QoS-aware Energy-Efficient Algorithms for Ethernet Link ... - UVigo

QoS-aware Energy-Efficient Algorithms for Ethernet Link Aggregates in Software-Defined Networks

Pablo Fondo Ferreiro Miguel Rodríguez Pérez Manuel Fernández Veiga September 15, 2018

1

Context

Context Previous work on Aggregates of Energy Efficient Ethernet Links

Straightforward Solution Power off unused links • Slow response time • What about half used links? 2

EEE Links

Normalized Energy Usage (%)

100

• Formally IEEE 802.3az. • Low Power Idle (LPI) state. • Sleeping and waking up is not instantaneous.

80

60

40

20 EEE Link

0 0.001

0.01

0.1

1

Load

Low Power Mode

Quiet

𝑡r

Quiet

Waking up

𝑡r

........

Refreshing

𝑡s

Quiet

Refreshing

Sleeping

Active

𝑡w

Figure 1: Energy-Efficient Ethernet model. Retrieved from [1].

Active

3

Problem statement Goal Minimize energy consumption in bundles of EEE links leveraging SDN. λ1 λ

λ2

λ3

λ4

Normalized Energy Usage (%)

100

80

60

𝜆𝑖 = 𝜆/4

40

20

0

Equitable share 0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Load

4

Problem statement Goal Minimize energy consumption in bundles of EEE links leveraging SDN. λ1 λ

λ2

λ3

λ4

Normalized Energy Usage (%)

100

80

60

𝑖−1

𝜆𝑖 = min {𝐶, 𝜆 − ∑ 𝜆𝑖 }

40

20

0

𝑘=1

1 Link Bundle Ideal Behavior 0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Load

1-link bundle

5

Problem statement Goal Minimize energy consumption in bundles of EEE links leveraging SDN. λ1 λ

λ2

λ3

λ4

Normalized Energy Usage (%)

100

80

60

𝑖−1

𝜆𝑖 = min {𝐶, 𝜆 − ∑ 𝜆𝑖 }

40

20

0

𝑘=1

1 Link Bundle 2 Link Bundle Ideal Behavior 0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Load

2-link bundle

5

Problem statement Goal Minimize energy consumption in bundles of EEE links leveraging SDN. λ1 λ

λ2

λ3

λ4

Normalized Energy Usage (%)

100

80

60

𝑖−1

𝜆𝑖 = min {𝐶, 𝜆 − ∑ 𝜆𝑖 }

40

1 Link Bundle 2 Link Bundle 4 Link Bundle Ideal Behavior

20

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

𝑘=1

0.9

1

Load

4-link bundle

5

Problem statement Goal Minimize energy consumption in bundles of EEE links leveraging SDN. Theoritical solution Presented in [2], provides a • Packet level algorithm • Assumes real time access to individual occupation of each output port SDN Solution • Needs flow level operation • Cannot take real-time decisions based on queue occupation • Will use ONOS for portability 6

SDN Algorithm

SDN Application

Main Tasks • Flow identification • Flow characterization • Port allocation

7

Flow definition

Challenge Which fields of the packets will identify our flows? • We need: • Enough flows to distribute them along the bundle. • Few flows to keep flow tables small. • Flows with predictable demand.

• Two alternatives: Flow tagging vs field matching. • We will aggregate IP flows: • MAC flows can be insufficient (e.g., transit networks). • Transport flows would be excessive.

8

Flow rate estimation 100

EWMA α = 0.2 EWMA α = 0.4 EWMA α = 0.6 EWMA α = 0.8 previous

estimation error (Mbps)

90 80 70 60 50 40 30 20 10

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

sampling period (seconds)

Figure 2: Average error in the estimation of the flow rate for different periods.

Use rate of previous interval with sampling rate around 0.2 s

9

Port Allocation

In essence, a bin packing problem. Heuristics Greedy Corresponds to first fit decreasing. A flow level water-filling. Bounded Greedy Variation to reduce loses: Maximum usable capacity of a link: 1 − Conservative

𝑏𝑜𝑢𝑛𝑑 |𝑓𝑙𝑜𝑤𝑠|

• Balanced distribution among needed ports. • Safety margin to further avoid losses. • Note: Energy consumption raises very rapidly with traffic load.

10

Conservative Algorithm Behavior • Determines the number of needed links • Distributed flows evenly among the links Basis EEE energy usage rises rapidly with load. 2-bundle link

Normalized Energy Usage (%)

100

80

60

40

20

0

Ideal share Conservative share 0

5

10 Incoming traffic load (Gb/s)

15

20

11

Conservative Algorithm Behavior • Determines the number of needed links

2-bundle link 100 Normalized Energy Usage (%)

• Distributed flows evenly among the links Basis EEE energy usage rises rapidly with load.

80

60

40

20

0

4-bundle link

Ideal share Conservative share 0

5

10

15

20

Incoming traffic load (Gb/s)

Normalized Energy Usage (%)

100

80

60

40

20

0

Ideal share Conservative share 0

5

10

15

20

25

Incoming traffic load (Gb/s)

30

35

40

11

Conservative Algorithm Behavior • Determines the number of needed links

2-bundle link 100 Normalized Energy Usage (%)

• Distributed flows evenly among the links Basis EEE energy usage rises rapidly with load.

80

60

40

20

0

8-bundle link

Ideal share Conservative share 0

5

10

15

20

Incoming traffic load (Gb/s)

100

4-bundle link

Normalized Energy Usage (%)

Normalized Energy Usage (%)

100

80

60

40

80

60

40

20

0

20

0

Ideal share Conservative share 0

5

10

15

20

25

30

35

40

Incoming traffic load (Gb/s)

Ideal share Conservative share 0

10

20

30

40

50

Incoming traffic load (Gb/s)

60

70

80

11

Experimental setup • Topology: Two switches connected by 5 EEE interfaces 10 GBASE-T. • We have used real traffic traces retrieved from CAIDA [3]. • Baseline: Equitable algorithm.

• Metrics: • Energy consumption • Packet losses • Packet delay

12

Results: Energy consumption 98

100

94

energy consumption (%)

real energy consumption (%)

96

92 90

Greedy Bounded-Greedy Conservative Equitable

88 86 84 82

80

Greedy Bounded-Greedy Conservative Equitable

60

40

20

80 78

0

0.1

0.2

0.3

0.4

0.5

0.6

sampling period (seconds)

(a) 32.5 Gbit/s trace.

0.7

0.8

0.9

1

0

6.5

13.0

19.5

26.0

32.5

rate (Gbps)

(b) sampling period = 0.5 seconds.

Figure 3: Normalized energy consumption (buffer = 10000 packets).

• Theoretical bound for the consumption of the 32.5 Gbit/s: 78.5 %. 13

Results: Packet losses

18 16 14

1.6 1.4

12

packet loss (%)

packet loss (%)

1.8

Greedy Bounded-Greedy Conservative Equitable

10 8 6

1.2 1 0.8 0.6

4

0.4

2

0.2

0

10

100

1000 buffer size(packets)

(a) 32.5 Gbit/s trace.

10000

100000

Greedy Bounded-Greedy Conservative Equitable

0

6.5

13.0

19.5

26.0

32.5

rate (Gbps)

(b) buffer = 10000 packets.

Figure 4: Packet loss percentage (sampling period = 0.5 seconds).

14

Results: Packet delay

4500

10000

average delay (microseconds)

average delay (microseconds)

4000 3500 3000 Greedy Bounded-Greedy Conservative Equitable

2500 2000 1500 1000

1000

Greedy B-Greedy Conservative Equitable

100

10

500 0

0

0.1

0.2

0.3

0.4

0.5

0.6

sampling period (seconds)

(a) 32.5 Gbit/s trace.

0.7

0.8

0.9

1

1

6.5

13.0

19.5

26.0

32.5

rate (Gbps)

(b) sampling period = 0.5 seconds.

Figure 5: Average per packet delay (buffer = 10000 packets).

15

QoS-aware algorithms

Problem statement

Goal Provide low-latency service while reducing energy consumption. • The previous algorithms manage to reduce energy consumption. • However, they increase the delay of the packets. • We consider now the QoS latency requirements of the flows. • Two types of traffic: • Best-effort. • Low-latency.

• Modifications to the previous algorithms.

16

Solutions Spare Port

Port 1

low−latency flows best−effort flows

1. Apply energy-efficient algorithm to best-effort flows.

Port 2

2. Low-latency flows are allocated to the most empty port.

Port 3

Port 4

Figure 6: Spare Port.

Two Queues

low−latency flows

high−priority queue

best−effort flows

1. Apply energy-efficient algorithm to all the flows.

Port 1

low−priority queue

Port 2

2. Low-latency flows are allocated to the high-priority queue of the assigned ports.

Port 3

Port 4

Figure 7: Two Queues.

17

Simulations • Same topology: 5-link bundle of 10 GBASE-T EEE interfaces. • Real traces for best-effort traffic. • Synthetic traffic for low-latency packets. • Baseline: Conservative algorithm. • Parameters: • Buffer = 10 000 packets. • Sampling period = 0.5 seconds.

• Metrics: • Delay of low-latency packets. • Delay of best-effort packets. • Energy consumption.

18

Results: Delay of low-latency packets

1000

average delay (microseconds)

average delay (microseconds)

1000

100 Conservative Spare Port Two Queues 10

1

0.1

1

10 low-latency rate (Mbps)

(a) 32.5 Gbit/s trace.

100

1000

100 Conservative Spare port Two queues 10

1

0.1

1

10

100

1000

low-latency rate (Mbps)

(b) 45.5 Gbit/s trace.

Figure 8: Average delay of low-latency packets.

19

Results: Delay of best-effort packets 1100

average delay (microseconds)

1000

Conservative Spare Port Two Queues

900 800 700 600 500 400 300 200

0.1

1

10

100

1000

low-latency rate (Mbps)

Figure 9: Average delay of best-effort packets (32.5 Gbit/s trace).

20

Results: Energy consumption

energy consumption (%)

100

80

Conservative Spare Port Two Queues

60

40

20

0

0.1

1.0

10.0

100.0

1000.0

low-latency rate (Mbps)

Figure 10: Normalized energy consumption (32.5 Gbit/s trace).

21

Conclusions

Conclusions

• SDN can be leveraged to implement energy saving algorithms • Results match theoretical model • Provided low latency service based on QoS requirements Future work • Reuse edge allocations for inner switches. • Reduce control plane traffic (e.g., minimize flow re-allocations).

22

Thank you for listening! Email: [email protected]

References i

S. Herrería-Alonso, M. Rodríguez-Pérez, M. Fernández-Veiga, and C. López-García, “How efficient is energy-efficient ethernet?” in Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), 2011 3rd International Congress on. IEEE, 2011, pp. 1–7. M. Rodríguez Pérez, M. Fernández Veiga, S. Herrería Alonso, M. Hmila, and C. López García, “Optimum Traffic Allocation in Bundled Energy-Efficient Ethernet Links,” IEEE Syst. J., vol. 12, no. 1, pp. 593–603, Mar. 2018. “The CAIDA UCSD Anonymized Internet Traces 2016 — 2016/04/06 13:03:00 UTC.” [Online]. Available: http://www.caida.org/data/passive/passive_2016_dataset.xml