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.
fine-grained temporal constraints (the overrun <strong>of</strong> expected individual activity<br />
durations) may <strong>of</strong>ten take place. Therefore, temporal consistency states need to be<br />
kept under constant monitoring to detect <strong>and</strong> h<strong>and</strong>le potential violations in a timely<br />
fashion. For example, during the second step <strong>of</strong> data pre-processing, delays may<br />
occur in the activity <strong>of</strong> De-dispersion which needs to process terabytes <strong>of</strong> data <strong>and</strong><br />
consumes more than 13 hours computation time. Here, for example, we assume that<br />
the De-dispersion activity takes 14.5 hours (a delay <strong>of</strong> 90 minutes, i.e. around 10%<br />
over the mean duration), then given U ( SW1<br />
) <strong>of</strong> 15.25 hours, there would only be 45<br />
minutes left for the Accelerate activity which normally needs 1.5 hours. Another<br />
example is that during the third step <strong>of</strong> pulsar seeking, we assume that the overall<br />
duration for the FFT Seek, Get C<strong>and</strong>idates <strong>and</strong> Eliminate C<strong>and</strong>idates activities is<br />
108 minutes (a delay <strong>of</strong> 18 minutes, i.e. around 20% over the mean duration). In<br />
such a case, given U ( SW2)<br />
being 5.75 hours, there will probably be a 3 minutes<br />
delay if the subsequent activity completes on time. In both examples, time delays<br />
occurred <strong>and</strong> potential temporal violations may take place.<br />
Therefore, based on the above two examples, we can see that monitoring <strong>of</strong><br />
temporal consistency is very important for the detection <strong>of</strong> temporal violations.<br />
Effective <strong>and</strong> efficient strategies are required to monitor temporal consistency <strong>of</strong><br />
scientific cloud workflow execution <strong>and</strong> detect potential violations as early as<br />
possible before they become real major overruns.<br />
3) H<strong>and</strong>ling temporal violations. If temporal violations like the two examples<br />
mentioned above are detected, temporal violation h<strong>and</strong>ling strategies are normally<br />
required. However, it can be seen that the temporal violations in the two examples<br />
are very different. In the first example with U ( SW1<br />
) , even if we expect the<br />
Accelerate activity can be finished 10 minutes less than its mean duration (around<br />
10% less than the mean duration), i.e. finished in 80 minutes, there would still be a<br />
35 minute time deficit. Therefore, under such a situation, at that stage, some<br />
temporal violation h<strong>and</strong>ling strategies should be executed to decrease the duration <strong>of</strong><br />
the Accelerate activity to at most 45 minutes by, for instance, rescheduling the Taskto-Resource<br />
assignment or recruiting additional resources in order to maintain<br />
temporal correctness. Cleary, in such a case, a potential temporal violation is<br />
detected where temporal violation h<strong>and</strong>ling is necessary. As in the second example<br />
with U ( SW2)<br />
, the 3 minutes time deficit is a small fraction compared with the mean<br />
duration <strong>of</strong> 4 hours for the Fold to XML activity. Actually, there is a probability that<br />
it can be automatically compensated for since it only requires the Fold to XML<br />
activity to be finished 1.25% shorter than its mean duration. Therefore, in such a<br />
situation, though a temporal violation is detected, temporal violation h<strong>and</strong>ling may<br />
6