1 Limiting the Impact of Failures on Network Performance Joint work with Supratik Bhattacharyya, and Christophe Diot High Performance Networking Group,

1

High PerformanceSwitching and RoutingTelecom Center Workshop: Sept 4, 1997.

Limiting the Impact of Failures on

Network Performance

Joint work withSupratik Bhattacharyya, and Christophe Diot

High Performance Networking Group, 25 Feb. 2004

Yashar GanjaliComputer Systems Lab.Stanford University

[email protected]

http://www.stanford.edu/~yganjali

2

Motivation

The core of the Internet consists of several large networks (IP backbones).

IP backbones are carefully provisioned to guarantee low latency and jitter for packet delivery.

Failures occur on a daily basis as a result of Physical layer malfunction, Router hardware/software failures, Maintenance, Human errors, …

Failures affect the quality of service delivered to backbone customers.

3

Outline

Background Sprint’s IP backbone Data

Impact Metrics Time-based metrics Link-based metrics

Measurements Reducing the impact

Identifying critical failures Causes analysis Reducing critical failures

4

Background – Sprint’s IP backbone

IP layer operates above DWDM with SONET framing. IS-IS protocol used to route traffic

inside the network. IP-level restoration

When an IP link fails, all routers in the network independently compute a new path around the failure

No protection in the underlying optical infrastructure.

5

Data

IS-IS Link State PDU logs Collected by passive listeners from Sprint’s

North America backbone. Feb. 1st, 2003 to Jun. 30th, 2003.

SNMP logs Link loads recorded once in every 5

minutes. SONET layer alarms

Corresponding to minor and major problems in the optical layer

We are only interested in two alarms:SLOS, and SLOS cleared.

6

Link Failures in Sprint’s IP Backbone – 9408 Failures

7

Inter-POP vs. Intra-POP

ANA-2

ANA-3

ANA-1

ANA-4

8

Outline

Background Sprint’s IP backbone Data

Impact Metrics Time-based metrics Link-based metrics

Measurements Reducing the impact

Identifying critical failures Causes analysis Reducing critical failures

9

Inter-POP Link Failures in Sprint’s IP Backbone

10

Two Perspectives

For a given impact metric Time-based analysis: Measure the

impact of failures on the given metric as a function of time.

Link-based analysis: Measure the impact of failures on the given metric as a function of failing links.

11

Time-based Impact Metrics

1. Number of Simultaneous Link Failures

2. Number of affected O-D pairs3. Number of affected BGP prefixes4. Path unavailability5. Total rerouted traffic6. Maximum load

12

Number of Simultaneous Failures

13

Number of Simultaneous Failures

14

1. Number of Simultaneous Link Failures

2. Number of affected O-D pairs3. Number of affected BGP prefixes4. Path unavailability5. Total rerouted traffic6. Maximum load

Time-based Impact Metrics

15

Number of Affected O-D Pairs

A C F

B

D E

16

Number of Affected O-D Pairs

17

1. Number of Simultaneous Link Failures

2. Number of affected O-D pairs3. Number of affected BGP

prefixes4. Path unavailability5. Total rerouted traffic6. Maximum load

Time-based Impact Metrics

18

Number of Affected BGP Prefixes

19

Time-based Impact Metrics

1. Number of Simultaneous Link Failures

2. Number of affected O-D pairs3. Number of affected BGP prefixes4. Path unavailability5. Total rerouted traffic6. Maximum load

20

Path Unavailability

A C F

B

D E

21

Path Unavailability

22

Time-based Impact Metrics

1. Number of Simultaneous Link Failures

2. Number of affected O-D pairs3. Number of affected BGP prefixes4. Path unavailability5. Total rerouted traffic6. Maximum load

23

Total Rerouted Traffic

24

Time-based Impact Metrics

1. Number of Simultaneous Link Failures

2. Number of affected O-D pairs3. Number of affected BGP prefixes4. Path unavailability5. Total rerouted traffic6. Maximum load

25

Maximum Load Throughout the Network

26

Maximum Load Throughout the Network

96% of link failures were

not followed by an immediate

change in maximum load.

27

Time-based Impact Metrics

1. Number of Simultaneous Link Failures

2. Number of affected O-D pairs3. Number of affected BGP prefixes4. Path unavailability5. Total rerouted traffic6. Maximum load

28

Number of Failures per Link

29

Number of Affected OD Pairs per Link

30

Number of Affected BGP Prefixes per Link

31

Path Coverage

A C F

B

D E

32

Path Coverage of Links

33

Total Rerouted Traffic on a Link

34

Peak Factor of a Link

35

Link-based Impact Metrics

1. Number of Link Failures2. Number of affected O-D pairs3. Number of affected BGP prefixes4. Path coverage5. Total rerouted traffic6. Peak factor

36

Outline

Background Sprint’s IP backbone Data

Impact Metrics Time-based metrics Link-based metrics

Measurements Reducing the impact

Identifying critical failures Causes analysis Reducing critical failures

37

Critical Failures

For each time-based metric Removing failures occuring during 1-

5% of time improves the metrics by a factor of at least 5.

For each link-based metric Removing failures on 1-7% of links

improves the metric by a factor of at least 3.

38

Critical Time Periods

39

Critical Links

Any link which has a critical failures, is called a Critical Link.

We are interested in fixing such links.

40

Correlation of Critical Sets

41

Correlation of the Critical Sets

Metric Size 1 2 3 4 5 6 7 8 9 10

1) Simultaneous failures 11 -0.38

0.33

0.27

0.23

0.11

0.13

0.08

0.15

0.05

2) # of O-D pairs 9 - -0.37

0.21

0.25

0.12

0.14

0.06

0.09

0.06

3) # of BGP prefixes 6 - - -0.18

0.32

0.09

0.05

0.10.07

0.03

4) Path unavailability 5 - - - -0.41

0.14

0.11

0.08

0.12

0.04

5) Total rerouted traffic 6 - - - - -0.09

0.11

0.09

0.08

0.08

6) # of failures 2 - - - - - -0.29

0.31

0.25

0.17

7) # of O-D pairs 3 - - - - - - -0.29

0.30.18

8) # of BGP prefixes 2 - - - - - - - -0.13

0.19

9) Path coverage 6 - - - - - - - - -0.08

10) Total rerouted traffic 1 - - - - - - - - - -

Overall 23% of all links are critical.

42

Cause Analysis

Markopoulou et al. have used IS-IS update messages for characterizing link failures into the following categories [MIB+04]. Maintenance Unplanned

• Shared failures– Router-related– Optical-related– Unspecified

• Individual failures

About 70% of all unplanned

failures

43

Matching SLOS Alarms with IP Link Failures

Time IP link failure

SLOS~

20ms

SLOS Cleared~

12sec

58% of all link failures are due to optical layer problems.

84% of critical failures are due to optical layer problems.

44

Reducing Critical Failures

Replace old optical fibers/parts.

Optical Protection.

Push the traffic away. Also works for maximum load and

peak factor.

45

Performance Improvement

Time-based metrics Link-based MetricsMetric

% improvement

Metric%

improvement

# of failures# of affected O-D pairs

# of BGP prefixesPath unavailability

Total rerouted traffic

4136323929

# of failures# of affected O-D pairs

# of BGP prefixesPath coverage

Total rerouted traffic

4537294238

46

Reducing Link Down-time

Low-failure links: Failure are very rare. Damping doesn’t help.

High-failure links: Failure rate changes very slowly. Fixed damping is wasteful.

47

Adaptive Damping

Input: : time difference between the last two failures: threshold: constant

function Adaptive_Dampingbegin

if ( < )ADT := x ;

elseADT := 0;

end;

Output:ADT: Adaptive damping timer

48

Number – Duration Pareto Curve

49

Thank you!