Xiao Liu PhD Thesis.pdf - Faculty of Information and Communication ...
Xiao Liu PhD Thesis.pdf - Faculty of Information and Communication ...
Xiao Liu PhD Thesis.pdf - Faculty of Information and Communication ...
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
to 5 which is normally large enough to compensate for time deficit with the given<br />
compensation rate. The success rate for temporal violation h<strong>and</strong>ling is set as 80%.<br />
CSS TD is applied with default values as defined in [22]. The fixed confidence<br />
threshold for RA is set as 0.9 (i.e. select 10%) so as to be comparable to our strategy,<br />
denoted as AD . The initial probability threshold <strong>and</strong> the update rate for our strategy<br />
are set as 0.5 (mid-range initially) <strong>and</strong> 5% (a small percentile) respectively. This is a<br />
moderate <strong>and</strong> reasonable setting when there is no prior knowledge. Additionally, to<br />
reflect the effectiveness <strong>of</strong> temporal violation h<strong>and</strong>ling, the results <strong>of</strong> workflow<br />
execution under the natural condition without temporal violation h<strong>and</strong>ling (denoted<br />
as NIL ) are also presented.<br />
Here, before we demonstrate the experiments on temporal violation h<strong>and</strong>ling<br />
point selection in detail, we present an overview <strong>of</strong> the results on the cost for<br />
h<strong>and</strong>ling temporal violations in scientific workflow systems. The two representative<br />
temporal violation h<strong>and</strong>ling strategies investigated in our system are workflow<br />
rescheduling <strong>and</strong> extra resource recruitment. Workflow rescheduling is to<br />
compensate the time deficit by optimising the current activity-resource assignment.<br />
The major overhead involved in workflow rescheduling includes activity transfer<br />
(data <strong>and</strong> activity definitions) between resources, <strong>and</strong> the computation overhead <strong>of</strong><br />
the rescheduling strategy itself. In our SwinDeW-C scientific workflow system, for<br />
a typical workflow rescheduling scenario (with around 200 activities <strong>and</strong> 10<br />
resources), the average overhead for activity transfer is around 2 minutes <strong>and</strong> the<br />
computation overhead is around 0.2 minutes [56]. As for extra resource recruitment,<br />
it is to compensate the time deficit by employing additional resources for the<br />
violated workflow instances at runtime. Its major overhead includes the set-up time<br />
for new resources <strong>and</strong> the task transfer time. For a typical resource recruitment<br />
scenario (adding one new resource), the normal set-up time for a single resource is<br />
several minutes (like the Amazon EC2 Reserved Instances:<br />
http://aws.amazon.com/ec2/reserved-instances/) <strong>and</strong> the activity transfer time (for<br />
fewer than 5 activities in a local workflow segment) is around 0.3 minutes.<br />
Furthermore, to give a direct view for the savings on the total temporal violation<br />
h<strong>and</strong>ling cost, we record the reduced overhead for the 10 test cases in each round <strong>of</strong><br />
the experiment. For example, if we specify the basic time unit as one second, the<br />
results have shown that for a large scale scientific workflow (with over 10,000<br />
114