Title: Comparative simulative analysis of WDM LANs for avionics platforms
Full Citation
Permanent Link: http://ufdc.ufl.edu/UF00094700/00001
 Material Information
Title: Comparative simulative analysis of WDM LANs for avionics platforms
Physical Description: Book
Language: English
Creator: Reardon, Casey
Profumo, John
George, Alan D.
Publisher: Reardon et al.
Place of Publication: Gainesville, Fla.
Copyright Date: 2006
 Record Information
Bibliographic ID: UF00094700
Volume ID: VID00001
Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.


This item has the following downloads:

MILCOM2006 ( PDF )

Full Text


Casey Reardon, John Profumo, and Alan D. George
High-Performance Computing and Simulation Research Laboratory
University of Florida
Gainesville, Florida


With their almost unlimited potential for performance and
their decreasing costs, advanced optical components and
networks are now being seriously considered for
deployment in emerging avionics systems. Towards the
goal of developing an advanced avionics network that
features wave-division ,iilhpl % 'i g (WDM) for
performance that is highly scalable, dependable, protocol-
independent, and versatile, many disparate architecture
strategies need to be evaluated. Due to the high cost of
tested prototyping and integration with i~'I g systems,
a simulative approach is used in this study to analyze and
compare candidate WDM LAN architectures at a high
level. Using discrete-event simulation models developed
at the University of Florida, several contrasting
approaches are examined for ,i,,,i ing an optical
network architecture supportive of future avionics
requirements. Each architecture is evaluated in terms of
two application scenarios. The results from the
simulation experiments enable a high-level comparison of
comperingg architectures and provide insight for aerospace
network researchers and designers.


Advanced optical networks featuring wavelength division
multiplexing (WDM) have been targeted as the
technology of choice to realize the desired avionics
network of the future. There are numerous advantages
that WDM networks can offer, some of which include
almost unlimited bandwidth potential, resistance to
electromagnetic interference, and the potential for a
unified network with protocol independence. Despite
these many major advantages, there are just as many
challenges when considering optical network technology
to realize a high-performance, local-area network (LAN).

Traditionally, optical WDM links have been reserved for
long-haul links and high-bandwidth trunks. These cases
are drastically different than the avionics LAN
environment, where links are relatively short and
numerous, and signals need to be routed amongst a large
number of nodes. Additionally, the harsh environment of
aerospace platforms combined with the mission-critical
nature of avionics applications requires the network to be

highly reliable and fault-tolerant. Meeting all of these
requirements presents a difficult challenge to network
designers. Without a widely accepted solution to meet
these requirements, numerous ideas for network designs
need to be formulated and analyzed.

To investigate and realize such an advanced network,
virtual prototyping of potential ideas and architectures to
meet these requirements will be necessary. The
advantages of a computer modeling approach to network
design are obvious, especially when considering the cost
and delays associated with the fabrication and testing of
physical prototypes. In this paper, we evaluate a set of
potential avionics LAN designs using WDM technology.
Those designs are compared via simulative
experimentation and analytical analysis. The results and
analysis presented here provide a high-level comparison
of competing architectures, and we believe represents a
valuable step towards creating an optimized solution for a
WDM avionics network architecture.

The remainder of this paper is organized as follows:
Section 2 provides background information describing the
requirements of a military avionics network, and the
methods used for this research. Section 3 provides details
describing the optical network architectures proposed and
evaluated in this paper. Section 4 describes the
experiments used to compare the proposed architectures.
The results of these experiments are presented and
analyzed in Section 5. Finally, the conclusions from the
work presented are provided in Section 6.


Before attempting to design a network architecture for any
community, it is important to identify the needs and
requirements of the network users. We use the basic
guidelines proposed by the working session problem at the
2005 AVFOP conference when designing and evaluating
network architectures [1]. The requirements described
there included high-speed transmission between any two
nodes on the network, with the number of nodes scaling
up to a maximum of 256. While operation at full scale is
important, it will also be desirable that any network
architecture can be easily implemented for contemporary
avionics platforms, which may often contain far fewer

than 256 nodes. Finally, the network should be fully
functional under the presence of one or two faults, with
graceful degradation after three faults.

All of the network architectures analyzed in this paper
were modeled using the Library for Integrated Optical
Networks (LION). Developed at the University of
Florida, LION provides researchers with an extendable
tool to assess both lower- and upper-layer networking
issues simultaneously by providing a set of accurate
optical components within a powerful network simulation
environment. LION was created in a discrete-event
simulation environment called MLDesigner from
MLDesign Technologies. MLDesigner allows custom
models to be defined using C-based code. Numerous low-
level models can then be combined in a hierarchical
structure to realize complex systems. The components in
LION enable accurate modeling of timing and physical
effects inside optical devices. New and legacy network
protocols can be implemented on top of components to
realize and evaluate almost any system design.


In order to design and realize an optimized network
architecture to meet the current and future needs of the
military avionics community, a number of system designs
need to be evaluated. Thus, we have identified an initial
set of promising optical network designs, which have been
modeled and analyzed. These systems represent a wide
range of design approaches, featuring numerous
topologies and control protocols. Additionally, the extent
to which the networks rely on optics to function varies, as
will be clearly illustrated. Most of the proposed designs
also represent examples of network architectures that have
been previously developed for alternate networking
applications. The remainder of this section is spent
detailing each of the network architectures.

The first network design, the ring-ring architecture, is
illustrated in Figure 1. The basic design of the ring-ring
comes from the ROBUS network architecture [2]. This
network consists of multiple rings of local nodes,
connected by a separate master ring. At the local level,
network nodes are grouped together and connected using
redundant rings. The use of redundant rings allows this
network to maintain full connectivity under the presence
of multiple faults, a key advantage of this architecture.
Each node places data on the ring traveling in both
directions, while being able to receive data from both
directions on the ring. The use of bidirectional
transmission allows each ring to suffer one cut and still
operate at full functionality. Each local ring is equipped

with a ring-leader node, which provides the interface
between the local ring and master ring. This node is
responsible for routing wavelengths between rings, and
removing local traffic from the local ring. The ring-leader
will also perform additional functions, depending on the
control protocol. We allow each local ring to host up to
16 network nodes. Using 16 of these rings will provide
scalability to 256 nodes. A drawback of this approach is
the difficulty in maintaining acceptable optical signal
powers, as each optical signal must reach all local nodes
and ring-leaders. This limitation requires carefully tuned
optical amplifiers throughout the rings.

Figure 1: Ring-Ring Architecture

For the ring-ring architecture, two different control
protocols are employed in this study. The first is a static
time-division, multiple-access (TDMA) protocol. In this
protocol, each destination on the ring is assigned a
wavelength. The same wavelengths can be reused within
each group, so only 16 wavelengths are required for all
local traffic. Longer time slots are used for inter-group
traffic, since this traffic will travel longer distances thus
additional time is required to allow traffic to clear any
fiber links it was using. For smaller-sized systems, such
as a 16-node ring, a simple TDMA system can work well.
Unfortunately, pure static TDMA will break down when
the network scales, and we have to consider traffic
traveling between rings. For inter-ring traffic, we use a
compromise between a TDMA and reservation protocol,
while introducing an optical-electrical conversion. Each
ring is assigned one or two wavelengths to receive traffic
from remote nodes. The groups that must share a
wavelength to reach a remote group use TDMA to
eliminate contention. Within each ring, nodes request the
ring-leader for access to transmit on the desired non-local
wavelength. Once the ring-leader grants access, the
sending node can send during their group's time slot. The
transmitted data is collected and retransmitted by the
remote ring-leader.

The second control protocol we consider is a reservation-
based control protocol (RSVP). Again, each destination
on the ring is assigned a wavelength, and the same
wavelengths can be reused among groups. In this
scenario, nodes that wish to send data issue a request to
the ring-leader on a reserved wavelength. The controller
responds when the requested wavelength is available,
allowing the sender to transmit. When transmission is
complete, the sender notifies the controller again,
releasing its token for that wavelength. Inter-group traffic
is handled in the same way as described with TDMA
protocol. The one exception is that the group controller
does not wait for its assigned time slot to send remote data
to local nodes. The controller instead waits for all
pending requests to the destination to be satisfied, before
transmitting the received data to the destination.

architecture are considered here as well. Each local tree
functions the same as a local ring, and tree-leaders now
replace the ring-leader nodes from before.

The third candidate system represents a hybrid optical-
electrical network architecture. This network is based
largely off of the LAN architecture and solutions provided
by commercial companies such as Matisse Networks [5].
In this architecture, nodes are connected to switches with
electronic links, while optical links are reserved for
communication between switches. The switches in the
network are connected in a ring. By using common
components such as Ethernet of Fibre Channel devices in
this architecture, the costs of implementing this network
are very low. Costs are also minimized by using a very
minimal amount of optical devices. One drawback is the
difficulty of fault-tolerance, as each switch must be
duplicated to overcome most faults. Another drawback is
the limited bandwidth provided to each individual node
where electronic links are used, but this can be overcome
by allocating multiple ports to a single node or device.
Figure 3 provides an illustration of the hybrid architecture.

Figure 2: Optical Tree Architecture

Our second candidate architecture uses optical trees to
connect groups of nodes, while using a star coupler to
connect the distributed trees [3]. Similar to the ring-ring
system previously described, this architecture can be
viewed as a two-level design. Figure 2 provides an
illustration of the optical tree architecture. We allow 16
nodes to be connected on each tree, with 16 trees leading
to a network that can scale to 256 nodes. The tree-leaders,
which are placed at the root of each local tree, will be
asked to perform many of the same functions as the ring-
leader in the previous system. Thus, the tree-leaders
interface each tree with the central coupler, and perform
the necessary wavelength routing. Data produced by each
node is sent up the tree to the root, or tree-leader. The
tree-leader is responsible for routing the wavelength to the
appropriate tree. The tree architecture is not as inherently
fault-tolerant as rings, but there are two advantages the
tree architecture provides. Nodes on the trees will be
simpler, as they only need one interface with the tree.
Second, the lengths between each node and the root are
consistent, unlike in rings, thus time slots are used more
efficiently. Both control protocols used with the ring-ring

Figure 3: Hybrid Architecture

The switches in this case-study contain 32 local ports. By
using 32-port switches, only 8 switches are required for
the network to scale to 256 nodes. Each switch is also
outfitted with two optical ports, each with a laser
transmitter and set of optical receivers. By using multiple
receivers at each optical port, an entire wavelength can be
dedicated for transmission between each pair of switches.
Since optical components are limited to the switches only,
this will not be very costly. Each switch on the optical
ring will filter locally destined wavelengths, while
allowing all other wavelengths to pass through. For each
node, we have assumed a low-latency NIC which requires
4 .is of processing time per packet at each end, not
including any queueing delays within the NIC. While this
setting is faster than seen by most typical commercial
NICs, these values can be viewed as exclusive of



application and transport layer delays that may eventually
increase packet latencies (in all systems).

Our last system architecture uses optical switches to form
a Clos network, as illustrated in Figure 4. First proposed
by Charles Clos in 1953, a Clos network is a highly-
connected multi-stage switched network that provides
multiple paths between end ports, which minimizes or
eliminates blocking in the network [7]. In this system, we
use switches modeled after the time-slotted OSMOSIS
switch architecture developed at IBM, as part of the
DARPA HPCS program [4]. A simplified version of their
architecture is used and presented here.

Figure 4: Optical Clos Architecture

Our Clos network is comprised of eight 32-node perimeter
switches, plus three 8-port backbone switches. Each
connection between two ports includes both electronic and
optical links. The electronic link is used to inform the
switch of transfer requests from the nodes. These requests
include the desired destination and the number of required
time slots needed for the data. An arbiter within each
switch quickly processes these requests, and schedules
transfers in a simple round-robin format in our
implementation. When a requested transfer is scheduled
to take place, the arbiter informs the participating nodes of
this, including how many consecutive time slots the
sender is allotted. If the time slots allotted to the
transmitter are fewer than the number it requested, the
node waits until informed again by the arbiter it is allowed
to transfer. For each new message transmission, nodes
must submit a new request to the arbiter before sending
again. The optical switching is performed using the
broadcast and select approach detailed in [4]. The use of
purely optical switches, combined with a Clos topology
that provides multiple paths between perimeter switches,
provides the highest potential bandwidth of any
architecture proposed here. These advantages come at a
cost though, since the Clos would be the most costly
proposed architecture to implement. Currently the optical
switches themselves are complex and unproven, although

this is likely to change in the coming years as the
technology matures. The cost increases even more when
considering fault-tolerance in a Clos network, which
requires redundant switches all around the perimeter.
Furthermore, the Clos network requires a high degree of
cabling, which may be difficult to achieve in the tight
confines of many aircraft.


In an effort to generate useful results and analysis
predicting network performance, we attempt to create
experiments that will mimic real conditions on an avionics
platform. Thus, two experimental configurations
representative of actual avionics platforms have been
constructed to stimulate our simulation experiments. The
traffic sources in these configurations generate traffic that
is burst, periodic, or random in nature. These two
configurations are briefly described in this section.

The first configuration used in these experiments is our
military configuration. This system is largely based off of
the architecture of the F-22 Raptor, as documented in [6].
The Raptor uses a centralized architecture, where data
acquired from remote sensors and actuators are gathered
and processed at the core processing units. The core
processing system includes two common integrated
processors, or CIPs. Each CIP includes a network of data
and signal processors, memory and other units. Our
military configuration includes two CIP sets in our core
processing subsystem. In total, our military configuration
includes 97 nodes across seven subsystems. In the
baseline case, this system generates about 200 megabits of
traffic per second. Figure 5 gives a top-level diagram of
the military configuration design.

14 Nodes

J_ lj

__ -

16 Nodes

Radar 1 Radar2
(1 ) 16 Nodes 0<| 4 Nodes
Total Noe
Figure 5: Military Configuration

The second configuration used in these experiments is our
commercial configuration. The primary difference
between the two configurations is that the commercial


system is a more distributed architecture than the military
system, which was a centralized one. There is no central
unit in the commercial system used to process critical
data. Each subsystem instead takes on more responsibility
for processing local information. Subsystems are now
free to directly communicate with each other, without
going through central processing. Thus, the network will
see an increased variety of communication between
subsystems. The commercial configuration includes a
total of 84 nodes across eight subsystems. In the baseline
case, this system generates about 300 megabits of traffic
per second. Figure 6 gives a top-level diagram of the
commercial configuration design.

U0 *

9 Nodes

gure 6: Commercial Configuration


In this section the structure and results of the simulation
experiments are presented. The results for each of the
architectures are then discussed, along with comparative
analyses between the candidate architectures. Each of the
six candidate architectures was tested and evaluated using
four different experiments. The first two experiments
consisted of subjecting the architecture to one second of
traffic from the baseline military and commercial
configurations previously described. For the final two
tests, the traffic rates from all of the nodes were scaled by
a factor of 10 to represent demanding traffic loads on
future platforms. As before, one second of network traffic
was again used to stimulate the systems. In all of the
experiments, the payload size of each message was
uniformly distributed between 100 and 2,500 bytes.

The optical transmitters used in every scenario operate at
2.5 Gbps. Additionally, the electronic links used in the
Clos and Hybrid architectures operated at 1 Gbps. The
timeslot period used in both TDMA architectures was 5 us
for local traffic, and 6 us for traffic between rings or trees.
A 500 ns timeslot period was used with the OSMOSIS
switches in the Clos architecture.

The results from all of these experiments are presented in
the following tables and figures. Table 1 summarizes the
overall average packet latencies from each system under
each of the four experimental configurations. Tables 1
and 2 illustrate the overall packet latency for each system
in the baseline military and baseline commercial
experiments, respectively. In addition. Table 2 shows the
worst-case packet latencies for each experiment, which are
determined by taking the average latency of the ten
slowest packets in each test. This ten-packet average was
used so that the result of a single packet would not skew
the perception of the system's overall performance. At the
same time, with each system providing data from tens or
hundreds of thousands of packets, the average of the ten
slowest will generally not be far off from the slowest
packet overall.

Table 1: Average Packet Latency (us)
Ring- Ring- Tree- Tree- Hybrid Clos
Mil- 1,109 223 1,122 230 52 21
Mil- 93,075 65,880 92,973 60,955 78 24
Corn- 4,264 140 4,188 139 40 22
Corn- 113,249 113,640 116,841 113,072 44 23

Table 2: Worst-Case Latency (us)
Ring- Ring- Tree- Tree- Hybrid Clos
Mil- 20,045 3,167 20,045 3,453 365 124
Mil- 100,279 646,698 988,777 703,736 973 264
Corn- 291,607 2,121 289,717 2,203 143 94
Corn- 1,053,385 943,428 1,051,196 943,727 352 109

The results tabulated from all of our simulative
experiments reveal several clear trends. The first major
trend is that the Clos network exhibits the highest
performances in all four experimental setups. The average
packet latency for the Clos networks is less than half that
of the hybrid network in each setup, while performing
several orders of magnitude better than both RSVP and
TDMA cases. The results should not be surprising for
several reasons. The OSMOSIS architecture uses
switches with an optical backplane that can handle up to
60 Gbps of optical data in this configuration. As
importantly, the scheduling in OSMOSIS is very fast and
highly efficient, since each node has its own dedicated
link to the scheduler. This method means that bandwidth
is inefficiently used in most traffic scenarios, which is
illustrated by the low worst-case latency values.
Moreover, a Clos network provides high connectivity
between switches, which increases the total bandwidth the

system can support and provides multiple paths to allow
the network flexibility in how data is routed across it. All
of these factors combined lead to the high performance
observed. Even with these expectations, the performance
of the Clos network is still impressive, as 20 us latencies
for messages averaging 1300 bytes in size is very fast.


Tree-TDM Tree-RSVP Ring-TDM Ring-RSVP


Figure 8: Average Packet Latency
(Baseline Commercial Configuration)

The performance of the Clos network was constant across
all four experiments. This behavior was not observed with
any other candidate architecture, whose average packet
latencies saw very significant increases as the generated
traffic rates were elevated. The results suggest that the
Clos network not only offers the highest performance at
current traffic levels, but will provide superior
performance with even higher traffic demands. It should
also be noted that a Clos network can easily accommodate
additional nodes by simply connecting an additional
switch to the perimeter. Also, if performance was affected
due to contention through backbone switches, a fourth or
fifth switch can be added to the backbone to accommodate
additional bandwidth requirements. However, the high



1,122 1,109




230 223
200 52 21

Tree-TDM Tree-RSVP Ring-TDM Ring-RSVP Hybrid Clos

Figure 7: Average Packet Latency
(Baseline Military Configuration)

,500 188
140 140 40 22


performance and fault-tolerance do come at a significant
price, as described in Section 3.

The hybrid system consistently showed the second best
performance results across all systems. While the average
latencies of the hybrid system were over twice those
exhibited by the Clos network, it still performed orders of
magnitude better than the other four systems. The primary
reason for the large disparity is the efficiency of the
control protocols. Since traffic in data networks is often
bursty in nature, it is important that the network
architecture can handle traffic bursts. Architectures that
can quickly allocate resources these bursts are highly
desirable, such as switched networks which are common
and proven in high-performance networks. The Clos
network takes advantage of the fast arbitration the
OSMOSIS architecture was designed around. In the
hybrid system, the only contention occurs from buffering
packets at the switch and NIC ports.

Despite their efficiencies, the hybrid system lags behind
the Clos system for two major reasons. First, the Ethernet
switches do not provide nearly as much throughput as the
optical switches are capable of providing. Meanwhile, the
Ethernet switches modeled here do not switch as fast as
the optical switches. Despite this limitation, the operation
is highly efficient, as bandwidth in the switches is not
wasted when nodes are inactive. The hybrid system does
degrade in performance when traffic levels were
increased, especially on the military system where the
latency average increased 25 us. This result is largely due
to the strain put on the core processing subsystem's
switch, which sees traffic from all other subsystems.
Since the switching in the hybrid architecture is electronic,
the bandwidth of the switching backplane is limited
compared to optical switches. The strain on the core
processing is further illustrated by the high worst-case
latency in the 10x military experiment. This strain could
be alleviated by using smaller switch modules and
spreading the load at high traffic areas such as the core
processing system. By contrast, expanding the number of
switches significantly would increase inter-switch
communication, and would likely not be desirable for a
network implementation.

As described in Section 3, another major advantage of the
hybrid network is the potential low costs of deployment
on current and future platforms. Since Ethernet is used as
the primary transport in our experiments, this approach
allows the use of commodity components that are very
mature and cost-efficient.

After the Clos and Hybrid architectures, the RSVP control
protocol on both the ring and tree networks showed the

next best results. The RSVP protocol provided average
latencies around 140 and 230 us for the military and
commercial baseline configurations respectively. The
performance of the RSVP protocol suffered greatly when
the traffic rates were increased. These numbers are well
above the Clos and hybrid averages, but much better than
the TDMA averages. Even though the topologies used
with the RSVP protocol do not offer the bandwidth
capabilities of the Clos, they do provide bandwidth
potentials comparable to the Hybrid architecture, and thus
this limit is not the primary cause for decreased
performance. Instead, the long amount of time required to
complete arbitration is the major drawback. Arbitration is
significantly slowed by the TDMA used within the control
wavelength. The slow arbitration process can cause major
delays when several nodes need to send to the same node
at once, causing packets to back up when bursts occur.
Using a faster form of arbitration, such as using electronic
control channels as in OSMOSIS, would greatly increase
the performance potential of those architectures. This
method is almost certainly the most attractive way to
attempt to achieve the performance of the Clos
architecture with the fault-tolerance of rings.

The TDMA protocol showed the worst performance, and
the latencies grew super-linearly as the traffic rates were
increased. The worst-case latencies exceeded a full
second when both configurations' traffic levels were
increased. This outcome can be largely attributed to the
inefficiency of a static TDM-based protocol for a packet-
switched network. Each node has access to only a fraction
of the bandwidth that each node can receive. Thus, even
when there is only one node sending to a destination, it
only utilizes a fraction of the full bandwidth the
destination node can receive. As the number of nodes
increases, smaller fractions of bandwidth must be pre-
allocated, leading to highly inefficient performance,
especially in the presence of bursty traffic. Using more
complex strategies such as statistical TDMA can increase
the efficiency of this protocol, but not the several orders of
magnitude required to be comparable with the Hybrid or
Clos architectures. Thus TDMA is not an attractive basis
for designing a powerful and flexible avionics network.


To arrive at an optimized design for the "irresistible
network" that the military avionics community demands,
numerous designs from disparate paradigms must be
analyzed. A preliminary set of potential designs have
been presented and described in this paper. Those designs
were evaluated and compared via simulative experiments.
The results of those experiments revealed important
insights for designing a WDM avionics network.

While WDM optical networking technology provides a
powerful and effective solution for some networking
applications, a purely passive optical approach to a large
packet-switched LAN for avionics may not be an ideal
approach. Purely optical approaches for dynamically
establishing lightpaths and avoiding collisions between
large numbers of nodes on shared optical mediums thus
far leads to highly inefficient control protocols, and
disappointing performances. This outcome was evidenced
by the poor performance results observed with our ring-
ring and optical tree architectures, each using two
different control protocols. For such architectures to be
used in avionics LANs, significantly more efficient
control protocols will be required.

Without the functionality of buffering and active
switching that electronic networks can provide, a scalable
and flexible LAN with high-performance is difficult to
achieve. In contrast, architectures that use electronic
components to provide additional network functionality,
and thus leverage the strengths of both electronic and
optical network components, seem to offer more promise
for providing the flexible and powerful network solution
the avionics community is currently seeking. The Clos
and Hybrid switched architectures proposed here both fall
into this category, whose performance results were very
promising. Unfortunately, these implementations lead to
complex wiring demands, and require costly solutions to
achieve the desired levels of fault-tolerance.


[1] Avionics Fiber-Optics and Photonics (AVFOP)
Conference, MIT Lincoln Lab, Minneapolis, MN,
Sep. 20-22, 2005.
[2] Gardner, R., et al., "High-Performance Photonic
Avionics Networking Using WDM," MILCOM 1999,
Volume: 2, 31 Oct.- 3 Nov. 1999, pp. 958-962.
[3] Gerla, M., Kova, M., and Bannister, J., "Optical Tree
Topologies: Access Control and Wavelength
Assignment," Computer Networks and ISDN Systems
Journal, Vol. 26, No. 6-8, pp. 965-983, 1994.
[4] Hemenway, R., and Grzybowski, R., "Optical-Packet-
Switched Interconnect for Supercomputer
Applications," Journal of Optical Networking, Vol. 3,
No. 12, pp. 900, Dec. 2004.
[5] Matisse Networks, www.matissenetworks.com
[6] Spitzer, Carl, "The Avionics Handbook," CRC Press
LLC, Boca Raton, Florida, 2001.
[7] Clos, Charles, "A Study of Non-Blocking Switching
Networks," Bell System Technical Journal, March
1953, pp.406-424.

University of Florida Home Page
© 2004 - 2010 University of Florida George A. Smathers Libraries.
All rights reserved.

Acceptable Use, Copyright, and Disclaimer Statement
Last updated October 10, 2010 - - mvs