Information Technology Reference
In-Depth Information
Figure 4. Performance of Static Data Feeder scheduler implementations
in more realistic assumptions, and lower execu-
tion times.
the CPUs of the CE and the size and location of
the files the given job would process.
We proposed algorithms for generating the
job behaviour descriptions automatically after
monitoring its resource access. The job behaviour
descriptions generated after subsequent executions
can be composed into a complex description.
By using the complex description the proposed
scheduling algorithms take into account that jobs
can act in different ways when they process dif-
ferent files.
The scheduling of the job, the creation of
its behaviour description, the refinement of the
description and the maintenance of the complex
description are supported by our proposed ar-
chitecture. However, for various non-technical
reasons, it is hard to implement the architecture
in the proposed form in existing production Grid
systems. Therefore, we could implement the com-
ponents of the presented solution only partially.
SUMMARY AND CONCLUSION
In this paper we have presented scheduling algo-
rithms for parameter sweep applications in Grid.
The scheduling algorithms estimate the job finish-
ing times and select the target CE accordingly.
The key for the job completion time estimation
is the description of the behaviour of the job. We
have defined the job behaviour description so that
it characterises the resource access of the job: the
CPU consumption and secondary storage access.
However, the description of a job alone is not
enough to estimate its completion time; informa-
tion about the characteristics and state of each CE
is also required. Such information is the length of
the job wait queue of the CE, the performance of
Search WWH ::
Custom Search