Upload
helene
View
38
Download
0
Embed Size (px)
DESCRIPTION
Theophilus Benson *, Aditya Akella *, Aman Shaikh + *University of Wisconsin, Madison + ATT Labs Research. Demystifying Configuration Challenges and Trade -Offs in Network-based ISP Services. Adoption of Network-based Services. ISP Core. Adoption of Network-based Services. - PowerPoint PPT Presentation
Citation preview
1
DEMYSTIFYING CONFIGURATION CHALLENGES AND TRADE-OFFS IN NETWORK-BASED ISP SERVICES
Theophilus Benson*, Aditya Akella*, Aman Shaikh+
*University of Wisconsin, Madison+ ATT Labs Research
2
Adoption of Network-based Services
ISP Core
3
Adoption of Network-based Services
“By 2015, annual global IP traffic will reach 966 exabytes” [Cisco’11]
Customers adopting new applications ISPs upgrade and purchase new equipment Best effort ineffective for some applications
2010 2011 2012 2013 2014 20150E+00
2E+05
4E+05
6E+05
8E+05
1E+06
Time (in Years)
Size
of I
nter
net T
raffi
c(In
PB
)
4
Adoption of Network-based Services
Services are a crucial part of the Internet’s ecosystem
ISP Core
Host-based and network-based services “AT&T to spend $1 billion to ramp up enterprise services” Recover cost and improve application performance
5
Goals
Vision: Improve service integration/upgrades Simplify service management
Understand impediments Service configuration files Complexity of configuration
EdgeEdgeISP Core
6
Configuration is a crucial component
Configuration determines Customer functionality ISP interactions
Configuration is complex Most time consuming task [Feamster ‘05 ] Most error-prone
> 50% customer problems due to configuration errors [Yankee Group ‘04]
OSPF
BGP
7
Configuring a Service
PECEPE CE
acz
ISP Core
Control Plane
Data plane
acz
acz
8
Contributions
Analyzed 2.5 years of configurations Show how complexity evolves over time
Worsens over time Highlight the location of complexity
Complexity exists at the edge Identify the cause of complexity
Due to provisioning of new customers
PEPE
ISP CoreCE
EdgeCustomersEdge
Customers
9
Contributions
Identified potential ways to mitigate complexity Showed the impact of design choices on complexity
Vendor
Com
plex
ity
#1 #2 Routing Design
Com
plex
ity#1 #2 #3
10
Outline
Motivation Background Models and Data-Set Understanding Complexity Mitigating Complexity Conclusion
12
Configuring the Provider’s Edge
Complexity is due to: Dependent commands
PECEPE CE
ISP Core
Ip vrf blueRd 23234:100223Route-target import 1000:1Route-target export 1000:1!Interface ethernet1Ip address 128.105.82.66/30Ip vrf forwarding blueServices-policy output policy1!Policy-map policy1Policy 100 20 confirm-action transmit
VRF
Interface
Policy-map
a
13
Configuring the Control-Plane Core
PECEPE CE
Router bgp 65000Neighbor 129.168.6.6Neighbor 129.168.2.2!
Ip vrf blueRd 23234:100223Route-target import 1000:1Route-target export 1000:1!Router bgp 65000Neighbor 129.168.2.1Neighbor 129.168.2.1
ISP Core
acz
acz
acz
Complexity is due to: Dependent commands Maintaining consistency
14
Configuring the Data-Plane Core
PECEPE CE
acz
acz
Interface gigethernet1Ip address 128.105.82.66/30!Router ospf 2Network 128.105.82.0/24!
Ip vrf blueRd 23234:100223Route-target import 1000:1Route-target export 1000:1!Interface gigethernet1Ip address 128.105.82.65/30!
ISP Corea
Complexity is due to: Dependent commands Maintaining consistency
acz
15
Models and Data-Set
Requirements for Data Models
Quantify complexity of configuration Capture dependencies between commands Capture consistency across devices
Use complexity metrics [Benson ‘09] Motivated by software engineering techniques
Abstract away low level details Abstract groups of commands stanzas
16
Ip vrf blueRd 23234:100223!Interface ethernet1Ip address 128.105.82.66Ip vrf forwarding blueServices-policy input policy2!
17
Data Models
Referential Graph [Benson ‘09] Syntactic dependencies operators must track Network graph of dependent stanzas Metric: size of graph Larger graph more dependencies
Templates [Benson ‘09] Clone detection used to capture uniformity
VRF
Interface
Policy-map
18
Data-Set
Service % of Routers ( PE + Core)VPN 48%VPLS 27%VoIP 5%DDoS Prev. 31%Virtual Wire 25%
Diversity allows for a comprehensive study
Collected data from tier-1 ISP for 2.5 years 5 services: VPN, VPLS, VoIP, DDoS Prev., Virtual Wire Daily snapshots of router configuration files Metadata (per router): vendor, role and location
19
Understanding Complexity Provider Edge (PE) Complexity Control-Plane Core Complexity Data-Plane Core Complexity
PEPE
ISP CoreCE
20
Understanding Complexity Provider Edge (PE) Complexity Control-Plane Core Complexity Data-Plane Core Complexity
PEPEISP Core CE
21
PE Complexity over Time
Growth is due to worsening complexity New devices have less dependencies
Over time, configuration tasks become tricky
Dec 08
Dec 10
0E+0
2E+5
4E+5
6E+5
8E+5
1E+6
DDoS V-Wire VplsVPN VoIP
Time (in Months)
Tota
l # o
f Ref
. Lin
ks
Dec 08
Dec 10
0
500
1000
1500
2000
2500
DDoS V-Wire VplsVPN VoIP
Time (in Months)
Max
PE
Gra
ph S
ize
22
Understanding PE Complexity Which stanza contributes to VPLS growth?
Edge EdgeISP Core
VRF
Interface
Policy-map
Dec 08
Dec 10
0%20%40%60%80%
100%
Interface CoS VRF
Time (in Months)
Perc
enta
ge o
f Sta
nzas
Dec 08
Dec 10
0%20%40%60%80%
100%
Interface CoS VRF
Time (in Months)
Perc
enta
ge o
f Sta
nzas
VRF
23
Understanding PE Complexity Which stanza contributes to VPN growth?
Complexity caused by customer provisioning
Edge EdgeISP Core
VRF
Interface
Policy-map
Dec 08
Dec 10
0%20%40%60%80%
100%
Interface CoS VRF
Time (in Months)
Perc
enta
ge o
f Sta
nzas
Dec 08
Dec 10
0%20%40%60%80%
100%
Interface CoS VRF
Time (in Months)
Perc
enta
ge o
f Sta
nzas
26
Configuration Reuse over Time
Specialization leads to added complexity
• Reuse and specialization exists• Configuration overlap reduces over time
– Reduction due to specialized usage of service
0% 20% 40% 60% 80% 100%0
0.20.40.60.8
1
VPN 2008
Percentage of Reuse
CD
F of
Cus
tom
ers
0% 20% 40% 60% 80% 100%0
0.20.40.60.8
1
VPN 2008 VPN 2010
Percentage of Reuse
CD
F of
Cus
tom
ers
71%
62%88%
27
Understanding Complexity
Data-Plane Core: service-agnostic and simple Control-Plane Core: distinct across services
Growing number of adjacencies with PEs PE is the most complex
PEPE
ISP CoreCE
28
Mitigating Complexity Vendor Selection
Cost
Func
tiona
lity
Complexity
Time
Com
plex
ity
TimeC
ompl
exity
29
How to Compare Vendors
Different vendors different languages Language impacts complexity Difference in structure of functionality
Comparing vendor languages Configurations representing same policy Same customer same policy on all PEs
PECEPE CE
ISP CoreVendor1 Vendor2
30
Vendor Selection
Graph for vendor1 is consistently larger Vendor1 requires more stanzas for same policies Operators need to track more dependencies
Choice of vendor can reduce PE complexity
1 2 3 4 5 6 70.000.501.001.502.002.503.003.504.004.50
Anonymized Customer ID
Rat
io o
f Ref
eren
tial g
raph
(Ven
dor1
/Ven
dor2
)
1 2 3 4 5 6 70.000.501.001.502.002.503.003.504.004.50
Anonymized Customer ID
Rat
io o
f Ref
eren
tial g
raph
(Ven
dor1
/Ven
dor2
)
31
Conclusion
Studied the factors that impede services Complexity grows over time
Modifications become time consuming Most complexity lies in configuring customers
Varying requirements and specialized configuration Framework to systematically consider complexity
Choice of vendor can reduce complexity
32
Thank You
Theophilus Benson ([email protected])
"Complex systems are built out of a myriad of simple components"
33
34
35
Related Works
Customer usage patterns Traffic matrixes [Garimella’07,Qiu’09,] Control plane state [Raghunath’05]
Techniques for optimizing services Improve hardware usage [Kim’08,Raghunath’07]
Very little on configuration Modeling Class-of-Service [Sung’08] Root cause analysis [Mahimkar’10, Turner’10] Complementary to understanding complexity
PECEPE CEISP Core
36
Future Works
Automated simplification of configuration Utilizing insights on design choices
New abstraction layer Policy language for simpler configuration Translates language to configuration
Correlate complexity with manageability
37
38
Customer Provisioning
Customer provisioning causes complexity
Provisioning should slightly increase complexity Similar functionality for all customers
Small change in configurations Modular configuration languages
Reuse of configurations between customers
Dec 08
Dec 10
0E+0
2E+5
4E+5
6E+5
8E+5
1E+6
DDoS V-Wire VplsVPN VoIP
Time (in Months)
Tota
l # o
f Ref
. Lin
ks
39
Similarities Across Customers
Customers are not created equally VPLS: 2 types of customers VPN: Multiple types of customers
Groups of customers use services differently
Differences in requirements introduces complexity
0 5 10 15 20 250
0.2
0.4
0.6
0.8
1
VPLS VPN
Size of Referential Graph
CD
F of
Cus
tom
ers