Grid Resource Allocation - Grid Computing: Infrastructure, Service, and Applications

Information Technology Reference

In-Depth Information

it cannot obtain good performance in lightly loaded system. Ernemann and

Yahyapour [21] proposed a supply-and-demand mechanism that enables

the system to optimize different objectives for different participants. It pro-

vides a robust and l exible approach for handling failures as well as allow-

ing adaptivity during changes.

Figure 7.3 shows the general l ow of economic scheduling. According to

the site-autonomy policy, the total area is separated into a local domain

and a remote domain. In each domain, there is a MetaManager control-

ling the resources. It can act as a broker with the MetaManagers in the

remote domain, and is able to explore them by directory services or peer-

to-peer strategies. All users or clients send requests to a local MetaManager.

Each request can specify the requirements, job attributes, and objective

functions. The request is described by request-and-offer description

language, which is similar to classad, but has its own parameters. After

the local MetaManager received a request, it i rst analyzes the require-

ments of a job and i nds the matched local resources. Local offers are

generated, whereas the best one is kept for further processing. The local

scheduler then forwards the request to other MetaManagers. The remote

schedulers search their domains and send back their best offers to the

original domain. All the offers are compared and the i nal offer is selected.

Finally, the client receives the feedback on the resource allocation and

execution of jobs will be initiated. The MetaManager can reallocate the

schedule to optimize the objective or to recover from system changes. In

order to prevent permanent forwarding of the request, the maximum

number of hops are restricted. The original request is eliminated if no

offer is received in some particular deadlines.

8. Execution of a job

can be initiated by

the MetaManager

Client

Remote