Prioritized Contention Resolution Scheme for Grid Services over OBS
Networks
Burak Kantarci, Sema Oktug, and D.Turgay Altilar
Istanbul Technical University,
Department of Computer Engineering, Istanbul, Turkey
(bkantarci, oktug, altilar@itu.edu.tr)
Abstract
In this paper a novel contention resolution scheme,
namely Expanded Priority Vector (EPV) approach is
proposed to support differentiated Grid service over
OBS networks. EPV is an enhancement to a previously
proposed application-aware contention resolution
scheme for the OBS-based-Grid services in the
literature. By using the EPV, we keep all the
advantages of the application-aware contention
resolution. Moreover, we introduce fairness among the
contending jobs of the same class by taking care of the
two least significant digits of the vector. Those two
digits pay attention to the remaining distance to the
resource and the job size. By simulations, we show that
EPV based policy reduces the blocking probability for
all classes of submitted jobs.
1. Introduction
As a result of the high computational and storage
demand in scientific applications in the technical
community, local resources became suffering from less
efficient computational power. This phenomenon has
caused the philosophy of Grid Computing to develop
[1]. The most favorable example of an high
computational demand application is the particle
physics experiments. Those experiments require a
computational resource of several petabytes/year and
expected to rise up to several exabytes/year. Therefore
the users are expected to access and use remote
resources that are distributed in different locations
forming a grid.
Basically, the grid consist of heterogeneous
computing and storage units that are connected to each
other via LANs, WANs or metro networks, and that are
distributed into multiple administrative domains.
Besides these, the grid is also supposed to have no
central administration unit.
Several types of data and processing intensive grid
applications can be considered in order to make the
proposed architecture be based on. The first one is high
performance computing and visualization applications.
Multimedia video editing is a good example for this
kind of computationally intensive jobs and that require
TBytes of storage and TFlops of computational
resources. An online visualization application is
another type of job in which the computational
capacity requirement is as much as a few thousands of
GFlops, and a few tens of milliseconds of transmission
and processing latency even though the storage
requirement is not as much as the previous application
example.
Today, in order to build the communication backbone
for the grid, optical networking technologies are
employed [2]. Most of the optical networking
technologies are based on optical circuit switching
(OCS). In OCS, wavelength division multiplexing is
used on the fibers. In order to transmit a job from a
source to destination, the source has to set up a
connection with the destination over a pre-determined
routing path and a wavelength. Unless the connection
is terminated, the transmission resources are dedicated
to the source.
For a high performance computing and visualization
application, the required transmission times of jobs are
as short as a few hundred microseconds or tens of
milliseconds. However, in an OCS architecture, the
main overhead of the job submission into the grid
comes from the connection setup and release times that
are in hundreds of milliseconds. Besides these, since
the resources are devoted to a connection until the
connection is released, OCS leads to a waste of the grid
resources.
The data intensive user applications need high access
bandwidth. As we state above, dedicating the network
resources to a job submission in a connection oriented
structure increases the cost of constructing a grid
dramatically. Although the dedicated grid architecture
provides a job to be submitted at full wavelength
capacity, as a result of the grid service requests' non-
deterministic arrival, this may not be the optimal case.
In many of the grid applications, the job sizes are
significantly short, which leads to the holding time of
Communication Networks and Services Research Conference
978-0-7695-3135-9/08 $25.00 © 2008 IEEE
DOI 10.1109/CNSR.2008.55
232
Communication Networks and Services Research Conference
978-0-7695-3135-9/08 $25.00 © 2008 IEEE
DOI 10.1109/CNSR.2008.55
232
Communication Networks and Services Research Conference
978-0-7695-3135-9/08 $25.00 © 2008 IEEE
DOI 10.1109/CNSR.2008.55
234
Communication Networks and Services Research Conference
978-0-7695-3135-9/08 $25.00 © 2008 IEEE
DOI 10.1109/CNSR.2008.55
234
Communication Networks and Services Research Conference
978-0-7695-3135-9/08 $25.00 © 2008 IEEE
DOI 10.1109/CNSR.2008.55
234