Speculative Scheduling of Parameter Sweep Applications Using Job Behaviour Descriptions - Cloud, Grid and High Performance Computing: Emerging Applications - page 80

Structure of complex job behavior descriptions • the average time elapsed between two

subsequent file operations working on the
given file, <simple_description weight=”...”>

...
</simple_description>

• the minimum and maximum file positions

accessed by the job, and the number of
times the job changes these positions.

GENERATING JOB DESCRIPTIONS When the analyser is called with a new activity,

it refines the corresponding file access descrip-
tion by either refining the latest file area access
description of the file access description or by
adding a new file area access description . The
changes in the applied file access methods are
detected through the recalculation of the behaviour
parameters and the comparison of the new values
with the previous ones. If a parameter change is
larger than a specified threshold value, the actual
file area access description will be closed and a
new one will be added to the file access descrip-
tion . For example, if the maximum file position
would be needed to be updated in the case of a
decreasing sequential method, the analyser will
decide that the job stopped using the decreasing
sequential method and it will try to determine
the new method.
The detection of the behaviour changes is based
on the access log which the analyser maintains
for each file accessed by the job. An access log
entry holds the position and size of the datablock
accessed by the job and the time elapsed since the
last file access. The size of the access logs is limited
allowing the analyser to detect and determine the
file access method changes in O (1) time.
In order to determine the new file access
method, the analyser resets all behaviour charac-
terisation parameters and the access log. At this
point, the file access method is undetermined .
After the analyser has processed enough file ac-
cess operations and has filled the access log, it
determines the new method. Please note that the
analyser actually detects changes of file access
behaviour . This means that the new method is
not necessarily a different kind of strategy but
a file access method having different behaviour

Simple job descriptions are generated by the de-
scription generator deployed on the computers the
jobs will run on, the complex job descriptions are
maintained by the description repository service .
In the following the algorithms implemented by
these components are introduced.

Generating Simple Job Descriptions

The “ simple” description is generated during the
job run. The generator monitors the activity of
the job and re-fines the simple description when-
ever the job accesses a ”relevant” resource. Such
monitored activity is the computing (CPU usage)
and file I/O (usage of secondary storage).
The analyser generates simple job descrip-
tions by continuously processing the resource
access information obtained by monitoring. For
each file accessed by the job the analyser builds
a file access description , which consists of one
or more file area access description(s) . A file
area access description presents the file access
strategy used by the job when accessing a specific
part of a file. Throughout the job execution, the
analyser continuously keeps track of the file area
access strategies applied by the job. The analyser
recognises two kinds of file area access methods :
random and sequential . The latter can be both
increasing and decreasing .
Each of these methods is characterised by the
following behaviour parameters :

• the average size of the blocks accessed by

the individual file operations,