Learning Center
Plans & pricing Sign in
Sign Out
Get this document free

Grid computing


Report on Grid Computing

More Info
 Grid computing is a new model of super computing where the jobs are scheduled
and distributed in a network of normal personal computers. Rather than
configuring and developing a super computer, the system takes the help of the
individual computers in the network and performs the jobs. The performance
depends upon many factors such as the processing and the memory capability of
the nodes. Number of jobs already running in the nodes. Cost to route a job from
one system to another in a grid computing environment and so on. Therefore
merely scheduling a job parallel to other computers do not achieve the
performance gain desired.

Hence we take the help of grid computing system. A grid computing is a
combination of computer software and hardware and networking resources. The
objective of the project is to analyze the performance of two major job scheduling
techniques that is : 1) Space Shared and 2) Time shared techniques and to
demonstrate that the space shared techniques are of more efficient in terms of fast
execution of the jobs.
Present System

Grid commuting is relatively a new domain of research. Over the years the computation power
of the CPU has grown quite significantly ( nearly double in every 18 months). Hence a simple
question may easily arise is if the CPU is performing so fast what is the requirement of job
distribution to other nodes as the entire process may take complex resource management and
protocols. The analogy can be drawn from the way chips are designed. The upper limit of
amount of integration in any VVLSI circuit prohibits it from achieving ultra computation speed.
For example a 4 GHZ processor would require nearly 90 billion gates being integrated par gate
on a dual surface making it nearly an impossible aspect. A better way is to have many parallel
processors of some 2 GHZ speed and then design a collaboration technique amongst them.
The CPU will process binary level codes. Similarly in real time applications some information
may be distributed at different servers. They need to be downloaded or uploaded from such
servers. Performing all the tasks in a single system concurrently will not help improving the
speed. Rather if they are done parallel and then the accumulated result is transferred to the
node seeking the result, it becomes more efficient. The restriction of primary memory and the
network access mechanisms in the system makes it difficult for single computers to perform
such jobs. Therefore the Grid computing approach is the desired solution.
Functional Requirements
Process Model

The nodes are active processes. Each process must have it’s own memory. Therefore the node
must be array objects of the node type where the node must define all the features of the job
performing units. In a real time environment the nodes may be hybrid but for the simplicity of
the architecture design they all must be considered as belonging to the unique class.       The
hybrid nature then can be simulated by initializing the nodes with various different parameters
like the computational power, memory etc.

Communication Model

The job performing agents are all active processes as already being discussed. The processes
must be capable of communicating with each other with some communication techniques like
the RPC/ Message passing or Socket driven model. As the core concept adopted here is a
thread based approach where all the threads lie in the same physical memory, a memory
oriented message passing is desirable with data being copied directly into the data space of the

Job Distribution Model

The most critical section of the grid computing is to decide how a job must be distributed and
how far a job must be distributed. Given a set of N available nodes how many nodes and which
are the nodes should get the distributed jobs and what volume of the jobs precisely. The answer
to this question is solved by several complex mathematical models. Some of the algorithms take
into account of the average computation required by each operation of any job where as some
other takes memory into consideration. As the jobs that are being primarily design for system
are relatively low memory jobs, computational power should be the main matrix for job
distribution. The mathematical model for calculating the exact computational need is not
important here as the idea is to provide an interface for such a facility. An arbitrary threshold
based method should be considered in order to show that not all the nodes get the jobs and not
all the jobs are distributed with equality.


The Nodes in general are distributed over the network and the network could be distributed over
either a very small area or a large area. As the computations are performed in Pico seconds
these days and the network propagation delays are calculated in milliseconds, the network
nodes must be considered very close to each other and the delay must be negligible. Otherwise
the delay in transferring the job and getting a job back will be more delay inductive than the
result. As the nodes are considered closer, the routing should be direct. Hence no specific
routing algorithm is required. The situation here can also be considered here as a parallel
computing environment of multiple processor distributed over a common platform like the Intel
64-processor test platform for processor performance monitoring.





Job Distribution

            Figure: Simulator Block Diagram
Data Flow Diagram

                                         Job Router
Job Initiator(User)

                      Gridlet forming



                       Distributed      Network Resource Manager

  R1            R2          Rn

Figure: Data Flow Diagram
Use-Case Diagram




                                               Job Distribution


                    Figure: Use Case Diagram

To top