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
Despite growing popularity of small-scale clusters built out
of off-the-shelf components, there has been little research
on how these small-scale clusters behave under different
scheduling policies. Batch scheduling policies with back-
filling provide excellent space-sharing strategy for parallel
jobs. However, as the performances of uniprocessor and
symmetric multiprocessor have improved with time-
sharing scheduling strategies, it is intuitive that the per-
formance of a cluster of PCs with distributed memory may
also improve with time-sharing strategies, or a combination
of time-sharing and space-sharing strategies. Apart from
the batch scheduling policies, this research explores the
possibilities of using synchronized time-sharing scheduling
algorithms for clusters. This paper describes simulation of
the Gang scheduling policies on top of an existing batch
scheme. The simulation results indicate that time-sharing
scheduler for clusters could exhibit superior performance
over a batch policy.
1 INTRODUCTION
With the increase in processing power and decrease in the
prices of today’s commercial-off-the-shelf (COTS) PCs,
and the increase in the bandwidth of easily available and
affordable Ethernet, “home grown” clusters like Beowulf
cluster have been popularly built for High-Performance-
Computing and Parallel Computing environment. Such
parallel processing finds a wide range of usage among re-
searchers and practitioners. However, often the parallel
systems suffer from under utilization due to inappropriate
choice of scheduling policy. A scheduling policy is used to
settle the conflicts in resource acquisition when a job re-
quires more nodes than that are currently available.
In our previous studies (Rajaei and Dadfar 2005,
2006), various backfilling techniques, namely conservative
and aggressive backfilling (Srinivasan et al. 2002), multi-
ple-queue (Lawson and Smirni 2002) and lookahead
(Shmueli and Feitelson 2003), have been investigated.
Multiple queue backfilling suffers from over-fragmentation
of available nodes in a small-scale cluster (Rajaei and Dad-
far 2006), whereas other techniques seem to be more prom-
ising. The main drawback of backfilling is higher response
time for some jobs which are requesting more resources
particularly in case of aggressive backfilling, and some
jobs may suffer from overestimation of required processing
time.
Gang scheduling (Zang et al. 2003) overcomes the
problem of response time, while its disadvantage is the
global synchronization overhead needed to coordinate a set
of processes. Even though it incurs a heavy context switch
overhead, it may still serve as a viable scheduling alterna-
tive in small-scale clusters. Gang scheduling also offers
various tunable parameters, which can be dynamically
changed according to size and nature of workload. The
number of time slots, which are the interval allocated to a
job during which it runs without preemption, and their du-
ration can be fine-tuned dynamically. Gang scheduling also
supports dynamic prioritization of jobs.
This research simulates Gang scheduling based upon
various policies and tries to find its adaptability to small-
scale cluster. Arrival of the jobs to the system as well as
the payloads and other interesting attributes are randomly
generated. The generated jobs are sent to the simulated
scheduler who in turn activates the desired nodes based on
its current policy.
The goal of this research is to provide a framework
that can be used beyond the current simulation of schedul-
ing policies. We investigate whether a timesharing policy
is suitable also for cluster computing. If the answer is posi-
tive, then we could replace the batch scheduling with an
appropriate scheme satisfying the selection criteria. Other
parameters such as migration policies and cost perform-
ance analysis constitute important research elements.
The rest of this paper is organized as follows. Section
2 describes related work. Section 3 provides details of the
scheduling polices under investigation. The simulation
SIMULATION OF JOB SCHEDULING FOR SMALL SCALE CLUSTERS
Hassan Rajaei
Mohammad Dadfar
Pankaj Joshi
Dept. of Computer Science
Bowling Green State University
Bowling Green, OH 43403, U.S.A.
1195 1-4244-0501-7/06/$20.00 ©2006 IEEE