Internet Engineering Task Force B. Zhang
Internet-Draft J. Shi
Intended status: Informational The University of Arizona
Expires: July 11, 2013 J. Dong
M. Zhang
Huawei
Januray 10, 2013
Power-aware Routing and Traffic Engineering: Requirements, Approaches,
and Issues
draft-zhang-greennet-01
Abstract
Energy consumption of network infrastructures is rising fast. There
are emerging needs for power-aware routing and traffic engineering,
which adjust routing paths to help reduce power consumption network-
wide. This document gives a high-level analysis on the basic
requirements, approaches, and potential issues in power-aware routing
and traffic engineering.
Status of this Memo
This Internet-Draft is submitted in full conformance with the
provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF). Note that other groups may also distribute
working documents as Internet-Drafts. The list of current Internet-
Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."
This Internet-Draft will expire on July 11, 2013.
Copyright Notice
Copyright (c) 2013 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(http://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
Zhang, et al. Expires July 11, 2013 [Page 1]
Internet-Draft Green Routing and Traffic Engineering January 2013
carefully, as they describe your rights and restrictions with respect
to this document. Code Components extracted from this document must
include Simplified BSD License text as described in Section 4.e of
the Trust Legal Provisions and are provided without warranty as
described in the Simplified BSD License.
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 3
2. Requirements . . . . . . . . . . . . . . . . . . . . . . . . . 4
3. Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . 4
4. Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4.1. Hardware Support . . . . . . . . . . . . . . . . . . . . . 6
4.2. Software Support . . . . . . . . . . . . . . . . . . . . . 8
4.3. Impacts on Protocols . . . . . . . . . . . . . . . . . . . 9
4.4. Network Monitoring . . . . . . . . . . . . . . . . . . . . 11
5. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
6. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 12
7. Security Considerations . . . . . . . . . . . . . . . . . . . 12
8. Informative References . . . . . . . . . . . . . . . . . . . . 12
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . . 13
Zhang, et al. Expires July 11, 2013 [Page 2]
Internet-Draft Green Routing and Traffic Engineering January 2013
1. Introduction
Driven by exponential growth of Internet traffic, networks worldwide
are expanding their infrastructures at a fast pace by deploying more
high-capacity, power-hungry routers, which also leads to increasing
energy consumption. Besides operational costs and environmental
impacts, the ever-increasing energy consumption has become a limiting
factor to long-term growth of network infrastructure, due to
challenges in power delivery to and heat removal from router
components as well as the hosting facilities [Gupta03] [Epps06].
Today's ISP networks have redundant routers and links, over-
provisioned link capacity, and load-balancing traffic engineering.
As a result, routers and links operate at full capacity all the time
with low average usage, typically less than 40% of link utilization.
This practice makes networks resilient to traffic spikes and
component failures, but also makes networks far from energy
efficient. Though advances in hardware design have made individual
routers more energy efficient over the years, there is still has a
long way to go before routers become energy-proportional. Recently
researchers have started to look beyond a single router or linecard
for network-wide solutions towards energy proportionality. Power-
aware routing and traffic engineering has been proposed to improve
network's energy efficiency, for example, aggregating traffic onto a
subset of links and putting the other links with no traffic into
sleep. As demonstrated in several research works, this approach has
the potential to save a significant amount of energy [GreenTE]
[Nedevschi08] [Chabarek08]. Designing a practical protocol, however,
has been challenging, because making routing protocols power-aware
brings fundamental changes to the routing system and the entire
network, thus it is a complicated task with involvement of hardware
support, protocol design and operations, and network monitoring.
This document gives a high-level analysis on power-aware routing and
traffic engineering, including solution requirements, existing
approaches, and potential issues. Power-aware routing and traffic
engineering exploits the over-provisioned feature of networks, so
there will be some impact on network performance and resilience. In
order to save energy without impacting network performance and
resilience too much, certain requirements should be met. When a
power-aware approach is implemented, new issues may arise, and should
be addressed to make the solution practical. While there are many
aspects of energy efficient networks, this document focuses on intra-
domain routing within a single ISP.
Zhang, et al. Expires July 11, 2013 [Page 3]
Internet-Draft Green Routing and Traffic Engineering January 2013
2. Requirements
The high-level idea of power-aware routing or traffic engineering is
to adjust routing paths based on traffic level. When traffic level
is high, use more links to carry the traffic; when traffic level is
low, merge traffic to a small number of links so that other links can
be put to sleep or reduce rate in order to save power. The
fundamental requirement for any energy-efficient network solution is
to save power without negative impacts on network operations, which
can be broken down as follows.
o The network should retain enough resiliency against node and/or
link failures.
o The network should have enough spare standby capacity or be able
to react quickly enough to traffic spikes in order to minimize
packet losses due to links/routers being in low power states.
o QoS metrics such as end-to-end delay should be kept at a desired
level.
o The operation of other protocols should not be interrupted, or at
least other protocols should be able to adapt without being broken
after links/routers change their power states.
While this document focuses on routing and traffic engineering, it
requires support from underlying hardware and system for energy
management capability, which is the topic of the IETF Energy
Management Working Group [EMAN-WG]
3. Approaches
In the last couple of years a number of power-aware protocols have
been proposed in research. Instead of listing them individually,
here we categorize the solutions along three different dimensions.
Link Sleep vs. Rate Adaptation
Sleeping and rate adaptation are two major ways to save energy in
computer systems. Many hardware, including line cards and chassises,
consumes a significant amount of power when they stand by without
doing any actual work. When put into sleep mode, they will consume
only a little power. Thus putting an idle component to sleep is a
common way to save energy. If there is a need to use this component,
it can be waken up and become usable after a transition time. The
longer a component is in sleep mode, the more power saved. A power-
aware protocol adjusts routing paths to increase the sleep time for
Zhang, et al. Expires July 11, 2013 [Page 4]
Internet-Draft Green Routing and Traffic Engineering January 2013
certain links in the network.
A network interface often supports multiple data rates. Operating at
a lower data rate usually consumes less energy, though the actual
rate-power curve varies from device to device. Rate-adaptation-based
approaches operate interfaces at lower data rates when the traffic
demand is low and increase the data rate when traffic demand is high.
Thus the routers can save power during low utilization period.
These two approaches are also related in the case of "bundled links"
[Fisher10]. A bundled link is a virtual link comprised of multiple
physical links. A sleep-based approach can put some physical links
into sleep to save power, which is same as conducting rate adaptation
on the virtual link with adjustment unit of a physical link.
Configured vs. Adaptive
The key in power-aware routing and traffic engineering is to adjust
routing paths in response to traffic changes, so that the power state
of routers (or router components) will also change accordingly to
achieve energy saving. Different approaches differ at the
granularity of the adjustment.
Some approaches take the long-term traffic average as input, and
output a routing configuration that is applied to the network
regardless of short-term traffic variation. This is mostly useful
when network traffic exhibits a stable, clear pattern, e.g., diurnal
pattern where traffic is high during work hours and low during off
hours. It can only exploit the target traffic pattern; it cannot
react dynamically to short-term traffic changes to either save energy
(by putting links to sleep) or avoid congestion (by waking links up),
but the design and implementation should be simple.
Another type of approaches is to adapt to traffic changes dynamically
on much smaller time granularity. This approach may be able to save
more energy and have better performance because it is more
responsive, but the design and implementation usually are more
complicated. This approach needs to continuously collect traffic
data in order to adjust routing dynamically. The adjustment may be
done periodically or whenever significant traffic changes are
observed.
Distributed vs. Centralized
In distributed solutions, routers make power-aware adjustment
decisions, such as link sleep/wake-up and rate increase/decrease,
locally without a central controller. These routers need to exchange
information in order to achieve consistent network states.
Zhang, et al. Expires July 11, 2013 [Page 5]
Internet-Draft Green Routing and Traffic Engineering January 2013
Distributed approach fits the Internet operation model well but its
design is the most challenging. Traditional routing does not respond
to traffic variation while power-aware routing does, and it needs to
do so without causing loops or congestions.
In centralized solutions, a controller computes the routing paths
considering the network topology and traffic demand, and informs
routers how to adjust their routing paths. A centralized server
usually has more complete information, more computation power, and
more memory and storage than routers, thus it may make better
decisions than distributed approach. The server locates in the
network NOC and can be backed up by server replicas. Nevertheless,
this approach requires high reliability of the server.
Both distributed and centralized solutions may find their places in
ISP networks. For example, centralized solution can be integrated
into the Path Computation Element (PCE) framework [PCE-WG]. There
can also be hybrid designs, e.g., using a centralized solution based
on long-term traffic pattern, and distributed mechanisms to handle
short-term traffic variations.
4. Issues
4.1. Hardware Support
In order to save power, routers and switches should support low power
states, and make available control primitives to enter or leave low
power states. To reduce the impact on network performance, routers
and switches should have the ability to change power states quickly.
These are the hardware support needed by power-aware protocols.
Sleeping State
Most sleep-based approaches require routers and switches, or a
component of them such as a line card, to support sleeping state.
While most components can go to sleep very quickly, they also need to
be able to wake up quickly. Besides, entering and leaving sleeping
state often incurs extra energy draw, which need to be kept small.
Different designs may have different requirements of the transition
time between power states. In uncoordinated sleeping approach,
upstream routers intentionally buffer packets for a very short period
of time to allow downstream routers longer sleep time. This approach
can only allow a component to sleep for a few milliseconds, otherwise
the buffering may cause too much extra delay. Hence this approach
requires a very short transition time and low penalty power. In
coordinated sleeping approaches, where routers coordinate on which
paths to use and when to put links to sleep, a component usually can
Zhang, et al. Expires July 11, 2013 [Page 6]
Internet-Draft Green Routing and Traffic Engineering January 2013
sleep much longer, for seconds, minutes or even longer. Therefore
their requirement on transition time and power is more relaxed.
Common energy management scheme at the individual component such as
line card is sleep-on-idle (SoI) and wake-on-arrival (WoA). When a
link is idle for a short period of time, it goes into sleep; when a
packet arrives, it wakes up. Power-aware protocols manipulate
traffic paths so that some links will have much longer idle time than
default routing.
The hardware is also expected to minimize potential packet loss
during the transition between power states. Especially in WoA, the
first packet is susceptible to loss. The two ends of the link can
coordinate, e.g., one end sends a dummy packet to the other end to
inform about the link wakeup, or if they don't coordinate, the
receiver end should have the capability to buffer incoming packets
before the interface wakes up to process these packets.
Multiple Data Rates
CMOS based silicon supports Dynamic Voltage Scaling (DVS), so
clocking an interface at a lower frequency, and operating at lower
data rate can save considerable amount of energy. This calls for a
need for router interfaces to support multiple data rates. If an
interface could support more data rates and incur low penalty power
on a change, there are more opportunity to save energy. Furthermore,
it will also help if an interface supports different sending and
receiving data rates.
The transition between different data rates needs be quick and on-
the-fly. Most Ethernet cards supports auto-negotiation of data
rates, which happens when a cable is plugged in and takes hundreds of
milliseconds. Auto-negotiation is not suitable for changing data
rate to save energy, because buffer would be filled up during the
negotiation period and leads to packet loss. A fast mechanism for
initiating and agreeing upon a link data rate change is necessary.
Electrical Damage
Many electronic devices are not designed to be turned on and off
frequently. When a device is waken up, the in-rush current may
damage or destroy a component and related circuits. Hardware that is
more friendly to power management is needed.
Optical Component Support
Electrical components consumes much more energy than optical
components in network routing infrastructure. Therefore, many power-
Zhang, et al. Expires July 11, 2013 [Page 7]
Internet-Draft Green Routing and Traffic Engineering January 2013
aware routing or traffic engineering approaches are designed with
electrical devices in mind. However, the number of optical
components in ISP networks can also be large. Care should be taken
when adopting existing approaches to optical networks. For example,
optical receivers cannot be turned off when WoA is needed.
Furthermore, there is room for more power saving when optical
components are explicitly considered in the approach. It's possible
to turn off an optical receiver while maintaining the ability to wake
it up when needed, by maintaining another route towards the other end
that a control packet could be delivered.
4.2. Software Support
There are many different power-aware approaches. They need different
input datasets, and generate different instructions. Software is
required to collect necessary input, as well as deliver and execute
resulting instructions.
Topology and Traffic
Many power-aware approaches require knowledge of global topology on a
centralized server or on each router. It's fairly easy to satisfy
this requirement by running a link state routing protocol such as
OSPF. If a network running OSPF has OSPF areas configured, power-
aware approaches can only be deployed within one area, or some other
way to collect global topology is needed.
Many approaches require knowledge of link utilization on a local
router, its neighbors, or all routers. Routers may need to maintain
necessary counters to calculate this information, and exchange or
announce them. Routers then need to categorize packets and maintain
a separate set of counters for each interesting category. Some
approaches such as GreenTE require network-wide traffic matrix.
There are two ways to obtain this information: infer from link
utilization, or collect directly. We can infer traffic matrix from
global topology and link utilization by using gravity model and
tomographic method [TM]. This method requires some computation
power, but needs least amount of data exchange, so it is particularly
useful when traffic matrix is only needed on a centralized server.
However, the accuracy of this method is not guaranteed, especially
when traffic engineering is in place that causes traffic pattern to
deviate from the gravity model, or multicast is enabled which creates
multiple copies of packets. We can also collect traffic matrix
directly. There is a cost on ingress routers: an ingress router
needs to identify the egress node, and maintain one counter per
egress. Identifying egress is not an extra cost in many cases,
because many approaches need to know egress to select a feasible
route. Maintaining per-egress counters, as well as sending them to
Zhang, et al. Expires July 11, 2013 [Page 8]
Internet-Draft Green Routing and Traffic Engineering January 2013
the centralized server in centralized approaches, is a high cost.
Traffic Splitting
Some approaches may route traffic between an ingress-egress pair
along multiple paths, according to certain split ratio. To avoid
out-of-order arrival which impacts TCP performance, traffic splitting
is usually based on the hash of some fields in the packet header,
such as source-destination IP pair. In a small network such as a
company, there are big flows between some IPs, while there is little
traffic on most other IPs. In this case, hash-based splitting has
significant bias.
Timeliness of Solutions
Traffic engineering approaches take network topology and traffic
information (in the form of link utilization or traffic matrix) as
input, and outputs a solution including which links should be
sleeping and what rate should links be operated on. Most traffic
engineering approaches run on a centralized server. Traffic demand
changes over time, and network topology may even change due to link
failure. It takes time to collect traffic information from the
entire network, and time is also consumed while computing the
solution. Thus, the solution, when comes out, is based on network
topology and traffic information of sometime earlier, and it may not
still be applicable to current situation. Prediction of future
traffic information may help in some situations.
4.3. Impacts on Protocols
Power-aware routing and traffic engineering is a tradeoff between
energy consumption and network resilience. They save power by
turning off or slowing down some links, which were previously over-
provisioned to obtain better resilience. Any power-aware approach
will cause loss of network resilience to some extent. Sleeping based
approaches has another impact. Traditionally, a link is either up or
down. An up link can transmit packets, and a down link cannot. A
third state, sleeping, is added by power-aware protocols. A sleeping
link cannot transmit packets right away, but it can be waken up when
needed. The introduction of a third sleeping state has its impact on
protocols that maintain their own states about network links.
Congestion after Traffic Surges
Traffic engineering approaches usually take traffic information at
certain time, and a solution contains a routing scheme that could
accommodate such traffic on a reduce topology with some links
sleeping or operating at lower rate. This routing scheme usually
Zhang, et al. Expires July 11, 2013 [Page 9]
Internet-Draft Green Routing and Traffic Engineering January 2013
keeps link utilization under certain threshold, so that there is some
safe margin in case traffic increases. However, because a solution
is computed periodically, congestion is still possible when traffic
increases to a level that exceeds the safe margin within one
adjustment period. To address this issue, some method of fast
readjustment is needed. When a traffic increase is observed, the
routing scheme should be slightly changed to accommodate this
traffic, probably waking up or increase rate on a few links.
Network Partition on Link or Node Failure
Many sleep based approaches will result in a topology with very low
redundancy level. These reduced topologies are vulnerable to link
and node failures, which are quite common in large networks. Those
approaches should be improved by adding a constraint of redundancy
level. A redundancy level of 2, which could protect from single link
failure, is a reasonable value. It's possible to incorporate power
aware feature into MRT to achieve energy saving while remain the
network 2-disjoint [I-D.ietf-rtgwg-mrt-frr-architecture]. Once a
partition is detected, it's easy to repair by waking up all sleeping
links. But this causes a sudden increase on power consumption, which
is sometimes undesirable. A local algorithm to select a subset of
sleeping links that could repair the partition is needed. The
selection doesn't need to be optimal, because waking up a small
subset is much better than waking up all sleeping links.
Sleeping Link State in Routing Protocols
Sleeping links should be handled separately in routing protocols. A
sleeping link should be advertised as up, probably with a tag stating
it's sleeping. No HELLO messages should be sent over a sleeping
link, so no HELLO messages could be received from a sleeping link.
Missed HELLO messages on a sleeping link should not cause the link to
be treated as down state. As a consequence, if a sleeping link
fails, the failure would not be detected until the router attempts to
wake it up. To detect a failure earlier, it may be desirable to wake
up the link and probe it periodically (using a long interval such as
every hour). No control message should be sent over a sleeping link.
This may cause the network to converge slower than usual, because LSA
flooding takes more hops. Fortunately most power-aware approaches
have network diameter constraints, so convergence time should be
comparable.
IP Multicast on Reduced Topology
IP Multicast works by building one or more trees on available links.
If any link in a multicast tree goes to sleep, some receivers cannot
receive multicast packets for a noticeable period of time, until IP
Zhang, et al. Expires July 11, 2013 [Page 10]
Internet-Draft Green Routing and Traffic Engineering January 2013
multicast automatically repairs the tree. So, if a link is part of a
multicast tree, it should not be put to sleep.
One solution is keeping all links that are contained in multicast
trees active. If there are many multicast trees that don't have much
overlap, a major portion of links would be forced active by
multicast, and power saving potential is greatly limited. Another
solution is explicitly modifying multicast trees in a power-aware
approach. This is not an easy way to go. There should be a delay
constraint on each multicast tree, and there're possibly a large
number of multicast trees. After a multicast tree is modified,
utilization of multiple links will change.
A third solution is making IP multicast power-aware. When a
multicast tree is being built, energy consumption is taken into
account, such that IP multicast would attempt to use as few links as
possible as long as delay constraint could be satisfied. After that,
these links used by IP multicast will not go to sleep.
4.4. Network Monitoring
Network operators demand a monitoring solution when deploying
anything. The most important metrics are: How much energy is saved?
How much impact is there on network performance? Measurement of
Energy Consumption. The IETF eman WG [EMAN-WG] is working on
defining energy objects in network devices, and monitoring and
controlling their states. When a device is running on full power
state, the power demand is recorded as full power demand. When a
power-aware approach is deployed, actual energy consumption is
measured. The amount of saved energy is the full power demand
multiplied by elapsed time during the measurement of actual energy
consumption subtracted by actual energy consumption.
In centralized periodical adjustment approaches, the centralized
server should have knowledge of current applied solution (which is
based on previous traffic information) and current traffic
information. It can then calculate what link utilization and delay
would be when this traffic is routed on current applied solution, as
well as the performance as if this traffic is routed without power-
aware consideration. It's not trivial to measure the impact in other
cases.
SNMP MIBs are needed to standardize monitoring. Software for
operations products such as System Center Operations Manager needs to
integrate power-aware routing and traffic engineering to existing IT
monitoring architecture.
Zhang, et al. Expires July 11, 2013 [Page 11]
Internet-Draft Green Routing and Traffic Engineering January 2013
5. Summary
Power-aware routing and traffic engineering has great potential to
improve network energy efficiency while maintain network services at
desired levels. Its effectiveness, however, depends on various
supports from hardware and software, and more importantly, protocol
designs that address operational issues. This document is a first
step towards developing practical power-aware protocols.
6. IANA Considerations
This document has no actions for IANA.
7. Security Considerations
This draft is a discussion on the Internet's necessity to follow an
evolutionary path towards the future. There is no direct impact on
the Internet security.
8. Informative References
[Chabarek08]
Chabarek, J. and et al. , "Power Awareness in Network
Design and Routing", IEEE INFOCOM 2008.
[EMAN-WG] "IETF Energy Management Working Group", 2012,
.
[Epps06] Epps, G. and et al. , "System Power Challenges", 2006,
.
[Fisher10]
Fisher, W. and et al. , "Greening Backbone Networks:
Reducing Energy Consumption by Shutting Off Cables in
Bundled Links", Green Networking 2010.
[GreenTE] Zhang, M. and et al. , "GreenTE: Power-Aware Traffic
Engineering", ICNP 2010.
[Gupta03] Gupta, M. and S. Singh, "Greening the Internet", ACM
SIGCOMM 2003.
[I-D.ietf-rtgwg-mrt-frr-architecture]
Atlas, A., Kebler, R., Envedi, G., Csaszar, A.,
Zhang, et al. Expires July 11, 2013 [Page 12]
Internet-Draft Green Routing and Traffic Engineering January 2013
Konstantynowicz, M., White, R., and M. Shand, "An
Architecture for IP/LDP Fast-Reroute Using Maximally
Redundant Trees", draft-ietf-rtgwg-mrt-frr-architecture-01
(work in progress), March 2012.
[Nedevschi08]
Nedevschi, S. and et al. , "Reducing Network Energy
Consumption via Sleeping and Rate- Adaptation", USENIX
NSDI 2008.
[PCE-WG] "IETF Path Computation Element Working Group", 2012,
.
[TM] Roughan, M., Thorup, M., and Y. Zhang, "Traffic
Engineering with Estimated Traffic Matrices", IMC 2003.
Authors' Addresses
Beichuan Zhang
The University of Arizona
Email: bzhang@cs.arizona.edu
Junxiao Shi
The University of Arizona
Email: shijunxiao@cs.arizona.edu
Jie Dong
Huawei
Email: jie.dong@huawei.com
Mingui Zhang
Huawei
Email: zhangmingui@huawei.com
Zhang, et al. Expires July 11, 2013 [Page 13]