Power-efficient Electronic Burst Switching for Large File
Transactions
Ilijc Albanese
1
, Sudhakar Ganti
2
and Thomas E. Darcie
1
1
Department of Electrical and Computer Engineering, University of Victoria, Victoria, BC, Canada
2
Department of Computer Science, University of Victoria, Victoria, BC, Canada
Keywords: IP Networks, Burst Switching, High Speed DWDM Optical Transmission, Optical Network, Overlay
Network, Low Power Networking, Electronic Buffering, Router Power Consumption.
Abstract: Much of the growth in bandwidth demand and power consumption in today’s Internet is driven by the
transport of large media files. This work presents a power-efficient overlay network specifically designed
using electronic burst switching for these large files. The two approaches are presented in which electronic
bursts or media frames (MF) containing >1Mb are routed in a manner similar to UDP or concatenated into
periodic semi-transparent chains and routed using a two-way reservation protocol. Utilization, blocking,
delay and buffer size are compared to UDP/IP by means of simulation. Both approaches dramatically reduce
header-related power consumption. Concatenation also reduces significantly the amount of buffer space
required. A representative router design is evaluated showing a potential energy saving of roughly 80%
relative to standard IP routers.
1 INTRODUCTION
Despite ongoing improvements in bandwidth
capacity and power efficiency, power consumption
in router networks continues to be a concern.
Estimates suggest that Internet-based
communication technologies consume between 2-
3% of power generated globally and that this
number is increasing at a rate of 16-20% per year
(Fettweis and Zimmermann, 2008). As a result,
extensive effort is ongoing to develop new
techniques for minimizing power consumption in
future Internet technologies.
Of particular interest is power consumed in
electronic routers. Numerous studies have shown
that a significant portion of power consumed can be
attributed to header processing on each packet. With
the increasing popularity of bandwidth intensive
applications such as streaming video and the sharing
of large files, studies support the general observation
that file sizes are growing rapidly. Given the large
number of IP packets required for these transactions,
header related power consumption is a key
contributor to the rapid growth of power
consumption in routers. Energy efficiency improves
if larger packets are used for large file transfers. In
response, the maximum IP packet size was extended
from 1500 bytes (B) using Jumbo and Super Jumbo
frames pushing packet sizes to 64KB. However,
these specifications are not in widespread use due to
problems related to backwards compatibility
(psc.edu, 2012) and network latency arising from
integration of very large packets with smaller
packets within the same links. Integration of large
(up to 19.5KB) and standard packets within the same
network was considered (Divakaran and Altman,
2009) showing that the use of larger packets may
reduce power consumption and computational load
required from the network hardware. However,
coexistence of large and small packets in the same
links results in unfairness and higher drop rates for
both types of packets, leading to inefficient
bandwidth and computing power utilization.
This raises the question as to the potential value
of using a separate overlay network for the traffic
associated with large file transfers. While clearly the
introduction of a new overlay network would have
to be predicated on compelling value, it would be
unwise to suggest that, given power consumption
and scaling considerations of IP, a next generation
non-IP switching/routing approach cannot exist.
What might be the form of this new overlay
network? On one extreme, numerous optical
43
Albanese I., Ganti S. and E. Darcie T..
Power-efficient Electronic Burst Switching for Large File Transactions.
DOI: 10.5220/0004357400430052
In Proceedings of the 2nd International Conference on Smart Grids and Green IT Systems (SMARTGREENS-2013), pages 43-52
ISBN: 978-989-8565-55-6
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
networking approaches offer up to an entire
wavelength for some time, through which GigaByte
(GB) files can be delivered. Optical burst (OBS)
(Jue and Vokkrane, 2005) or flow switching (OFS)
(Chan, 2010), or user-controlled end-to-end
lightpaths (e.g. CAnet4, MONET, CORONET and
GRIPhoN (Mahimkar et al., 2011)) have been
explored fully. For example, an OBS approach
(Yong et al., 2010) uses concatenated data bursts
where the data units are organized as non-contiguous
and non-periodic series of concatenated timeslots
(bursts), which are then handled as a whole in an all-
optical network infrastructure. These optical
approaches are not embraced by industry, in part
because the power efficiency of optical switching is
questionable (Tucker et al., 2009), (Tucker, 2006)
and optical buffers, widely used in most OBS
proposals, have not yet offered a commercially
viable alternative to electronic buffers. Also, while
capable of supporting large bandwidth, targeted
implementations are in the interconnection of
specialized nodes rather than broadly distributed
Internet users.
Hybrid architectures have been studied in which
both electronic and optical switching are combined
(Aleksic et al., 2011) to simultaneously handle
packets, bursts and TDM circuits. A large reduction
in the power consumption is achieved by selecting
adaptively which part of the node to activate based
on a per-flow evaluation of the data to be routed
while the other blocks are put in sleep or low-power
mode. While potentially powerful, this approach
requires the complex integration of disparate
switching and control elements, some of which (like
OBS and optical delay lines) have not proven
compelling individually.
A more incremental overlay network approach is
electronic burst switching (EBS) (Peng et al., 2010).
Following the OBS model, bursts are assembled at
edge burst switches and switched electronically at
core switches. It was concluded that using large
bursts (> 1 Mb) may lead to reduction in header-
related power consumption in core switches, but the
power consumed by burst assembly negates much of
the advantage gained in core switching.
In this paper we continue along the path of EBS.
Users share the bandwidth of an overlay network,
which we presume to be statically provisioned, using
electronic switches or routers specifically designed
to handle large file transactions. Unlike (Peng et al.,
2010), we eliminate burst assembly at edge switches
and consider direct end-to-end delivery of large
“media” frames (MF) (roughly 1-10Mb) to users
through an overlay to next-generation optical access
networks. Free from the constraints of coexisting
with highly granular and dynamic IP traffic, this
EBS overlay network can be designed specifically
for the efficient delivery of the large data
transactions that did not exist when the Internet was
conceived. Compared to traditional IP routers,
switch reconfiguration can be far less dynamic since
only very large packets are supported. Unlike
proposed optical alternatives, this can be
accomplished using available electronic buffers in a
form that is entirely compatible with today’s highly
efficient cross-point switch arrays.
Our objective is to enable a significant reduction
in power consumption of network hardware while
optimizing the use of resources. We first explore
routing MFs using a standard UDP protocol (MF-
UDP). UDP is selected for this study, rather than
TCP, as this avoids numerous complexities that add
little insight to a comparison with conventional IP
and, as discussed later, gives the best case scenario
for IP. Based on simple hardware considerations,
network performance simulations and comparison
with traditional UDP, we arrive at the anticipated
conclusion that router power consumption is reduced
dramatically, but performance is otherwise
unaffected and larger buffers are needed.
We then consider using concatenations of MFs
into periodic semi-transparent chains (MFCs) and
the scheduled transmission of these MFCs using a
two-way reservation mechanism. While such a
scheduling mechanism would be inappropriate for
traditional IP traffic, the large size of each MFC
(e.g. 1 GB) makes scheduling both manageable and
worthwhile. Also, the structure of an MFC makes it
easy to condense information on its configuration
with minimal control plane information, minimizing
the amount of information to be processed at each
node to schedule the chain and reducing the
probability of control plane collisions. Simulation
results show increased utilization efficiency and
decreasing buffer requirements in comparison to
MF-UDP as well as standard UDP. An MFC router
is designed based on a commercially available cross-
point switch array and power consumption is
estimated to be roughly 20% of that of a standard IP
router.
The paper is organized as follows: Section 2
provides an overview of the reference network
architecture in the context of transactions of large
files. Section 3 compares, using OMNeT++
simulation, traditional UDP to MF-UDP in
supporting representative large transactions. In
Section 4, chains of media frames (MFCs) are
introduced, along with exemplary admission control
SMARTGREENS2013-2ndInternationalConferenceonSmartGridsandGreenITSystems
44
and scheduling algorithms, and compared to UDP
and MF-UDP. Router implementation is discussed in
Section 5, where implications on power
consumption are considered and a representative
MFC router is evaluated.
2 NETWORK TOPOLOGY
AND LARGE FILE FLOWS
Our discussion is framed by the reference network
architectures shown in Figures 1 and 2. Fig. 1 shows
a hierarchical network representative of the current
state-of-the-art comprising various sizes of routers
(access, edge, and core) connected to a transport
network through various sizes of add-drop
multiplexers.
Figure 1: Reference Architecture.
Primary examples of large-file transactions can be
super- imposed onto this reference architecture.
These examples include: 1) Regionally-cached
download: In this case large media files are
downloaded from regional cached distribution
servers at the end points of Content Delivery
Networks (CDN), through Metro and access to end-
users. Driven by rapid proliferation of video-related
download applications, these downloads represent a
significant fraction of traffic growth, and therefore
are the focus of this study. Other important
transactions include: 2) Source-to-cache distribution:
To deliver and update content to CDN servers, large
files must be distributed from sources, typically
across a core network, to the regional cache; 3) End-
to-end file transfers: For peer-to-peer applications,
or for files for which widespread distribution is not
anticipated, the regional cache is bypassed and files
are transacted through Metro and access, possibly
across the core network, directly to an end-user.
An overlay network designed specifically for
large file transactions might look like Fig. 2. Each of
the “media” boxes parallels a present-day IP
counterpart and supports the origination or
termination (media client interface (MCI)), access
bypass (media access bypass (MAB)), admission
(media access router (MAR)), and efficient end-to-
end routing (media backbone router (MBR)), in
accordance with the principles described below for
MFs. In addition, it is convenient to consider
regional storage of MFs in a media frame cache
(MCa).
Figure 2: Media Frame Overlay Architecture.
Our intention is not to restrict application to
specialized nodes (e.g., campuses), but rather to
reach broadly distributed consumers, methods must
be established to transport MFs through access.
Two possible solutions are represented in Fig. 2.
A conservative approach involves an application
operable between the MCI and the MAR such that
using traditional broadband access alone, MFs and
MFCs are assembled or disassembled at the MAR.
This functionally is parallel to the burst assembly
routers proposed for use in OBS and EBS. A
preferred approach involves engaging the evolution
of optical access standards, where standards for 10
Gb/s PON have emerged recently, to enable higher
dedicated bandwidths perhaps through wavelength
overlays. MFs and MFCs would then be assembled
at the user end point or client directly.
3 MEDIA FRAMES VERSUS
CONVENTIONAL IP
We first explore the issues associated with migrating
very large file transfers to a separate overlay
network in which the standard unit of bandwidth is a
media frame (MF) containing roughly 1-10 Mb of
data plus overhead. In (Peng et al., 2010) it was
concluded that although using large packets would
Power-efficientElectronicBurstSwitchingforLargeFileTransactions
45
lead to considerable power savings, increasing the
frame size beyond 1Mb would only marginally
increase the energy efficiency of an EBS node.
However, using larger frames also reduces the
required reconfiguration speed of the switch fabric,
minimizing requirements on switching speed and
inefficiency introduced during transitions.
Overhead may include address, priority,
concatenation details (specifying MFCs, as
discussed later), coding, guard time and
management information. Given the very large
capacity within each MF, considerable header
information (e.g. 10 KB) can be included with
minimal impact on throughput. Structure may be
defined to facilitate error correction, segmentation,
security, file compression, and easy assembly from a
large numbers of smaller IP packets.
An obvious method for networking with MFs is
to adopt the same concepts as used with TCP/UDP,
allowing each MF to be routed in accordance with
predefined routing tables through a connectionless
queuing network. While the dynamics are
considerably different than with < 1500 B UDP
packets, the underlying issues are the same.
To explore this in detail, both UDP and MF-UDP
network simulators were built using OMNeT++
(OMNeT++, 2012) and the performances compared
in terms of link utilization, buffer space occupied
and delay per GB of data transferred. UDP was
simulated using 1500B packets and MF-UDP using
1Mb packets. Droptail queues were used for both
1Mb and 1500B UDP packets. For standard UDP
packets the maximum buffer size was set to 1000
packets, corresponding to roughly 1.5MB. Buffers of
the same size were used for MF-UDP.
A dumbbell topology was considered for the
simulations (Figure 3). While more complex
topologies could be simulated, this represents the
case of our reference network (Figure 2) with a
congested link between multiple source servers and
end users.
The capacity of each link is set to 10Gib/s (i.e.
10*2^30 Gb/s according to IEC standard) and each
source is offering the network an average load of
roughly 1.33 Gb/s. Various load conditions were
tested by activating more source-destination pairs
and the offered load was made to vary from 25% to
150% of the bottleneck link capacity (corresponding
to from 2 to 12 source-destination pairs). Each
source transmits data to one destination only and all
sources compete for the same bottleneck link.
Compared to UDP, the router switch fabric
becomes far less dynamic. Reconfiguration occurs
far less frequently (by 3 orders of magnitude).
Figure 3: Topology Used for Simulations.
It remains to be seen if an MF-router can be
designed to exploit this less dynamic reconfiguration
and negligible header processing with a sufficiently
large net increase in power efficiency to justify a
separate overlay network. This is addressed further
in Section 5.
4 CONCATENATED MEDIA
FRAMES
While UDP-like MF routing dramatically eases
processing (header and switch reconfiguration) over
traditional IP, performance is otherwise essentially
the same. We now consider the potential impact of
concatenation of MFs into larger structured chains
(MFCs). Concatenation creates single entities that
would contain an entire large transaction, for
example a multi-GB movie download. Discussion
centers on two key considerations: scheduling and
transparency. Scheduling and admission control
become worthwhile for such large transactions and
these can be used to increase resource utilization and
minimize buffer size. However, serving such large
transactions continuously in time introduces
significant latency for waiting transactions and is
incompatible with the lower end-user client and
access network throughputs. Making each MFC
partially transparent overcomes both problems. We
limit our discussion here to a simple functional
description of a representative methodology,
including MFC structure, transparency, signaling
and scheduling algorithm, then compare
performance to conventional UDP and the MF-UDP
described in Section 3.
4.1 Media Chain Transparency
A MFC with transparency degree 3 (defined as the
number of interstices between two consecutive MFs
in a chain plus 1) is illustrated in Figure 4.
Using periodic semi-transparent chained data
structures provides five primary functions. First,
given the large size of each MF, concatenations of
multiple MFs without transparency would introduce
substantial latency by occupying network resources
SMARTGREENS2013-2ndInternationalConferenceonSmartGridsandGreenITSystems
46
Figure 4: An Example of How Two MFCs, MF-UDP
Frames and Voids Fit in a Channel.
for substantial durations. The use of transparency
allows servicing multiple MFCs without introducing
long delays, albeit at a slower data rate. Secondly, a
fixed transparency simplifies scheduling of large
amount of data with minimal computational load.
Third, the semi-transparent structure of each MFC
allows the use of buffering as a means to affect the
relative timing of an MFC with respect another in
order to interleave MFCs competing for the same
link, but without buffering entire chains. Our hope is
that this use of the buffer will enable significant
savings in buffer space occupied relative to the UDP
and MF-UDP of Section 3. Fourth, we anticipate bit
rates of 10Gb/s in access networks at a time when
bandwidths in core networks will be 100Gb/s.
Transmission from access must then be up-shifted in
rate, resulting in, for this case, 10% time occupancy.
This would be accommodated naturally by a core
transparency degree 10 times higher than that used
in the access. Finally, for file transfer applications
one desires to send as much information as fast as
possible. For this transparency is a disadvantage.
However, streaming applications have evolved to
deliver segments of large files spread out over long
time periods. A large transparency degree can be
used to emulate streaming delivery while retaining
the other worthwhile attributes described above.
4.2 Signalling and Control
Signaling is needed to establish and update the
network state, schedule MFC transmission,
acknowledge receipt, and many other functions.
Information about an entire MFC, including length,
scheduling information, priority, etc. can be easily
contained within each MFC, MF in an MFC, or a
small separate control packet. Signaling could be
‘in-band’ using periodic time slots within the MF
transport structure, or ‘out-of-band’. Our preference
is to exploit the ubiquitous availability of traditional
IP networks for out-of-band signaling. In what
follows, we assume that each of our media access
and backbone routers (MAR, MBR in Fig. 2) are
able to signal through a suitable IP network.
Global Control: Since each signaling event
corresponds to of order 1 GB of data, the number of
signaling events is small. It is then reasonable to use
a centralized ‘state server’ to provide each router
with global path, timing, and occupancy
information. Each router updates status to the state
server regularly, and the state server is able to
calculate paths and approve initiation of a request
for MFC scheduling, as discussed below. The state
server must know the topology and is then able to
make globally informed decisions to queries from
routers. It is also useful to know the propagation
time between nodes for efficient scheduling.
Numerous methods can be implemented, like the
ranging protocol used in PON, to estimate these
times and report them to the scheduling server.
Distributed Control: Each router communicates
directly with its neighbors and the state server. Each
router continually updates the state server of status
and load, and MARs request path and approval for
MFCs from the state server. Approval does not
guarantee success, but suggests high probability.
Communication between routers along the path
determines ultimate success, as described below.
This minimizes latency in each MFC request-grant
negotiation.
4.3 Scheduling
The objectives of scheduling are to organize
transmission of concatenated chains in such a way as
to minimize hardware complexity and power
consumption, to maximize link utilization and to
minimize buffer requirements. All of these
objectives can be addressed through the use of an
Expected Arrival Time (EAT
i
) of the MFC to the
next node in its path. This is computed based on the
physical distance between the nodes, which is
assumed to be known by each scheduling node, and
included in a control packet CP (< 1 Kb). The CP is
used to reserve resources for its associated MFC
along its path. Many variations of the scheduling
algorithm can be considered. For purposes of
simulation, we defined an example that comprises
the following basic steps:
- MFCs assembled at end user machine.
- A request packet containing at least the length of
MFC, transparency degree, source and destination
address is sent to media access router (MAR).
- MAR queries state server for path, propagation
time associated with each hop in path.
- MAR estimates a Time-to-Transmit (TT)
parameter. All routers along path use TT to search
Power-efficientElectronicBurstSwitchingforLargeFileTransactions
47
for available time slots. TT is determined based on
transmission and propagation delay from user to
the MAR, hop distances along path and processing
time for control packets at each node.
- MAR generates control packet (CP) and sends it to
next node along path. CP contains
sender/destination address, length of MFC,
transparency degree, expected arrival time
(EAT[Ni]), and ID that associates each CP to an
MFC. EAT[Ni] indicates to the node receiving the
CP the amount of time, after the reception of the
control packet, before the arrival of first bit of the
MFC .
- EAT field in CP is updated before forwarding CP
to next hop node, in order to account for the
additional transmission, buffering and processing
delays at each node. This continues until
destination (egress MAR) is reached or until the
reservation process fails.
- If reservation succeeds, confirmation packet is sent
over IP network directly to source node which
starts transmission of MFC. If reservation fails, a
“NACK” packet is sent over reverse path to free
resources and source will retry after random back-
off time.
Since the estimated EAT for the MFC is
computed using the physical hop delay (known
globally) and the expected arrival time is carried in
the CP, there is no need for network-wide
synchronization. A local timer at each node keeps
track of the time differences between the reception
of the CP and the expected arrival time of its relative
MFC. The guard time in each MF compensates for
the time uncertainties in this estimation process.
4.4 Simulation
The scheduling algorithm was implemented (details
to be published) using OMNeT++ and the same
topology used in Section 3. Each MF is assumed to
contain 1Mb of data and each MFC is defined as the
concatenation of 8000MF (i.e. 1GB of data). A fixed
transparency degree of 8 was chosen for all MFC
simulations. The performances of the MFC-based
transport are then compared to those of UDP sources
using MF (1Mb) and standard UDP packets
(1500B).
Given the functioning of the algorithm presented
in Section 4.3, the payload bits arrive at the node
only if the reservation process was successful. In
order to compare this to a connectionless protocol as
in Section 3 (UDP) it was assumed that each source-
destination pair would attempt to transmit on
average the same amount of data. Hence for the
UDP cases (both standard and MF-UDP) the
“offered load” is the load physically reaching the
bottleneck node. For the MFC case, the offered load
is computed based on the number of reservation
attempts per second. The maximum buffer size
allowed for the bottleneck router (router 1 of Figure
3) was set to be 1.5MB for all cases.
4.5 Simulation Results
Link Utilization: As can be seen in Figure 5 the
link utilization is very close for all 3 approaches
tested for load conditions up to ~74%. Beyond this
point the higher cost of dropping larger frames
comes into play and the bandwidth efficiency of
MF-UDP is reduced. For load greater than ~90%
load packet dropping also affects standard UDP and
its utilization drops below that of MFC. MFC
utilization is ~9% higher for higher loads.
Figure 5: Link Utilization Vs Offered Load.
The highly structured MFC and the scheduling
algorithm allow interleaving large amounts of data
with link utilization similar to that of time division
multiplexed systems and performance consistently
better than both other cases under high load
conditions.
Packet Dropping and Blocking: Packet
dropping (two UDP cases) and blocking (MFC) are
presented Figure 6. The random nature of the arrival
for UDP packets allows for the possibility of filling
the queue at the bottleneck node even if maximum
load has not yet been reached. In the reservation-
oriented MFC system call blocking only occur after
the maximum bottleneck link capacity has been
effectively reached.
It is important to note that when a packet is
dropped, payload bits are discarded and these may
have already used resources along their path (buffer
space, switching power, bandwidth, etc.). When an
MFC request is blocked, we are simply rejecting an
attempt to transfer the data and not the data itself.
SMARTGREENS2013-2ndInternationalConferenceonSmartGridsandGreenITSystems
48
Figure 6: Packet Dropping and Call Blocking Vs Offered
Load.
This advantage of reservation-based mechanisms in
terms of efficiency of resource usage is well known.
Delay: To compare the delay performance a
1GB transaction was taken as a reference quantity.
For MFCs, upon failing a resource reservation
attempt, the source simply backs off for a random
amount of time before re-attempting the reservation
procedure. The back-off time is exponentially
distributed with an average equal to the duration of
an entire chain. Re-transmission attempts were also
taken into account in the delay performance
measurement, as shown in Figure 7. Up until 75%
load, the delay experienced by the MFC system is
virtually identical to that of both UDP cases. Beyond
this point using MFCs reduces delay. Besides
offering equal delay performances to UDP for the
majority of the range of operation of the network,
MFC also provides reliable data transfer, which
cannot be achieved by UDP due to the statistical
nature of arriving packets.
Figure 7: Delay (per GB of data transferred) Vs Offered
Load.
Buffer Size: As shown in Figure 8, the buffer size
required at low load for MFC and MF-UDP is higher
than that occupied by standard UDP (with MF-UDP
occupying the largest buffer space). As the load
increases (>88%) MFC requires considerably less
buffer space than both MF-UDP and standard UDP,
which beyond a certain load quickly start to fill
buffers up to the maximum capacity.
Figure 8: Buffer Occupancy Vs Offered Load.
The periodic data structure of the MFC and the
scheduling algorithm bound the required buffer
space to about 2-5 times less than that of both UDP
approaches.
Figure 9: Buffer Occupancy Vs Offered Load for Various
Frame Sizes for MFC.
For both UDP and MF-UDP the buffers are required
to avoid dropped packets, while for MFC the buffers
are used to align MFCs in time with scheduled slots.
This results in a buffer size that depends much more
on the size of the MF within the MFC than on the
offered load. This dependency is shown in Figure 9
where various media frame sizes were tested for
MFC. Varying frame size results in virtually
identical performances in terms of blocking
probability and utilization but a significant increase
of the required buffer space. Similarly, reducing the
MF size for each MFC can reduce the buffer space.
Increasing the frame size for the MF-UDP case
would result in buffer sizes that would simply
become impractical. In the MFC case using frames
larger than 1Mb may enable a relaxation of the
reconfiguration speed for the switch fabric as well as
a reduction of the CPU utilization (see section 5.1)
while keeping the buffer size limited.
Power-efficientElectronicBurstSwitchingforLargeFileTransactions
49
5 DISCUSSION
AND ANTICIPATED BENEFITS
Results for MFC indicate a considerable advantage
in terms of buffer size as well as a large reduction in
processing power with respect to UDP. A
comparison with TCP would have been useful in
that, unlike UDP and more like MFC, TCP can
guarantee the correct delivery of the file transferred.
TCP acknowledgements and retransmissions would
have made the delay per GB much higher and
bandwidth utilization much lower than that achieved
with UDP. In other words, for our simulation study
UDP is the best-case scenario for IP networking.
5.1 Impact on Router Power
Consumption
In (Aleksic, 2009) the issue of power consumption
in large scale networks is addressed concluding that
power consumption of header related functions is far
larger than that consumed by the switching fabric.
Apparently, most power is consumed in data
processing functions (i.e. header parsing, address
lookup, etc.), which must be carried out for each
packet traversing a router.
MF-UDP and MFC offer significant reductions
in the number of packets to be processed. Consider a
state-of-the-art router working at 1 Tb/s. Depending
on the manufacturer such a router can consume
about 4 KW (CRS-3, 2012), (T1600, 2012), or an
overall energy per bit of roughly 4 nJ/b.
Estimates of the power consumption of the
various functional blocks of a similar IP router can
be found in (Tucker et al. 2009). From this study it is
clear that, other than power supply and cooling
blocks, which are largely dependent on the energy
needed by the other blocks, the forwarding engine
consumes the most power, using about 32% of the
energy supplied to the router.
Assuming an average packet size of 10 Kb
(caida.org, 2012) means that a 1 Tb/s router will
have to perform header related operations
approximately 10^8 times per second (CRS-3,
2012). Organizing data in MFCs carrying roughly 1
GB of payload can reduce this by many orders of
magnitude.
In addition, the required processing speed is
reduced so that a much slower processing unit can
handle the same data throughput. From the study
presented in (Wang et al.,2006) it is also reasonable
to say that the use of large frames, together with
organizing the data within the sources in large
blocks (i.e. MFs) will allow a significant reduction
in CPU utilization in terms of number of memory
accesses and IRQs that the CPU has to handle. This
may lead to additional power savings in data storage
servers.
Given the large size of the MFs, requirements on
the reconfiguration speed are significantly relaxed,
as well as the amount of control information needed
to drive the switch fabric, with further impact on the
power consumption. Further reduction in the amount
of control information is achieved when using
MFCs, as a result of the predictability of the periodic
payload.
5.2 Power Consumption of Example
MFC Router
A schematic of an MFC-based router based on a
commercially available cross-point switch array is
shown in Figure 10.
Figure 10: Concatenated Media Frame Router.
Input data is converted into the electrical domain
and data streams from each channel are de-
multiplexed into their constituent MFCs. Each MFC
is then delayed by the amount indicated in its
associated CP. The number of buffer queues needed
depends on the number of chains the device is able
to handle and is given by # of dedicated buffer
queues = # of input channels * # of MFCs per
channel. The total buffer size also depends on the
number of simultaneous flows the device must
handle and the transparency degrees of the chains.
At the output of the buffer stage, chains
competing for the same output channels are
synchronized in order to allow interleaving. MFCs
are then passed to the switch fabric, which simply
routes each MFC to the appropriate output with no
further buffering or processing.
In order to estimate the power consumed by our
proposed router shown in Figure 10, the VSC3144-
11 has a switching capacity of 1.2 Tb/s with a power
consumption of about 20 W (Vitesse, 2012).
SMARTGREENS2013-2ndInternationalConferenceonSmartGridsandGreenITSystems
50
Regarding the buffer stage, values for the power
consumption of electronic memories are largely
dependent on implementation and size. If we assume
our MFC router would use a memory with similar
size and structure to that used in (CRS-3, 2012),
using the data on power consumption for a router
from (Tucker et al., 2009) gives roughly 200 W for
the buffer stage power consumption. Similarly we
can estimate the power consumption of O/E/O
blocks (including Tx/Rx equipment) to be about
280W.
The forwarding and routing engine, using about
32% and 11% of the total energy consumed by a
standard IP router, respectively Tucker et al., 2011)
(i.e. a total of ~720W (CRS-3, 2012) ), are grouped
in the “control” block of Figure 10 and, as a result of
the drastic reduction in the number packet processed
will be most likely reduced to negligible levels.
The input de-multiplexing stage, which de-
multiplexes MFCs from each channel, can be
modeled by (Tucker, 2011):
E
DeMUX
E
0
Lo
g
2
k
(1)
Where
0
E
is the energy per bit for a 1:2 de-
multiplexer and k is the number of output ports.
Assuming 2010 technology,
0
E
= 10 pJ (Vitesse,
2012). We can set k = 144 in Equation (1) obtaining
a total energy for the input stage of about 71W or 71
pJ/b.
Aside from power supply and cooling equipment,
the power consumed by the MFR would be roughly
571W. Power supply and cooling would consume
about 33% of the power consumed by the rest of the
device (Peng et al., 2010), leading to total power
consumption for the MFR of roughly 759W. This is
roughly 19% that of a state-of-the-art IP router
working in the same throughput range and load
conditions (CRS-3, 1012). This is dominated by the
power consumed by the buffer stage. Recognizing
that these buffers function more like slowly
reconfigurable delay lines than the buffers used in
typical IP routers, further study may reveal even
more significant reductions in power consumed.
6 CONCLUSIONS
In this paper we study the potential advantages of
using an overlay network in which only large media
frames (MFs) (1-10Mb) or concatenated frames
(MFCs) are used to efficiently transfer large files.
Numerical simulation is used to compare the use of
MFs using a traditional UDP routing protocol (MF-
UDP) to traditional UDP. Little difference is
observed in delay, utilization and throughput, while
the large packets dramatically reduce header-related
processing load.
In an effort to reduce buffer size and improve
resource utilization, a reservation-based networking
approach is developed using MFCs and compared to
MF-UDP and UDP through simulation. A
reservation system is defined for scheduling MFC
transmission, eliminating wasted network resources
since data leaves the source only if service is
guaranteed. Results show that buffer size can be
reduced by at least a factor of 2 under high load
conditions and the scheduling of large transactions
can increase efficiency to close to 100%. Further
advantages of using periodic, semi-transparent
MFCs is the ability to schedule large amount of data
with minimal header processing and to reduce the
reconfiguration speed requirements of the switch
fabric, ultimately reducing power consumption.
A representative MFC router is designed and
power consumption is estimated, under conservative
assumptions, to be roughly 20% that of a traditional
IP router.
Other ongoing studies include methods to allow
coexistence of scheduled MFCs and directly routed
MF-UDP within the same routers and links, and
extension of the simulations to other network
topologies. Our objective is to more definitively
articulate the cost-benefit trade-off, where the cost is
the rather large barrier associated with the creation
of a new overlay network.
In summary, the use of very large packets (MFs)
and concatenations of these (MFCs) offers an
interesting path to more power-efficient networking
for the dominant and rapidly growing portion of
Internet traffic that comprises very large (i.e. 1 GB)
transactions.
REFERENCES
G. Fettweis, E. Zimmermann. (2008) “ICT Energy
Consumption –Trends and Challenges”, 11
th
International Symposium on Wireless Personal
Multimedia Communications (WPMC 2008)
Psc.edu (2012). Retrieved October 25, 2012 from
http://staff.psc.edu/mathis/MTU/AlteonExtendedFram
es_W0601.pdf
Divakaran, D. M.; Altman, E. (2009) Post, G.; Noirie, L.;
Primet, “From Packets to XLFrames: Sand and Rocks
for Transfer of Mice and Elephants”, in IEEE
INFOCOM Workshops 2009.
J. P. Jue, V. M. Vokkarane, (2005)Optical Burst
Power-efficientElectronicBurstSwitchingforLargeFileTransactions
51
Switched Networks”, Ed. Springer 2005.
Chan, V. W. S. (2010) "Optical flow switching", Optical
Fiber Communication (OFC), collocated National
Fiber Optic Engineers Conference, 2010 Conference
on (OFC/NFOEC), vol., no., pp.1-3, 21-25 March 2010.
Ajay Mahimkar, Angela Chiu, Robert Doverspike, Mark
D. Feuer, Peter Magill, Emmanuil Mavrogiorgis, Jorge
Pastor, Sheryl L. Woodward, and Jennifer Yates.
(2011). Bandwidth on demand for inter-data center
communication. InProceedings of the 10th ACM
Workshop on Hot Topics in Networks (HotNets-X).
ACM, New York, NY, USA, , Article 24 , 6 pages
Yong Liu, Kee Chaing Chua, Gurusamy Mohan. (2010)
Achieving High Performance Burst Transmission for
Bursty Traffic using Optical Burst Chain Switching in
WDM Networks”, IEEE Transactions on
Communications, Vol.58, Issue 7 pp. 2127-2136, 2010.
Tucker, R. S., Parthiban, R., Baliga, J., Hinton, K., Ayre,
R. W. A., Sorin, W. V. (2009) Evolution of WDM
Optical IP Networks: A Cost and Energy Perspective,
IEEE J. Lightwave Technology, Vol. 27, Iss. 3, Feb.
2009, pp. 243-252.
R. S. Tucker. (2006) “The role of optics and electronics in
high-capacity routers”, J. Lightwave Technol., vol.
24, pp. 4655 - 4673, 2006.
Slavisa Aleksic, Matteo Fiorani, and Maurizio Casoni
(2011) “Energy Efficiency of Hybrid Optical
Switching”, ICTON 2011
S. Peng, K. Hinton, J. Baliga, R. S. Tucker et.al.
(2010)”Burst Switching for Energy Efficiency in
Optical Networks”, OSA/OFC/NFOEC 2010
OMNeT++ (2012) Discrete Event Simulation Tool,
http://www.omnetpp.org.
Aleksic, S. (2009) "Analysis of Power Consumption in
Future High-Capacity Network Nodes," Optical
Communications and Networking, IEEE/OSA Journal
of, vol.1, no.3, pp.245-258, August 2009.
CRS-3, single shelf system cisco data sheet (2012)
(http://www.cisco.com/en/US/prod/collateral/routers/p
s5763/CRS-3_4-Slot_DS.html)
T1600 Juniper networks data sheet (2012) (http://
www.juniper.net/techpubs/en_US/release-independent/
junos/topics/reference/specifications/t1600-specifica
tions-power-requirements.html )
caida.org (2012). Data retrieved October 25, 2012, from
www.caida.org
Wilson Yong Hong Wang, Heng Ngi Yeo, Yao Long Zhu,
Tow Chong Chong, Teck Yoong Chai, Luying Zhou,
Jit Bitwas. (2006) “Design and development of
Ethernet-based storage area network protocol”,
Computer Communications, Volume 29, Issue 9, 31
May 2006, Pages 1271-1283, ISSN 0140-3664,
10.1016/j.comcom.2005.10.004
Vitesse VSC3144-11, Data Sheet. (2012). Retrieved online
November 2, 2012 from http://www.vitesse.com
Tucker, R. S. (2011), "Scalability and Energy
Consumption of Optical and Electronic Packet
Switching," Lightwave Technology, Journal of, vol.29,
no.16, pp.2410-2421, Aug.15, 2011.
SMARTGREENS2013-2ndInternationalConferenceonSmartGridsandGreenITSystems
52