Content uploaded by Khaled Elleithy
Author content
All content in this area was uploaded by Khaled Elleithy
Content may be subject to copyright.
Content uploaded by Khaled Elleithy
Author content
All content in this area was uploaded by Khaled Elleithy
Content may be subject to copyright.
Content uploaded by Khaled Elleithy
Author content
All content in this area was uploaded by Khaled Elleithy
Content may be subject to copyright.
A New Approach to Avoid Deadlock in Parallel Discrete Event Simulation
(PDES) System
Hemen Patel, Syed S. Rizvi, and Khaled M. Elleithy
Department of Computer Science, University of Bridgeport, Bridgeport 06604, USA
{hemenp, srizvi, elleithy}@bridgeport.edu
Tel: (203) 576-4703
Fax: (203) 576-4766
Keywords:
Conservative distributed simulation, Null
messages, logical process with null message, look ahead
.
Abstract
A conservative distributed simulation requires all logical
processes (LPs) to follow the causality constraint requirement.
This implies that all event-messages are processed in strictly
timestamp order. Apart from the timestamp of each event
generated by LPs, synchronization between all LPs is the
second most important requirements. Finally, there must not
be a deadlock in the distributed environment. A deadlock may
occur when there is no events present in the queue of LP. In
such case, to avoid deadlock, Chandy-Misra-Bryant presented
an algorithm called
Null Message Algorithm
(NMA) [3].
These null messages are passed as an event-message to other
LPs and it stored in one of queues of LPs. This null message
indicates that till the time stamp of that null message, all other
events in the queue which have lesser time stamp than null
message’s time stamp are safe to process. It means that there
won’t be any arrival of any events from that logical process
until current simulation time is equal to the time stamp of the
null message. With the time stamp of the null message, a
Lookahead value is added to the time stamp of that null
message. This Lookahead value can be measure on certain
kind of parameters such as delay to transmit a message,
propagation delay, etc. therefore, calculating value of
Lookahead is the most important part as Lookahead value
affects the performance of the conservative distributed event
simulation. Proper value of Lookahead can reduce the number
of null messages which decreases the traffic of the network. In
this paper, we demonstrate some calculation on the Lookahead
which shows the performance of the distributed event
simulation.
1. INTRODUCTION
The term conservative refers to the causality constraint. In
such kind of distributed simulation, events which are
processed already can’t be rolled back. It shows that all events
presented in the queue of LP must be in sorted time stamp
order or an arrived event’s time stamp must be greater than the
biggest time stamp event presented in the logical process’
queue. All event messages violating causality constraint are
rejected by the receiving logical process. In the simulation,
each logical process has logically different queues for the rest
of logical process present in the distributed environment. All
event messages received by the receiving logical process are
stored in that logical queue which is meant for the sending
logical process. If one of those queue becomes empty then
simulation stops and it is said that deadlock occurred in the
network.
To avoid deadlock, we transmit null messages at certain
simulation time or at certain occurrences of events. Hence,
now it is required to avoid deadlock, we have to send null
message indicating the safe time to process events. Chandy-
Misra-Bryant presented an algorithm called
Null Message
Algorithm
(NMA)
.
[3] This algorithm sends either null
message or event message at each simulation time completion.
In NMA, value of the Lookahead is added to the time stamp
of the null message. This value of Lookahead is critical as it
depends on the transmission time and propagation time of the
distributed environment. We have to choose value for the
Lookahead carefully in order to avoid deadlock and reduce the
number of the null messages. In NMA, value for the
Lookahead is chosen inappropriate. This is one of the main
problems present in NMA. Therefore, numbers of null
messages are increased to more. In this paper, we present a
model that will avoid deadlock and also calculate appropriate
value for the Lookahead.
2. RELATED WORK
There are two approaches in the distributed environment to
process events. One is the conservative approach and the
second one is optimistic approach. Both approaches have their
own advantages and disadvantages. An optimistic approach
has roll back mechanism which is not present in the
conservative approach. Likewise, conservative approach has
causality constraint which does not exist in the optimistic
approach. An approach is different but goal of both
approaches are same to simulate processes in a distributed
environment [4].
Many algorithms are there in conservative algorithm for the
distributed environment simulation.
Chandy-Misra-Bryant
presents an algorithm based on null message which itself is
called null message algorithm.[3] One drawback of that
algorithm is that the timestamp of the null message is chosen
randomly due to which there are null messages in the network
when deadlock is occurred in the network. Bain and Scott tr y
to simplify network topology to re solve problem of null
messages overhead. [1] All these works are done to optimize
the performance of the conservative distributed event
simulations.
3. PROPOSED MODEL
When conservative distributed algorithm is there, there are
few parameters where everyone has to take care of them.
Those parameters are like synchronization, deadlock,
propagation delay, transmission delay, latency, etc. Evaluating
the performance of a distributed simulation algorithm, we
have to consider all those parameters as well as overhead of
null messages. We have made some assumption in order to
make model clearly explain and defined analytically. Those
assumptions are described in the next paragraph of this part. I
have defined certain terms which are mentioned in the table 1
of this paper.
Fig. 1 represents an internal architecture of logical
processes. For the sake of simplicity, we assume that the
simulation system consists of 3 LPs that can directly
communicate with each other. However, in practice, the LP
may reach to an arbitrary value of N. An LP can not only
schedule an event for the neighboring LPs (remote events
generation) but can also schedule an event for itself (local
event generation). Since LPs are connected with each other
using a mesh topology, each LP must have n-1 number of
logical queues. In Fig.1 the total numbers of LPs are 3, it
implies that each LP contains 3-1 that is 2 queues in it. When
LP1 schedule event for LP2, it sends a message to LP1 with
the time stamp. That message is received by LP2 and it is
saved in the logical queue for the LP1 which is present in the
LP2. Likewise, scheduling of events takes place and
simulation cycles finishes. Each message must have causality
constraint, it also applies here. We assumed that Lookahead
value is calculated based on transmission time, and the
propagation time. Therefore, the value of the Lookahead is
directly propositional to the value of transmission time and
value of propagation time. Therefore the equation for the
Lookahead is as below:
L = T
transmission
+ T
propagation
(1)
Transmission time and propagation time is dependent on
the distributed network. Therefore, fix value for them can’t be
taken generally for the all distributed network. Furthermore,
We calculate value of L for the each two neighboring LPs
which are used for calculating timestamp of the null messages.
Also We have assumed that depending upon topology of the
network, value of the Transmission Time and Propagation
Time is calculated dynamically. Therefore, there will be no
furthermore derivation for the Lookahead value.
One more assumption is that, every event message present
in the queue has predefined simulation duration in the unit of
simulation time. Hence, we can able to know that how much
simulation time that all events present in the queue takes.
Therefore, we can calculate that time of duration by the
following equation:
Total
left duration
= ∑ T
event
(2)
This Total
left duration
is calculated for each logical queue
present in the LP. At every simulation time; value for Total
left
duration
is updated. When it is found that Total
left duration
is twice
of the look ahead value, at that simulation time, a signal is
sent to the neighboring LP that queue is going to empty. After
receiving signal from empty queue one’s LP, receiving LP
simply sends a null message to the empty queue one’s LP. By
the simulation time that null message reaches to the empty
queue one’s LP, all remaining events which was before at the
time of sending signal, has been processed by empty queue
one’s LP. And null message will be processed at the next
simulation time.
Hence, we can reduce wait time or ideal time for the LPs.
Let’s make it more clearly with an example. Suppose LP1 has
a logical queue for LP2 in itself. When value of Total
left duration
for the queue LP2 in LP1 is equal or less than value of Look
Ahead L, at that simulation time a signal is sent from LP1 to
LP2 indicating that a queue is going to become empty. At the
receiving end, LP2 sends null message with time stamp T null +
L.
T
ABLE
I
SYSTEM PARAMETERS AND DEFINITIONS
L Look Ahead
T
transmission
Transmission time : Time taken to
generate (send) message
T
propagation
Propagation time : Time taken to travel
from source LP to destination LP by a
message
Total
left duration
Total time of duration of each event
present in the logical queue
T
event
Simulation time that an event take to
execute
T
null
Time stamp of Null Message
Fig1. Logical Processes architecture
To complete this whole process, it takes almost twice of L
simulation time. It implies that after 2L simulation time; null
message reaches to the LP1. And at the same time, LP1 has no
events left in queue. Therefore, LP1 process that null message.
Hence, wait time for LP1 is reduced to nearly 0 simulation
time. All these steps are shown in the Figure 2, 3, 4 and 5. It
clear shows that idle time for LP1 is zero.
Algorithm
While (simulation is not over)
{
If(
Total
left duration
<= L)
{
Send Signal to LP;
}
.
.
Process events which are in Queue;
.
.
}
Given pseudo code is modified version of the original null
message algorithm. While comparing original modified null
message algorithm with original algorithm, we can see the idle
time for LP in the original algorithm is more than idle time for
LP in the modified algorithm. Figure 6, 7, 8 show original
algorithm idle time. In the case of original algorithm LP’s idle
time is 9 for the given example. From this example, it is clear
that original algorithm has more idle time for LPs than my
proposed model which is modified original algorithm.
4. PERFORMANCE EVALUATION
For the sake of performance evaluation and experimental
verifications, we present two cases that incorporate a verity of
different parameters such as Lookahead, frequency of
transitions, and non uniform distribution of Lookahead among
the LPs.
4.1. A Single LP goes dormant
In such case, if dormant LP can send Null Messages to other
LPs then my proposed model works fine. As dormant LP may
receive event scheduled by other LPs, it will become non-
dormant. And if dormant LP can’t send Null Messages to
other LPs then my proposed model can’t work. As dormant LP
can’t send Null Message and deadlock is still remain in the
distributed environment.
Signal Sent
1 1 1 1
xxxx
LP 1 T=1 LP 2
Fig. 8. Scenario at T=1; L=4
Signal Received
111
x x x x
LP 1 T=2 LP 2
Fig. 8. Scenario at T=2; L=4
1 1 Null Msg Sent
x x x x
LP 1 T=3 LP 2
Fig. 8. Scenario T=3; L=4
Null 1 Null Msg Recvd
x x x x
LP 1 T=4 LP 2
Fig. 8. Scenario t=4; L=4
sends signal
x x x x
LP 1 T=1 LP 2
Fig. 4. Original Algorithm Scenario T=1; L=4
Null Msg Sent
x x x x
LP 1 T=5 LP 2
Fig. 4. Original Algorithm Scenario T=5; L=4
Null Null Msg Recvd
x x x x
LP 1 T=9 LP 2
Fig. 4. Original Algorithm Scenario T=9; L=4
4.2. Multiple Output Lines per LP with Non-Uniform
Distribution of Lookahead
For this simulation, we assume that we have single LP that
has O number of output lines where each output line of an LP
can have different value of Lookahead (L). Fig.9 shows the
null message transmission with the following simulation
parameters: simulation time = 500 sec, L is non-uniformly
distributed per output lines (O) of an LP. The numbers of
output lines are varied from 1 to 10. Also, it should be noted
that the value of Lookahead is chosen randomly within the
range of 0 to 1 and assigned to each output line at run time.
This random selection may control the generation of
unnecessary null messages as long as the value is chosen
appropriately.
4.3. Multiple LPs with Multiple Fixed Output Lines with
Different Lookahead Value
For this simulation, we assume that we have multiple LPs
that can have fixed number of output lines where each line of
an LP can have different value of Lookahead (L). Fig.10 shows
the null message transmission with the following simulation
parameters: simulation time = 500 sec, L is non-uniformly
distributed per output lines (O). The numbers of LPs are varied
from 1 to 20 as show in Fig.10. Also, it should be noted that
the value of m and O are both varying quantity for this
particular scenario. This random selection may control the
generation of unnecessary null messages as long as the values
are chosen appropriately. In harmony with our expectation, the
number of null messages increases due to an increase in
number of LPs. However, this increase in null messages is
limited and controlled due to random behavior of Lookahead.
This can also be considered as irregular networks due to the
non uniform distribution.
5.
CONCLUSION
From my modified algorithm, we can reduce idle time for
the LPs. But number of messages or signals transferred is
increased. In other words, line utilization is increased. Idle
time for the LPs is most likely equal to zero but there may be
effect of distributed environment to the idle time. Though idle
time won’t be the more than Lookahead value of the LP for the
proposed model, which is again less than the original
algorithm because for the original algorithm idle time is almost
2 times of the Lookahead value.
REFERENCES
[1] W. L. Bain, and D. S. Scott, "An Algorithm for Time
Synchronization in Distributed Discrete Event Simulation",
Proceedings of the SCS Multiconference on Distributed
Simulation, 19, 3 (February), pp. 30-33, 1988.
[2] J. K. Peacock, J. W. Wong, and E. Manning, ‘‘Synchronization of
Distributed Simulation using Broadcast Algorithms,’’ Computer
Networks, Vol. 4, pp. 3-10, 1980.
[3] K. M. Chandy and J. Misra, "Distributed Simulation: A case study
in design and verification of distributed programs", IEEE
Transactions on Software Engineering, SE-5:5, pp. 440-452,
1979.
[4] Syed S. Rizvi, Khaled. M. Elleithy and Aasia Riasat, “Optimizing
the Performance of NMA for Conservative Synchronization in
Parallel and Discrete Event Simulation”
2 4 6 8 10 12 14 16 18 20
400
600
800
1000
1200
1400
1600
1800
2000
Number Of output lines (O) per LP
Null Message Transmission per LP
Null Messages where 0<L<1 (100 Runs)
Null Messages where 0<L<1 (300 Runs)
Null Messages where 0<L<1 (500 Runs)
Null Messages where 0<L<1 (700 Runs)
Fig.9. Multiple Output Lines per LP with Non-Unifor
m Distribution of
Lookahead versus null message transmission
2 4 6 8 10 12 14 16 18 20
600
700
800
900
1000
1100
1200
1300
1400
1500
Number Of LPs
Null Message Transmission
4 output lines per LP where 0<L<1 (100 Runs)
6 output lines per LP where 0<L<1 (300 Runs)
8 output lines per LP where 0<L<1 (500 Runs)
10 output lines per LP where 0<L<1 (700 Runs)
Fig.10
. Multiple LPs and Multiple Fixed Output Lines with
non-
uniform distribution of Lookahead value versus null
message transmission
