23
Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland Networked Systems Laboratory

Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

Embed Size (px)

Citation preview

Page 1: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

Identifying and Using Energy Critical Paths

Nedeljko Vasić

with Dejan Novaković, Satyam Shekhar, Prateek

Bhurat, Marco Canini, and Dejan Kostić

EPFL, SwitzerlandNetworked Systems Laboratory

Page 2: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

2

NETWORKDatacenter

DATACENTERNETWORK

20% of total serverenergy consumption(3 TWh in US in 2006)

Networking Energy

Several TWh/year for major telcos (Telefonica 4.5 TWh,Verizon 9.9 TWh)

Page 3: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

Causes of networking energy consumption

• Network redundancy– Achieving high availability

• Bandwidth overprovisioning– Tolerate traffic variations (address lack of QoS)

3

Page 4: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

Typical utilization

levels

Energy-(un)proportionality

4

0

20

40

60

80

100

0 20 40 60 80 100

Pow

er (%

of p

eak)

Utilization (%)

Ideal energy-proportionality

Existing networking hardware

Page 5: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

Networking energy outlook

• More demands will result in further increases– Video streaming, Cloud computing

• CMOS reaching a plateau in power-efficiency – Cooling costs of new equipment will increase

• 1 MW for latest Cisco platform, CRS-1

5

Rate of traffic increase > rate in which underlying technologies improve their energy efficiency

Power is not limitless (60 Amps per rack)

Page 6: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

Goal: Energy-Proportional Networked Systems

6

0

20

40

60

80

100

0 20 40 60 80 100

Utilization (%)

Goal

Pow

er (%

of p

eak)

Ideal energy-proportionality

Page 7: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

Make all devices energy-proportional?

7

Utilization (%)

Pow

er (%

of p

eak)

Ideal energy-proportionality

Performance penalties

Always-on components

Page 8: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

8

Dynamically matchresources to the demand make the networkenergy-proportional

Sleeping saves energy

Network-wide energy-proportionality

Page 9: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

Routing table computation

• Goal: The minimal set of network elements (w.r.t. power) that will satisfy the current demands

• Routing that minimizes energy consumption– Multi-commodity flow problem, but with additional

constraints for energy objective:Links + routers (switches) powered on/stand by

– Problem is computationally intensive– Heuristics take several minutes for small topologies

When traffic demand changes, optimal routing changes!

9

Page 10: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

How often is recomputation needed?

10

0

1

2

3

4

5

6

May-28 Jun-1 Jun-5 Jun-9

Dem

and

[Gbp

s]

Time

traffic demands

[Geant2 - European academic network, 15-day trace]

Page 11: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

11

How often is recomputation needed?

B

C

AD

Tim

e15

min

.

Page 12: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

How often is recomputation needed?

12

0

1

2

3

4

5

6

May-28 Jun-1 Jun-5 Jun-9

Dem

and

[Gbp

s]

Time

traffic demands

Routing table recomputed 3-4 times per hour!(state-of-the-art)

[Geant2 - European academic network, 15-day trace]

0

1

2

3

4

5

May-28 Jun-1 Jun-5 Jun-9

Rec

ompu

tatio

n ra

te [p

er h

our]

Time

trace granularity (upper bound)

Page 13: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

13

• Long computation time might lead to energy waste or congestion

• Adjusting routing upon each change in traffic patterns often leads to oscillations

• Network operators would like to base their designs on longer time scales

Issues with recomputation

Page 14: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

14

Can we precompute routing configurations?

B

C

AD

Tim

e15

min

Page 15: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

Can we precompute routing configurations?

15

Too many routing configurations

Fraction of time an energy-optimal routing configuration is used (Geant2 trace)

One routing configuration used

60% of the time

Page 16: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

Energy Critical Paths

16

1 2 3 4 50

20

40

60

80

100

120

GeantFatTree

Number of alternative paths

CDF

of O

ptim

al p

aths

in

clud

ed

Just a few precomputed paths offer near-optimal energy savings

Page 17: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

REsPoNse (Responsive Energy-Proportional Networks)

17

Energy-AwareTraffic Engineering

Runtime Traffic Measurement

[COMSNETS ‘09]

ComputingEnergy-Critical

PathsTraffic Analysis

Online components

Offline components

Ser

vice

-Lev

el O

bjec

tives

Page 18: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

REsPoNse

Always-on paths provide a routing that can carry low to medium amounts of traffic at the lowest energy consumption

On-demand paths start carrying traffic when the load is beyond the capacity offered by the always-on paths

Failover paths are designed tominimize the impact of single failures

18

REsPoNse

REsPoNse-lat

REsPoNse-heuristic

REsPoNse-ospf

Page 19: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

19

REsPoNseTE in action

DestSrcA B

C D

• Online effort to shift traffic to (in)activate on-demand paths• Intermediate routers mark packets with link load • Edge routers collect load info only on alternative paths

Load(AO) > Tr

TE

Page 20: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

Evaluation Questions

• How energy-proportional is REsPoNse?– ISP topologies– Datacenter networks

• How quick is REsPoNseTE in shifting traffic?• What is the impact of traffic aggregation on

application performance?

20

Page 21: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

Responsiveness/Energy-Proportionality (Geant)

21

• Replayed a 15-day trace• 2 power models

– Today– Future (static power is

significantly reduced) 0

20

40

60

80

100

120

May-28 Jun-1 Jun-5 Jun-9

Pow

er [%

orig

inal

net

wor

k po

wer

]

Time

ospfREsPoNse

REsPoNse (Alternative HW model)

0

1

2

3

4

5

6

May-28 Jun-1 Jun-5 Jun-9

Dem

and [G

bps]

Time

traffic demands

REsPoNse saves 30% - 45% with adding only 1 carefully precomputed routing table

P

ower

(% o

f orig

inal

)

100

80

60

40

20

Dem

ands

(GBp

s)

5

4

3

2

1

Page 22: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

Impact on application performance

80

85

90

95

100

105

epr-lat50 InvCap50 epr-lat100 InvCap100

Per

cent

age

(%)

22

0

0.5

1

1.5

2

epr-lat50 InvCap50

Tim

e (s

)

Application performance and e2e latency under REsPoNse is comparable to OSPF at both low and high utilization levels.

Live Modelnet experiment with P2P-VoD (BulletMedia [IPTV ‘07])

REsPoNse OSPF REsPoNse OSPF low utilization high utilization

REsPoNse OSPF low utilization

P

erce

ntag

e (%

) 100

95

90

85

80

Bl

ock

retr.

late

ncy

(s)

1.5

1

0.5

0

Page 23: Identifying and Using Energy Critical Paths Nedeljko Vasić with Dejan Novaković, Satyam Shekhar, Prateek Bhurat, Marco Canini, and Dejan Kostić EPFL, Switzerland

23

Conclusion

• REsPoNse• Key idea: hybrid offline/online approach

based on existence of energy critical paths• Properties

• Stable• Incrementally deployable• Responsive

REsPoNse enables power/cooling for the common case