Content uploaded by Stephen Bush
Author content
All content in this area was uploaded by Stephen Bush on Mar 21, 2013
Content may be subject to copyright.
arXiv:cs/9904012v1 [cs.NI] 21 Apr 1999
Active Virtual Network Management Protocol
Stephen F. Bush
General Electric Corporate Research and Development
KWC-512, One Research Circle, Niskayuna, NY 12309
bushsf@crd.ge.com (http://www.crd.ge.com/people/bush)
Abstract
This paper introduces a novel algorithm, the Active
Virtual Network Management Protocol (AVNMP), for
predictive network management. It explains how the
Active Virtual Network Management Protocol facili-
tates the management of an active network by allow-
ing future predicted state information within an active
network to be available to network management algo-
rithms. This is accomplished by coupling ideas from op-
timistic discrete event simulation with active network-
ing. The optimistic discrete event simulation method
used is a form of self-adjusting Time Warp. It is
self-adjusting because the system adjusts for predictions
which are inaccurate beyond a given tolerance. The
concept of a streptichron and autoanaplasis are intro-
duced as mechanisms which take advantage of the en-
hanced flexibility and intelligence of active packets. Fi-
nally, it is demonstrated that the Active Virtual Net-
work Management Protocol is a feasible concept.
1. Network Management and Active
Networks
The problem this paper addresses is the complex-
ity of managing large and rapidly growing communi-
cation networks. Network management consists of a
wide variety of resp onsibilities including configuration
management, performance management, fault manage-
ment, ac c ounting management, and security manage-
ment. A network management system must be able
to monitor, contr ol, and report upon the sta tus of all
of these areas. This is usually performed using a stan-
dards based management protocol such as the Common
Management Information Protocol (CMIP) [5] or the
Simple Network Management Pro toc ol (SNMP) [9]. A
goal of network management is to pro-actively detect
problems in each of these areas. This means detecting
such events as performance problems and faults before
they occur. This is a c c omplished by the Active Virtual
Network Management Protocol.
Active networks [11] are a relatively recent concept
in communication networks. Active networks are capa-
ble of executing general purpose code within packets as
the packets are transmitted through intermediate net-
work nodes. A framework for supporting the execution
of general purpose code within packets as they travel
through a network is an on-going research effort. Thus
active networks differ from today’s c ommunications
networks becaus e active networks offer a computational
service in addition to a data transport service. In cur-
rent communication netwo rks non-executable data is
passively forwarded through the traditional commu-
nication layers; intermediate devices such as bridges
and routers only access the data link or network head-
ers of packets. In active networks, inter mediate de-
vices can execute generic code within active packets
as they travel through the network. The ability for
communication networks to perform such computation
offers o ppo rtunities for gr eat advantages in such areas
as efficiency, rapid protocol development and deploy-
ment, and network flexibility. However, active net-
works also add additional complexity, particularly in
network management and security. The goal of the
Active Virtual Network Manage ment Protocol is to use
the advantages active networks provide in order to han-
dle the additional complexity in network management.
The Active Virtual Network Management Proto-
col caches predicted values within a State Queue and
makes them available to a standard network manage-
ment interface such as the Simple Network Manage-
ment Protocol (SNMP) [9] as shown in Figure 1. Be-
cause time is appended to the Object Identifier, a
series of Get-Next requests will return all the pre-
dicted cached values of a Management Information
Base (MIB) object. Also note in Figure 1 that the
SNMP agent has the capability to reside within a
packet. Because it is an active network, packets are
capable of issuing management r e quests and respond-
ing to such requests.
getnextresponse 1.3.6.1.x.x.x.x.future
Network Management Client
Active Packet
getnext 1.3.6.1.x.x.x.x.now
Managed Object
Figure 1. Obtaining a Future MIB Object
Value.
The predictive capability provided by the Active
Virtual Network Mana gement Protocol facilitates the
development of a variety of predictive applications from
mobile wireless location management and network se-
curity to improved Quality of Service (QoS). Mobile
systems, especially those using the Global Positioning
System can predict their location. This information
can be propagated through the system via the Active
Virtual Network Management Protoco l. An example of
predictive mobile wireless location management is de-
scribed in [4]. Because the mobile host’s loca tio n can
be predicted, the setup information req uired for hand-
off can be cached ahead of time providing dramatic
increases in the speed o f hand-off and thus improved
QoS. In the area of network security, given a set of
vulnerabilities within a communications network, the
most probable path an attacker will follow can be de-
termined. The Active Virtual Network Management
Protoco l can thus incorporate the effect of an attack
in its prediction and the consequence s of a real or an-
ticipated attack can be propagated through the sys-
tem before it occ urs. In regards to Quality of Service,
time sensitive applications requir e a predictable Qual-
ity of Service [1]. The offered load at the input to the
network can be predicted and the Active Virtual Net-
work Management Protocol will transparently propa-
gate the predicted load through intermediate network
devices. An example of using predictive load manage-
ment is des cribed in [3]. An example of a method for
predicting network traffic based on Wavelets [8] shows
promise and can be used to implement the Driving Pro-
cess (DP) within the Active Virtual Network Manage-
ment Protocol. The Active Virtual Network Manage-
ment Protocol driving process is described in Section
2.1.
2. Active Virtual Network Management
Protocol Description
The Active Virtual Network Management Protocol
algorithm encapsulates each Physical Process within a
Logical Process as illustrated in Figure 2. A Physical
Process is nothing more than an executing task im-
plemented by program code. An e xample of a Physical
Process in a mobile wireless environment is the Rapidly
Deployable Radio Network [10] beam table computa-
tion ta sk. The be am table c omputation task ge nerates
a table of complex weights which controls the angle
of radio beams based on position input. A Logical
Process consists of the Physical Process and additional
data structures and instructions which maintain mes -
sage order and correct operation as the sy stem executes
ahead of real time. These structures are illustrated in
detail in Figure 2. As an ex ample, the beam table com-
putation Physical Process is encapsulated in a Logical
Process which ma intains generated beam tables in its
State Queue and handles rollback due to out-of-order
input messages or out-of-tolerance real messag e s as ex-
plained in Section 2.2. A Logical Process contains a Re-
ceive Queue (QR), Send Queue (QS), and State Q ueue
(SQ) as shown in Figure 2. The simulation compo-
nent and simulation cache on the left side of Figure 2
represent the execution and state based upo n virtual
messages. The real time component and cache repre -
sent the execution and state based upon real mess ages.
The Rece ive Queue maintains newly arriving messages
in order by their Receive Time (TR). The Send Queue
maintains copies of previously sent messages in order
of their send times. The state of a Logical Process
is periodically saved in the State Queue. The Logi-
cal Process also contains its notion of time known as
Local Virtual Time (LVT) and a Tolerance (Θ). The
tolerance is the allowable deviation between actual and
predicted values of incoming messages. For example,
when a real message enters the beam table computa-
tion Logical Process the position in the message value
is compared with the po sition w hich had been cached
in the State Queue of the Logical Process. If these val-
ues ar e out of tolerance, then corrective action is ta ken
in the form of a rollback as explained in Section 2.2.
The Active Virtual Network Management Protocol
Logical Process has the contents shown in Table 1, the
message fields are shown in Table 2, and the message
types ar e listed in Table 3 where t is the real time at
the rec eiving Logical Process. Active Virtual Netwo rk
Management Protocol messages contain the Send Time
(TS), Receive Time (TR), Anti-toggle (A) and the ac-
tual message value itself (M). The Receive Time (TR)
is the time this message is pr edicted to be valid at the
Real-time
Cache
Virtual
Messages
Virtual
Messages
Real-time
Messages
Logical Process
If comparison true, then
continue sending virtual
messages.
If comparison false, then rollback and
send anti-messages.
Anti-messages
Comparison
Simulation
Component
Input
Output
Active Virtual Network Management Protocol (AVNMP) Algorithm
State Queue (SQ)
Real-Time
Messages (RT)
Real Time (RT)
Send Time (TS)
Receive Time (TR)
Anti-Toggle (A)
Sender (S)
Receiver (R)
Message (M)
Real
Time
Λ = maximum difference
between two clocks
(PP)
Θ = maximum beyond which a rollback
occurs back to last known state value
■ Two times message
overhead of conventional
non-predictive network.
With optimization
approximately the
same overhead as
conventional system.
■ Performance determined
by amount of lookahead
and amount of tolerance.
LVT-Local
Virtual Time
(Faster)
Simulation
Cache
Receive
Queue
(QR)
Real
Time
Component
Send Queue (QS)
Figure 2. The AVNMP Logical Process.
destination L ogical Proces s. It is not the link transfer
time from source to destination Logical Process. The
Send Time (TS) is the time this message was sent by
the originating Logical Process. The “A” field is the
anti-toggle and is used for creating an anti-message to
remove the effects of false messages as describe d later.
A messag e also contains a field for current Real Time
(RT). This is used to differentiate a real message from
a vir tua l messa ge. A message which is generated and
time-stampe d with the current time is called a real mes-
sage. Messages which contain future event information
and are time-stamped with a time grea ter than current
time are called virtual messages. If a message arrives
at a Logical Process out of order or with invalid infor-
mation, it is called a false message. A false message
will cause a Logical Process to rollback.
Structure Description
Receive Queue (QR) Messages o rdered
by receive time (TR)
Send Queue (QS) Messages o rdered
by send time (TS)
Local Virtual
Time (LVT) LV T = inf RQ
State Q ueue (SQ) States are periodically saved
Sliding Lookahead
Window SLW = (t, t + Λ]
Tolerance (Θ) Allowable deviation
Table 1. Active Virtual Network Management
Protocol Logical Process Structures.
Field Description
Send Time (TS) LVT of sending process
when mess age is sent
Receive Time (TR) Scheduled time message
is to be received
Anti-toggle (A) Identifies messag e as
normal or anti-message
Message (M) Contents of the message
Real Time (RT) GPS time message originated
Table 2. Active Virtual Network Management
Protocol Message Fields.
2.1. Driving Process
The Active Virtual Network Management Protocol
algorithm requires a Driving Process (DP) to predict
future events and inject them into the system. The
Virtual Message RT > t
Real Message RT ≤ t
Table 3. Active Virtual Network Management
Protocol Message Types.
driving process acts as a source of virtual messages for
the Active Virtual Network Management Protoco l sys-
tem. Virtual messages ar e injected at a rate of λ
vm
messages per unit time and e ach virtual message has
a lo okahead of ∆
vm
time units. Logical Processes re-
act to virtual messages. For example, in the case of
mobile wireless networking, the Global Positioning Sys-
tem receiver process runs in rea l-time providing c urrent
time and loca tion information and has been modified
to inject future predicted time and location messages
as well. Figure 3 shows an example how an active net-
work would appear with Logical Processes and Driving
Processe s deployed. Notice that the driving processes
define the scope of the system, that is, the degree to
which the system is predictable. For example, if only
a particular route is of interest for load prediction pur-
poses, then driving processes may be sent only to adja-
cent tributary nodes of the path. One of the objectives
of this research is to enable the logic al and driving pro -
cesses to automatically and dynamically locate them-
selves in optimal positions within the network.
PP - Physical Process
DP - Driving Process
LP - Logical Process
RM - Real Message
VM - Virtual Message
t+2
t+10
t+5
t
Host
Switch
RM
Router
Switch
Router
LP
PP
PP
LP
LP
PP
DP
DP
VM
VM
VM
Figure 3. Driving and Logical Processes
within a Communications Network.
2.2. Rollback
A rollback is triggered either by messages arriving
out of order at the Receive Queue of a Logical Process
or by a predicted value previously computed by this
Logical Process which is b e yond the allowable tole r-
ance. These are known as false messages because both
types of messages are sources o f error. In either case,
rollback is a mechanism by which a Logical Process re-
turns to a known correct state. The rollback occurs
just as in the original Time Warp algorithm [6]. There
are three phases. In the first phase, the Logical Pro -
cess state is restor e d to a virtual time strictly earlier
than the Receive Time of the false message. In the sec-
ond phase, anti-messages are sent to cancel the effects
of any invalid messages which had been generated be-
fore the arrival of the false message. An anti-message
contains e xactly the same contents as the original mes-
sage with the exception of an anti- toggle bit which is
set. When the anti-message and original message meet,
they are bo th annihilated. The final phase consists of
executing the Lo gical Process forward in virtual time
from its rollback state to the vir tua l time the false mes-
sage a rrived. No messages are cance le d or sent between
the virtual time to which the Logical Process rolled
back and the virtual time of the false message. Because
these messages are correct, there is no need to cancel
or re-send them. This increases performance and it
prevents additional rollbacks. Note that another false
message or anti-message may arrive before this final
phase has completed without causing problems.
3. Streptichrons
In the Active Virtual Network Ma nagement Pr oto-
col architecture describ e d thus far, there is a one-to-one
correspo ndence between virtual messages and real mes-
sages. While this correspondence works well for a dding
prediction to pro tocols using a relatively small portion
of the total bandwidth, it would clearly be beneficial to
reduce the message load, esp e c ially when attempting
to add prediction of the bandwidth itself. Ther e are
more compact forms of representing future behavior
within an active packet besides a virtual message. For
relatively simple and easily modeled systems, only the
model parameters need be sent and used as input to the
logical proc e ss on the appropriate intermediate device.
Note that this a ssumes that the intermediate network
device’s Logical Process is simulating the device op e r-
ation and contains the appropriate model. However,
because the payload of a virtual message is exactly the
same as a real message, the payload of the virtual mes-
sage can be passed to the actual device and the re sult
from the actual device is intercepted and cached. In
this case, the Logical Proce ss is a thin layer of code be-
tween the actual device and virtual messages prima rily
handling rollback. An entire executable lo ad model can
be included within an active packet generated by the
Driving Process and executed by the Logical Process.
When the active packet reaches the target intermediate
device, the load model provides virtual input messages
to the Logical Process and the payload of the virtual
message passed to the actual device as previously de-
scribed. A
bend
z
}| {
Strepti
time
z
}|{
chron is an active packet fa cilitating
prediction as shown in Definition 1 which implements
any of the above mechanisms.
Streptichr on
∆
=
Input Model (Monte-Carlo) Model
Model Parameter s (Self Adjusting)
Virtual Message (Self Adjusting)
(1)
4. Autoanaplasis
self
z
}|{
Auto
adjusting
z
}| {
anaplasis is the self-adjusting characteristic of
streptichrons. One of the virtues of the Active Virtual
Network Management Protocol is the ability for the
predictive system to adjust itself as it operates. This
is accomplished in two ways. When real time reaches
the time at which a predicted value had been cached,
a comparison is made between the re al value and the
predicted value. If the values differ beyond a given tol-
erance, then the Log ical Process rolls backward in time.
Also, active packets which implement virtual messages
adjust, or refine, their predicted values as they travel
through the network. As a spe c ific example consider
a streptichron in the form of a simple virtual message
which anticipates load as illustrated in Figure 4. Al-
though the virtual message r e presents a messa ge ex-
pected to exist in the future, the virtual message im-
plementation exists in real time providing feedback to
the predictive system. For example, as the virtual mes-
sage travels from one intermediate node to another, the
packet computes its transfer time and compares it with
the predicted state at that time of the intermediate log-
ical process causing a rollback if it is out of tolerance.
Also, as the streptichron travels through the network
and as real time approaches the streptichron’s Receive
Time, the streptichron refines its predicted value.
5. Class Hierarchy
The software architecture and state of the c ode un-
der development are shown in Figure 5. Time is crit-
ical in the architecture of the Active Virtual Network
Management Protocol system; thus, most classes ar e
derived from class Date. Class AvnmpTime handles
SmallState
Error = Actual - Predicted
Predicted
Virtual Message
Source
DataCode
Node OS
Network Device (switch, router, hub, etc...)
Destination
Update Data
Data
Adjusted
DataCode
Virtual Message
SNMP Client
Figure4.AStreptichronRefinesitsPrediction
as it Travels Through the Network.
relative time operations. Class Gvt uses active the Gvt-
Packets class to calculate global virtual time. Class
AvnmpLP handles the bulk of the proce ssing including
rollback. Class Driver generates and injects real and
virtual messages into the system. The PP clas s either
simulates, or accesses, an actual device on behalf of the
Logical Process. The PP class may not need to simu-
late the device because the payload of a virtual message
is exactly the same as a real message; thus, the pay-
load of the virtual message can be passed to the actual
device and the r e sult from the actual devic e is inter-
cepted and cached. In this case, the Logical Process is
a thin layer of code between the actual device accessed
by the PP class. The GvtPacket class implements the
Global Virtual Time packet which is exchanged by all
logical a nd driving process e s to deter mine globa l vir-
tual time. The Streptichron class has been discussed in
Section 4. Currently only the virtual message form of
a streptichron ha s b e e n implemented. The active pack-
ets have been implemented in both ANTS [11] and KU
SmartPackets [7].
6. Performance Analysis
This section analyzes the benefit of Active Virtual
Network Management Protocol in terms of speedup
based upon accurately predicting system behavior.
There are many factors which influence s peedup includ-
ing out-of-order message pro bability, out-of-tolerance
state value probability, rate of virtual messages enter-
ing the system, task execution time, tas k partitioning
into Logical Processes, rollback overhead, prediction
accuracy as a function of distance into the future which
predictions are attempted, and the effect of parallelism
and optimistic synchronization. All of these factors are
considered in this section beginning with a direct anal-
ysis using the definitions from optimistic simulation.
KU SmartPackets
State
Small
Streptichron
GvtPacket
Abstract Class
Date
AvnmpTime
Gvt
AvnmpLP
Driver PP
SNMP Client
SNMP Agent
Figure 5. Active Virtual Network Management
Protocol Class Hierarchy.
The definition of Global Vir tua l Time (GVT) can be
applied to determine the relationship among expec ted
task execution time (τ
task
), the real time at which the
state was cached (t
SQ
), and real time (t). Consider the
value (V
v
) which is cached at real time t
SQ
in the State
Queue (SQ) resulting from a particular predicted event.
The state queue values may be repeatedly added and
discarded as Active Vir tua l Network Management Pro-
tocol opera tion proceeds in the presence of rollback. As
rollbacks occur, values for a particular predicted event
may change, converging to the real value (V
r
). For
correct operation of Active Virtual Network Manage-
ment Protocol, V
v
should approach V
r
as t approaches
GV T (t
1
) where GV T (t
1
) is the GV T of the Active Vir-
tual Network Management Protocol system at time t
1
.
Explicitly, this is for all ǫ > 0 there exists δ > 0 such
that |GV T (t
1
)− t| < δ implies |f (t)−f (GV T (t
1
))| < ǫ
where f (t) = V
r
and f (GV T (t
1
)) = V
v
. f (t) is the
prediction function of a driving process. The purpose
and function of the driving process has been explained
in Sectio n 2. Because Active Virtual Network Manage-
ment Protocol will always use the correct value when
the predicted time (τ) equals the current real time (t)
and it is assumed that the predictions will become mor e
accurate as the predicted time o f the event approaches
the current time, the reasonable assumption is made
that lim
τ →t
f(τ ) = V
v
. In order for the Active Vir-
tual Network Management Protocol system to always
look ahead, for all t GV T (t) ≥ t. This means that
for all n ∈ {LP s} and for all t LV T
lp
n
(t) ≥ t and
min
m∈{M}
{m} ≥ t where m is the receive time of a
message, M is the set of messages in the entire s ystem
and LV T
lp
n
is the L ocal Virtual Time of the n
th
Logical
Process. In other words, the Local Virtual Time (LVT)
of each Logical Process (LP) must be greater than or
equal to real time and the smallest message Receive
Time not yet processed must als o be grea ter than o r
equal to real time. The smallest message Receive Time
could c ause a rollback to that time. This implies that
for all n, t L V T
dp
n
(t) ≥ t. In other words, this implies
that the Local Virtual Time (LVT) of each driving pro-
cess must be g reater than or equal to real time. An
out-of-order rollback occurs when m < LV T (t). The
largest saved state time such that t
SQ
< m is used to
restore the state of the Logical Pr ocess, where t
SQ
is
the real time the state was saved. Then the expected
task execution time (τ
task
) can take no longer than
t
SQ
− t to complete in order for GV T to remain ahead
of real time. Thus , a constraint be tween expected
task execution time (τ
task
), state save time (t
SQ
), and
real time (t) has been defined. There are three pos-
sible cases to consider when determining the speedup
of the Active Virtual Network Management Pro tocol
over non-lookahead sequential execution. In this sec-
tion we will determine the speedup given each of these
cases and their respective probabilities. These cases
are illustrated in Figures 6 through 8. The time that
an event is predicted to occur and the result cached is
labeled t
virtual event
, the time a real event occurs is la-
beled t
real event
, and the time a result for the real event
is calculated is labeled t
no−avnm p
. In the Active Vir-
tual Network Management Protoco l, the virtual event
and its result can be cached before the real event as
shown in Figure 6, between the real event but before
the real e vent result is calculated shown in Figure 7,
or after the rea l event result is ca lculated as shown in
Figure 8. In each ca se, all events are considered rela-
tive to the occurrence of the real event. It is assumed
that the real event occurs at time t. A random variable
called the lookahead (LA) is defined as LV T − t. The
virtual event occurs at time t − LA. Assume that the
task which must be executed once the real event oc-
curs takes τ
task
time. Then without the Active Virtual
Network Management Pro tocol the task is completed
at time t + τ
task
.
timet-LA
t
virtual event t
no-avnmp
t+
task
τ
real event
t
t
Figure 6. Active Virtual Network Management
Protocol Cached before Real Event.
timet t-LA
t
real event
t
virtual event t
no-avnmp
t+
task
τ
Figure 7. Active Virtual Network Management
Protocol Cached later than Real Event.
timet
t
real event
τ
task
t+
no-avnmp
t virtual event
t
t-LA
Figure 8. AVNMP Cached slower than Real
Time.
The prediction rate of a Logical Process is the rate
of change of the pr ocesses’ Local Virtual Time with re-
sp e c t to real time, thus P R is the Active Virtual Net-
work Management Protoc ol speedup. The prediction
rate has been derived in [2] and is defined in Equation 2 .
S
parallel
is the speedup due to parallelism, ∆
vm
is the
lookahead per virtua l message, λ
vm
is the rate at which
virtual messages are injected by a driving process, X
and Y are random variables representing the propor-
tion of out-of-order and out-of-tolerance messages re-
sp e c tively, τ
task
is the expected amount of time taken
by the physical process to process an input message,
and τ
rb
is the expected time taken by a Lo gical Process
to ro llback. Equation 2 includes the time to predict an
event and cache the result in the State Queue (SQ).
In [2 ] the expected value of X has been determined
based on the inhere nt sy nchronization of the Logical
Process topology. It is shown in [2] that X has an ex-
pected value which varies with the rate of hand-offs in
the mobile wireless location management application.
It is clear that the propor tio n of out-of-order messages
is dependent on the Logical Process (LP) architecture
and the par titioning of tasks into Logical Processes.
Thus, it is difficult in an experimental implementation
to vary X. It is easier to change the tolerance rather
than change the Logical Process architecture to eval-
uate the performance of the Active Virtual Network
Management Protocol. For these rea sons, the analy-
sis proce e ds with P R
X,Y |X=E[X]
. Since the prediction
rate is the rate of change of Local Virtual Time with
respect to time, the value of the Local Virtual Time is
shown in Equatio n 3 where C is an initial offset. This
offset may occur bec ause the Active Virtual Network
Management Pr otocol may beg in running C time units
befo re or after the rea l system. Replacing LV T in the
definition of LA with the right side of Equation 3 yields
the Eq uation for lookahead shown in Equation 4.
LV T
X,Y |X=E[X]
= (3)
λ
vm
(∆
vm
S
parallel
− τ
task
− (τ
task
+ τ
rb
)E[X]−
(∆
vm
S
parallel
−
1
λ
vm
)Y )t + C
LA
X,Y |X=E[X]
= (LV T
X,Y |X=E[X]
− 1)t + C (4)
The probability of the event in which the Active Vir-
tual Network Management Protocol result is cached
befo re the real event is defined in Equatio n 5. The
probability of the event for which the Active Virtual
Network Manage ment Protocol result is cached after
the real event but before the result would have been
calculated in the no n-Active Virtual Network Manage-
ment Protocol system is defined in Equation 6. Finally,
the probability of the event for which the Active Vir-
tual Network Management Protocol result is cached
after the result would have been calculated in a non-
Active Virtual Network Management Protocol system
is defined in Equation 7.
P
cache
= P [LA
X,Y |X=E[X]
> τ
task
] (5)
P
late
= P [0 ≤ LA
X,Y |X=E[X]
≤ τ
task
] (6)
P
slow
= P [LA
X,Y |X=E[X]
< 0] (7)
The goal of this analysis is to determine the effect
of the proportion of out-of-tolera nce messages (Y ) on
the speedup of a Active Virtual Network Management
Protoco l system. Hence we assume that the proportion
Y is a binomially distributed random variable with pa -
rameters n and p where n is the total number of mes-
sages and p is the probability of any single message
being out of tolerance. It is helpful to simplify Equa-
tion 4 by using γ
1
and γ
2
as defined in Equations 9 and
10 in Equation 8.
LA
X,Y |X=E[X]
= γ
1
− γ
2
Y (8)
γ
1
= (λ
vm
∆
vm
S
parallel
− λ
vm
τ
task
− (9)
λ
vm
(τ
rb
+ τ
task
)E[X]) − 1 )t + C
γ
2
= λ
vm
(∆
vm
S
parallel
−
1
λ
vm
+ τ
rb
)t (10)
The early prediction probability as illustrated in
Figure 6 is shown in Eq uation 11. T he late predic-
tion probability as illustrated in Figure 7 is shown in
Equation 12. The probability for which Active Vir-
tual Network Manage ment Pro tocol falls behind real
time as illustrated in Figure 8 is shown in Equation
13. The three cases for determining Active Virtual
Network Manage ment Pro tocol speedup are thus de-
termined by the probability that Y is greater or less
than two thresholds.
P
1
(t) = P
cache X,Y |X=E[X]
= P
Y <
γ
1
− τ
task
γ
2
(11)
P
2
(t) = P
late X,Y |X=E[X]
= P
γ
1
− τ
task
γ
2
≤ Y ≤
γ
1
γ
2
(12)
P
3
(t) = P
slow X,Y |X=E[X]
= P
Y >
γ
1
γ
2
(13)
The three probabilities in Equations 11 through 13
depend on (Y ) and real time because the analysis as-
sumes that the lookahead increases indefinitely which
shifts the thresholds in such a manner as to increase
Active Virtual Network Management Protocol perfor -
mance as real time increases. However, in this analysis,
it is assumed that a sliding lookahead window exists
on each Logical Process to control the lookahead rate
locally. The Active Virtual Network Management Pro-
tocol algorithm holds processing of virtual messages
once the end of the Sliding Lookahead Window (SLW)
is reached. The hold time occurs when LA = Λ where
Λ is the length of the Sliding Lookahead Window. Once
Λ is reached, processing of virtual messages is discon-
tinued until real-time re aches L ocal Virtual Time. The
lookahead versus real time including the effect of the
Sliding Lookahead Window is shown in Figure 9. The
dashed arrow represents the lookahea d which increases
at rate P R. The solid line returning to zero is looka-
head as the Logical Process delays. Because the curve
in Figure 9 from 0 to t
L
repeats indefinitely, only the
area from 0 to t
L
need be considered. For each P
i
(t)
i = 1, 2, 3, the time average over the lookahead time
(t
L
) is shown by the integral in Equation 14.
P
X,Y |X=E[X]
=
1
t
L
Z
t
L
0
P
i
(t)dt (14)
The probability of each of the events shown in Fig-
ures 6 through 8 is multiplied by the speedup for
each event in order to derive the average speedup.
For the case shown in Figure 6 , the speedup (C
r
) is
provided by the time to read the cache over directly
computing the result. For the remaining cases the
sp e e dup is P R
X,Y |X=E[X]
which has been defined as
LV T
X,Y |X =E[X]
t
as shown in Equation 15. The analyt-
ical results for speedup are graphed in Figure 10. A
high probability of out-of-tolerance rollback in Figure
P R
X,Y
= (2)
λ
vm
∆
vm
S
parallel
− τ
task
− (τ
task
+ τ
rb
)X − (∆
vm
S
parallel
−
1
λ
vm
)Y
t
h
LA
0
real time
Λ
Λ
t
L
Figure 9. Lookahead with a Sliding Looka-
head Window.
10 results in a speedup of less than one. Real messages
are always processed when they arrive at an Logical
Process. Thus, no matter how late Active Virtual Net-
work Management P rotocol re sults are, the sy stem will
continue to run near real time. However, when Active
Virtual Network Management Protocol results are very
late due to a high proportion of out-of-tolerance mes-
sages, the Active Virtual Network Management Pro-
tocol system is slower than real time because out-of-
tolerance rollback overhead proces sing occurs. Anti-
messages must be s e nt to cor rect other Logical Pro-
cesses which have processed messages which have now
been found to be out of tolerance from the current
Logical Process. This ca uses the speedup to be less
than one when the o ut-of-tolerance probability is high.
Thus, P R
X,Y |X=E[X]
< 1 for the “slow” case shown in
Figure 8.
Equation 16 shows the c omplete Active Virtual Net-
work Management Protocol per fo rmance utility. The
surface plot showing the utility of Active Virtual Net-
work Management Pro tocol as a function of the propor-
tion of out-of-tolerance mes sages is shown in Figure 11
where Φ
s
, Φ
w
, Φ
b
are one. The parameters are chosen
to accommodate the mobile wireless network applica-
tion desc ribed in [2]. Virtual messages are injected by
the Global Positioning System at a rate of 0.03 per
millisecond (λ
vm
= 0.03) with a lookahead o f 30.0 mil-
liseconds (∆
vm
= 30.0). The expected time to create
a beam table is 7.0 milliseconds (τ
task
= 7.0). The ex-
pected rollback time is 1.0 milliseconds (τ
rb
= 1.0) and
the speedup in reading from cached results over com-
puting the beam table is 100 (C
r
= 100). The y-axis is
the relative marginal utility of speedup over reduction
0
5
10
15
20
25
30
35
40
45
50
0 0.2 0.4 0.6 0.8 1
AVNMP Speedup
Expected Proportion of Out-of-Tolerance Messages (E[Y])
AVNMP Speedup Analysis
Figure 10. AVNMP Speedup.
in bandwidth overhead SB =
Φ
s
Φ
b
. Thus if bandwidth
reduction is much more important than speedup, the
utility is low and the propo rtion of rollback messages
would have to be kept below 0.3 in this case. How-
ever, if speedup is of higher priority relative to band-
width, the proportion of out-of-tolerance rollback mes-
sage values can be as high as 0.5. If the proportion of
out-of-tolerance messages becomes too high, the utility
becomes negative because prediction time begins to fall
behind real time.
The effect of the proportion of out-of-order and
out-of-tolerance messages on Active Virtual Network
Management P rotocol speedup is shown in Figure 12.
This graph shows that out-of-tolerance rollbacks have a
greater impact o n speedup than out-of-order rollbacks.
The reas on for the greater impact of the proportion of
out-of-tolerance messa ges is that rollbacks caused by
such messages always cause a process to rollback to
real time. An out-of-order rollback only req uires the
process to rollback to the previous saved state.
Figure 13 shows the effect of the pr oportion of vir-
tual messages and expected lookahead per virtual mes-
sage on spee dup. This graph is interes ting because it
shows how the proportion of virtual messages injected
into the Active Virtual Network Management Proto-
col system and the expected lookahead time of each
message can affect the speedup. The real and virtual
message rates are 0.1 messages per millisecond and the
remaining pa rameters remain a s previously stated.
η ≡ (15)
P
cache X|X=E[X]
C
r
+ (P
late X|X=E[X]
+ P
slow X|X=E[ X]
)P R
X,Y |X=E[X]
U
AV N MP
=
P
cache X|X=E[X]
C
r
+ (16)
(P
late X|X=E[X]
+ P
slow X|X=E[ X]
)P R
X,Y |X=E[X]
Φ
s
−
P [|AC
t
(Λ)| > Θ]Φ
w
−
λ
v
λ
rb
+ λ
v
λ
r
!
Φ
b
AVNMP Utility
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.5
1
1.5
2
-10
0
10
20
30
40
50
60
E[Y]
SB
AVNMP Utility
Figure 11. Overhead versus Speedup as a
Function of Probability of Rollback.
Effect of Rollback AVNMP
0
0.5
1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
5
10
15
20
25
30
E[X]
E[Y]
Speedup
Figure 12. Effect of Non-Causality and Toler-
ance on Speedup.
Effect of Virtual Message Timing on AVNMP
30
35
40
45
50
55
60
65
70
75
80
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
5
10
15
Expected Lookahead per Message
Virtual Message Rate
Speedup
Figure 13. Effect of Virtual Message Rate and
Lookahead on Speedup.
7. Summary
This paper ha s introduced a novel algorithm, Active
Virtual Network Management Protocol (AVNMP), for
predictive network management. It e xplained how the
Active Virtua l Network Management Protocol facili-
tates the mana gement of an active network by allow-
ing future predicted sta te information within an active
network to be available to network management algo-
rithms. This has been a c c omplished by coupling ideas
from optimistic discrete event simulation with active
networking. The optimistic discrete event simulation
method us ed is a form of s e lf-adjusting Time Warp.
The c oncept of a streptichron and autoanaplasis have
been introduced as mechanisms which take advantage
of the enhanced flexibility and intelligence of active
packets. Finally, it has been demonstrated that the
Active Virtual Network Management Protocol is a fea-
sible concept.
Future work on the Active Virtual Network Manage-
ment Protocol will focus on dynamic message repri-
oritization such that both real and virtual messages
are proce ssed with optimal priority as well as auto-
matic dynamic deployment of the Active Virtual Net-
work Management Protocol logical and driving pro-
cesses within a network. Wor k will also continue on
finding synergies between the self-adjusting Time Warp
algorithm and active networks in order to o ptimize the
Active Virtual Network Management Pro tocol.
References
[1] R. Braden, D. Clark, and S. Shenker. Integrated ser-
vices in the internet architecture: an overview, June
1994. RFC #1633.
[2] S. F. Bush. The Design and Analysis of Virtual Net-
work Configuration for a Wireless M obile AT M Net-
work. PhD thesis, University of Kansas, Aug. 1997.
[3] S. F. Bush, V. S. Frost, and J. B. Evans. Network Man-
agement of Predictive Mobile Networks. In IEEE Sym-
posium on Pl anni ng and Design of Broadband Net-
works, October 1996. Session 8, Montebello, Quebec,
Canada.
[4] S. F. Bush, S. Jagannath, J. B. Evans, V . Frost,
G. Minden, and K. S. Shanmugan. A Control and
Management Network for Wireless ATM Systems.
ACM-Baltzer Wireless Networks (WINET), 3:267,283,
1997. URL: http://www.ittc.ukans.edu/∼sbush.
[5] ISO. Open Systems Interconnection - Management
Protocol Specification - Part 2: Common Management
Information Protocol.
[6] D. R. Jefferson and H. A. Sowizral. Fast Concurrent
Simulation Using The Time Warp Mechanism, Part
I: Local Control. Technical Report TR-83-204, The
Rand Corporation, 1982.
[7] A. B. Kulkarni, G. J. Minden, R. Hill, Y. Wijata,
S. Sheth , H. Pindi, F. Wahhab, A. Gopinath, and
A. Nagarajan. Implementation of a Prototype Active
Network. I n OPENARCH ’98, 1998.
[8] S. Ma and C. Ji. Wavelet models for video time-series.
In M. I. Jordan, M. J. Kearns, and S. A. Solla, editors,
Advances in Neural Information Processing Systems,
volume 10. The MIT Press, 1998.
[9] M. T. Rose. The Simple Book, An Introduction to
the Management of TCP/IP Based Internets. Prentice
Hall, 1991.
[10] R. Sanchez, J. Evans, G. Minden , and V. S. F. K. S.
Shanmugan. RDRN: A Rapidly Deployable R adio
Network - Implementation and Experience. In ICUPC
’98, 1998.
[11] D. L. Tennenhouse, J. M. Smith, W. D. Sincoskie,
D. J. Wetherall, and G. J. Minden. A survey of active
network research. IEEE Communications Magazine,
35(1):80–86, Jan. 1997.