Proceedings of the 2006 Winter Simulation Conference
L. F. Perrone, F. P. Wieland, J. Liu, B. G. Lawson, D. M. Nicol, and R. M. Fujimoto, eds.
ABSTRACT
This paper presents the performance evaluation of a CMB
(Chandy-Misra-Bryant) protocol from the perspective of
execution time. The performance of each logical process in
simulation is measured. Our evaluation shows that logical
processes can have different behaviors and different proto-
cols can be used simultaneously in simulations. While
some logical processes may perform well using conserva-
tive protocols, others can use optimistic protocols because
otherwise most of the time these processes would be
blocked unnecessarily. In order to analyze the behavior of
the simulations some models were simulated using a CMB
implementation called ParSMPLX. These models showed
that each logical process of a simulation has a different be-
havior that makes it more suitable for a specific protocol,
increasing the performance.
1 INTRODUCTION
The simulation technique is a powerful tool to evaluate
performance of computer systems, however it can consume
a lot of time and computational resources. Even with fast
processors, simulations can take many hours to complete.
In order to reduce this time, simulations can be executed in
parallel machines or distributed systems (Dongarra et al.
2002, Kumar et al. 2004, Fujimoto 2003).
One parallel simulation approach is to decompose the
simulation model into logical processes and simulate each
one in a different processor. This approach can reduce the
simulation time for some applications, mainly those which
are easy to become parallel and have large computational
granularity.
Research has been done in distributed simulation tech-
nique and most of these studies focus on two well known
protocols: conservative and optimistic (Bruschi et al. 2004,
Xu and Chung 2004, Bauer et al. 2005, Bononi et al. 2005,
Curry et al. 2005, Lee, Luu and Konangi 2005).
Events in the conservative approach are executed only
when causality errors can be avoided, i.e., when there is
not an event with smaller timestamp than the first event in
the event queue.The optimistic protocol, on the other hand,
simulates all the events without being concerned about
causality errors. If an error occurs, the simulation returns to
a safe state (using rollbacks) and the simulation can con-
tinue.
Various papers have shown that choosing a parallel
simulation protocol is a difficult task because it depends on
a set of factors concerning the model and the computa-
tional platform (Alonso, Frutos and Palacio 1994, Choi and
Chung 1995, Xu and Chung 2004). Computational granu-
larity, model partitioning, load balancing, and lookahead
are a few factors that can influence the simulation per-
formance.
Attempting to predict the performance of a parallel
simulation is a difficult task because different models have
different characteristics. Some models can have good
speedup, even though some logical processes of the simu-
lation are slower than others.
This paper describes experiments that were carried out
with sequential (SMPLX) and conservative distributed
simulations (ParSMPLX) to find out the factors that can
affect the performance of a simulation. The results show
that each logical process of a simulation model has particu-
lar characteristics that make it more suitable for a specific
protocol (conservative or optimistic) depending on factors
such as blocked time, number of null messages exchanged,
and the type of blocking that the logical process can ex-
perience (necessary or unnecessary).
We explain the factors that affect the simulation per-
formance and use them to demonstrate how the perform-
ance can be analyzed in each one of the logical processes.
When these metrics are collected and analysed at run-time,
an opportunity is created for adapting the simulation for
increased speed. The performance we obtained with our
method differs from what has been reported in the litera-
ture.
PERFORMANCE EVALUATION OF A CMB PROTOCOL
Célia L. O. Kawabata
Centro Universitário Central Paulista
Rua Miguel Petroni, 5111
São Carlos, SP 13563-470, BRAZIL
Regina H. C. Santana
Marcos J.Santana
Sarita M. Bruschi
ICMC, Universidade de São Paulo
Av. Trabalhador São-Carlense 400
São Carlos, SP 13560-970, BRAZIL
Kalinka R. L. J. Castelo Branco
Centro Universitário Eurípides
de Marília
Av Hygino Muzzi Filho 529
Marília, SP 17525-901, BRAZIL
1012 1-4244-0501-7/06/$20.00 ©2006 IEEE