand)
• How could one provide SLA across multiple ISPs?

4
SLA Guarantee across ISPs

• Currently no US backbone ISP offers SLAs beyond
  its network boundary
• Why?
• Undisclosed network information
• BGP masks internal topology details
• No SLA interconnection policy
• No QoS class mapping among ISPs
• Lack of supporting ISP business model
• ISP financial relationship peer, wholesale,
  retail

5
Proposed Solutions
Request 100ms delay s?d
35
30
10
d
20
20
s
y
x
20
30
50
40

• Need a way to allocate customers requirement
  among transit ISPs
• SLA interconnection policies
• Least-effort policy ? selfish
• Most-effort policy ? generous
• Equal-distribution policy ? fair

6
Least-Effort Policy

• Let each ISP determine level of responsibility
  independently
• Expect selfish behaviour
• Each ISP provisions the least expensive path
  (lowest resource usage)
• Burden the last network
• Could cause a lot of unnecessary demand rejections

Request 100ms delay
35
30
10
d
20
20
s
y
20
x
30
50
40
7
Most-Effort Policy

• Force each ISP to take the highest level of
  responsibility
• Each ISP provisions the most expensive path
• Last network is free to choose any suitable path
• Fast resource saturation, especially in the first
  network

Request 100ms delay
35
30
10
d
20
20
s
y
20
x
30
50
40
8
Equal-Distribution Policy

• Allocate equal level of responsibility among
  transit ISPs
• Need to know n (number of transit ISPs)
• EqualThreshold constraint/n (delay, jitter)
• Or EqualThreshold n?constraint (availability,
  reliability)

Request 33ms delay
35
30
10
d
20
20
s
y
x
20
30
50
40
9
Performance Evaluation

• Simulations on a representative US IP backbone
  network
• 3 subnetworks, 3 border nodes
• Uniform optical protection (11, 13,
  unprotected)
• Uniform MPLS traffic demand among ISPs (each
  demand traverses at most 2 ISPs)
• Local SLA provisioning algorithm follows our
  scheme in Globecom03
• SLA parameters
• Bandwidth
• Availability
• Compare to global baseline case

10
Simulation Results 11 protection

• Most-effort worst performance because of early
  bw saturation
• Everything else follows the baseline global
  performance

11
Simulation Results no protection

• Most-effort earliest bandwidth saturation
• Least-effort latest saturation bc prematurely
  rejects request at low load
• Equal-distribution follows global case,
  manageable blocking prob

12
Summary of Findings

• With high intrinsic link availability
• The least-effort policy is most appropriate
• Without high intrinsic link availability
• Most-effort is good at low loads
• Equal-distribution is good at moderate loads
• Least-effort is good only at high loads
• Load-dependent
• But cannot switch policies as load changes
  because will have no bandwidth left
• No single policy works well over all load!

13
ISP Cost Structure

• Reexamine all 3 policies from the aspect of
  financial benefit
• Goal to recommend most profitable policy under a
  specific cost model
• Net profit ?i(Aiconn_acceptedi)
  (Btotal_bw_used)
• Consider
• ISP financial relationship
• Peer-peer ISPs
• Wholesale ISPs
• Retail ISPs
• Owning vs. leasing of network infrastructure
• Degree of price discrimination (in retail ISPs)

cost
revenue
i service levelAi revenue per connB cost
per bw
14
Peer-Peer ISPs

• No monetary exchange ? no revenue!
• Net profit (Btotal_bw_used)
• Plot total profit for all ISPs
• Least-effort is best because it allocates
  bandwidth most sparingly

This is just the upside-down bandwidth usage graph
Least-effort most profitable
15
Wholesale ISPs

• Revenue ? traffic volume, regardless of service
  level
• Net profit A?i(conn_acceptedi)
  (Btotal_bw_used), ? A/B
• Varying ? (ratio of revenue vs. cost)

Large ? (ISP owns the network)
Small ? (ISP leases bandwidth)
Equal-distribution most profitable
revenue dominates(after saturation)
cost dominates
16
Retail ISPs Availability-Based

• Revenue ? level of service availability offered
• Net profit ?i(Aiconn_acceptedi)
  (Btotal_bw_used)
• Varying A1A2Ak for different degree of price
  discrimination
• Small price discrimination ? degenerate to
  wholesale service

Small price discrimination
Equal-distribution most profitable
Least-effort rejects a lot of high-availability
connection!
Large price discrimination
17
Retail ISPs Bandwidth-Based

• Revenue ? level of bandwidth guaranteed
• Net profit ?i(Aiconn_acceptedi)
  (Btotal_bw_used)
• Least-effort is most profitable at high load bc
  other policies have reached saturation and can no
  longer accept new connection

Least-effort most profitable after load 60
Equal-distribution most profitable up to load 60
18
Summary of Findings

• Found two contenders
• Least-effort and equal-distribution
• Least-effort is attractive under the peer model
• Equal-distribution is attractive under
• Wholesale model ISPs own infrastructure
• Retail model availability-based revenue
  discrimination

19
Conclusion

• Introduced three simple policies to coordinate
  local SLA provisioning in multiple ISPs
• Performance evaluation shows no single policy is
  attractive over all loads
• Re-examined the policies under specific ISP cost
  models
• Found two attractive policies least-effort and
  equal-distribution
• Policy choice depends on
• Financial relationship among ISPs
• Relative magnitude of cost and revenue
• Type and degree of price discrimination