Xiao Liu PhD Thesis.pdf - Faculty of Information and Communication ...

commonly used to handle temporal violations. For example, in a scientific cloud
computing environment, the underlying cloud infrastructure can provide unlimited
scalable resource on demand at workflow runtime. Therefore, we can add a new
resource (i.e. a new virtual machine) to execute the local workflow segment for the
violated workflow instances. In that case, these workflow activities can be given
higher priorities (to decrease the waiting time) and/or higher execution speed (to
decrease the execution time), and hence the existing time deficits can be
compensated. To realise that, all the computation tasks and input data of these
activities need to be reallocated to the new resources.
Additional monetary cost and time overheads are required for adding a new
resource at workflow runtime. The monetary cost for adding a new resource in the
pay-for-usage cloud computing environment is mainly equal to the transfer cost
since it is normally free for setting up new resources; the total time overheads
consist of the transfer time for the data and the set-up time for a new service which
is normally around several minutes as in Amazon Elastic Compute Cloud (EC2,
http://aws.amazon.com/ec2/) given the load of the system and network.
Stop and restart: stop and restart consist of two basic steps. The stop step is to
stop the execution at the current checkpoint and store all the running data. The
restart step is to restart the application at a new set of resources. Although the
strategy is very flexible, the natural stop, migrate and restart approach to
rescheduling can be expensive: each migration event may involve large volume of
data transfers. Moreover, restarting the application can incur expensive startup costs,
and significant application modifications may be required (with human interventions)
for specialised restart code. As demonstrated in [28], the overhead for this strategy
includes at least the following aspects: resource selection, performance modelling,
computing resource overhead, application start, checkpoint reading. Among them,
the time for reading checkpoints dominates the rescheduling overhead, as it involves
moving data across the Internet and redistributing data to more processors.
According to their experiments, the average overhead for the strategy is around 500
seconds.
Processor swapping: the basic idea of processor swapping is as follows:
applications are launched with the reservations of more machines than actually used
at workflow build time; and then at runtime, constantly monitor and periodically
126