ECOC 2009, 20-24 September, 2009, Vienna, Austria
Paper 5.5.2
Energy Efficient Multilayer Traffic Engineering Bart Puype, Didier Colle, Mario Pickavet, Piet Demeester Ghent University – IBBT – IMEC, Ghent, Belgium,
[email protected] Abstract Simulation shows how energy efficiency aware multilayer traffic engineering taking into account network equipment power characteristics can optimize IP-over-optical networks for diurnal traffic variations, lowering total power requirements. Introduction
Diurnal traffic variations
Automatically switched multilayer IP-over-optical networks offer extensive flexibility in setting up endto-end lightpaths dynamically. These are used to construct the logical IP topology. Multilayer traffic engineering (MLTE) copes with varying traffic patterns, relying on online IP logical topology reconfiguration in addition to the more traditional rerouting. Although MLTE typically optimizes towards resource usage, QoS performance etc., energy efficiency has become equally important. Multilayer traffic engineering strategy The MLTE strategy first presented in [1] uses an IP layer cost function not only taking care of IP flow routing, but also logical topology construction. This is done by first routing the offered traffic on a virtual IP layer full mesh (a technique modeling the extensive flexibility of dynamic lightpath setup). The shortestpath routing however uses an IP link load dependent cost function (Fig. 1), which grooms traffic onto a subset of this virtual full mesh. Logical topology adaptation is then done by identifying unused edges in the virtual full mesh, and subsequently removing them from the actual IP logical topology. The load-dependent cost function parameters are identified in [1]. Of notice is low-load threshold (LLT), which serves to assign a higher cost for IP links having a load below this threshold, causing traffic 20
60
When looking at MLTE to increase energy efficiency, we can make the following observations. Firstly, router idle power is a constant power which cannot be optimized. Secondly, router client interfaces simply carry the offered traffic pattern into the network, power Pmax Pidle = 0.9 Pmax
Pidle = 0.25 Pmax
50
LLT
bw
40
Cost
# links
Power characteristics
70 # links cost
15
10
30 HLT
MC
20 LMR
5
LC:MC
LC
Although MLTE is suited for very large traffic variations, a typical pattern seen in access as well as core networks is a daily alternation between peak and off-peak volumes. In [2], the rate of bottom vs. peak volume is 50%. Depending on the type of network and devices accessing it, this rate may be 20% for WiFi networks, or even 10% for a network servicing smart phones [3]. Obviously, a large part of the background off-peak volume is always-on traffic such as P2P services. However, even on wired access network were such services are prevalent, diurnal variations are increasing with time [2]. Peak (day) traffic largely consists of more interactive services, which need increasingly larger bandwidths (e.g., consider gaining popularity of on-demand streaming video), making for a stronger diurnal traffic oscillation.
10
0
0 0
20
40 60 Load (%)
80
100
Fig. 1: IP/MPLS MLTE routing cost function
flows to be deviated away from them, until they can be removed from the logical topology mesh. The result is a logical topology suited to the offered traffic pattern, with an IP link load distribution as shown on the example histogram included on Fig. 1.
978-3-8007-3173-2 © VDE VERLAG GMBH
bwmax
Fig. 2: Interface power vs. bandwidth characteristics
without room for traffic engineering. This leaves the router trunk interfaces, whose traffic volume is for a large extent determined by the multi-hop routing behavior of the MLTE strategy. As the power requirement for such interfaces is not constant but depends on the carried traffic, the MLTE scheme can increase energy efficiency through optimized routing and logical topology construction. Interfaces found in commercial L3 routing equipment have an idle power requirement (Pidle) of around 90% of the maximum power rating (Pmax) [4]. For L2 Ethernet switching
ECOC 2009, 20-24 September, 2009, Vienna, Austria
equipment, a reduction in energy consumption by a factor of four is possible, simply by adjusting linerates according to carried traffic on the Ethernet interface [5]. We can therefore assume more energy efficient interfaces to have a Pidle of 25% of Pmax. This leads to interface power characteristics as shown on Fig. 2.
Paper 5.5.2
We assume two types of MLTE regimes. The first, ‘slow’, has logical topology updates slower than the diurnal variations (with a bottom/peak rate of 25%). For the ‘fast’ variant, we can perform MLTE actions in-between peak and off-hour periods. 45
On Fig. 3, we show the influence of the LLT parameter on resulting total power requirement of the logical topology and routing solution for the same sample traffic pattern over 14 nodes. Power requirement is normalized against Pmax. The typical value of LLT is 0.2 [1]. By varying LLT, the MLTE strategy will groom more or less aggressively; higher LLT yields longer multi-hop paths and vice versa.
High idle power
35
Low idle power
30 25 20 15 10 5 0
Regular MLTE, LLT = 0.4
50
MLTE with power metric, LLT = 0.4
Fig. 4: Comparison of MLTE performance without and with power metric, for various regimes
40
Ptot/Pmax
Total power Ptot/Pmax
40
Influence of LLT parameter
30 p_idle = 0.0 p_idle = 0.2 p_idle = 0.4 p_idle = 0.6 p_idle = 0.8 p_idle = 1.0
20 10 0 0
0.1
0.2
0.3
0.4
LLT
Fig. 3: Interface power vs. bandwidth characteristics
For equipment with high Pidle, higher LLT leads to a sparser logical topology and lower energy consumption. However, the exact opposite is the case when using interfaces offering low idle power requirements. These are very efficient at low bandwidth utilization rates, and therefore it is OK to set up a lot of lightly loaded IP links in the logical topology. In fact, multi-hop routing merely increases total transit traffic volume in this case, which explains the higher total power requirement. Power based metric From this, there is clearly room for specific optimization by somehow modifying the MLTE strategy. For the case of optical layer resource usage, we used a so-called optical metric in [1]. In the study presented here, we repeat this approach for a power based metric: we take the interface power characteristic (Fig. 2) and multiply it with the original cost function (Fig. 1) to reach a new cost function yielding automatic energy efficiency optimization. On Fig. 4 we simulate the regular MLTE approach (bars on the left), and the power metric enabled MLTE scheme (right) for a 14 node network, where 4 nodes have power efficient interfaces (Pidle = 25% of Pmax). Source-destination traffic is uniformly distributed between 0 and 50% maximum link bandwidth.
978-3-8007-3173-2 © VDE VERLAG GMBH
The lighter (top) part of each bar shows power required by interfaces with high Pidle. It can be seen that the power metric enabled scheme tends to avoid and shut down these inefficient interfaces, especially at off-peak (‘low’) volumes. Although the slow regime offers no reconfiguration at off-peak volumes, avoiding inefficient interfaces leads to better scaling of total power requirement with traffic volume. For the fast regime and regular vs. power metric, a reduction of 12% in power is seen for off-peak hours. The difference for peak traffic is minimal as no interfaces are run near their idle operating point, so differences in power efficiency are minimal also. Conclusions Energy efficiency aware MLTE serves to traffic engineer around, and shut down power inefficient parts of the network, leading to lower total energy consumption. Even for slow response regimes, perinterface power considerations lead to logical topologies scaling better with offered traffic volume. Acknowledgements This work was partially funded by the European Commission through the Network of Excellence BONE. References 1 B.Puype et al., Optical Cost Metrics in MLTE for IPover-Optical Networks, Proc. ICTON ’04, 1, 75-80 2 K.Cho et al., Observing Slow Crustal Movement in Residential User Traffic, Proc. ACM CoNext ’08 3 M.Afanasyev et al., Analysis of a Mixed-Use Urban WiFi Network, Proc. IMC ’08, 85-98 4 J.Chabarek et al., Power Awareness in Network Design and Routing, IEEE Infocom ’08, 457-465 5 B. Nordman et al., Reducing the Energy Consumption of Networked Devices, IEEE 802.3 tut. (2005)
