Figure 3b shows the difference between the predicting
values and the real values of the workload
on the cloud. We compared the proposed HMPM
prediction with some traditional methods, such as
average prediction (AP), weighted prediction (WP),
least squares prediction (LSP), and random prediction
(RP). The traditional methods only consider
the n-step values ahead of the current one, ignoring
the overall system’s current status, which increases
the possibility of missing a prediction. To overcome
this problem, the HMPM includes a two-level state
machine, in which the hidden state machine represents
the overall system’s state and the visible state
machine monitors the change of past-n-step values.
HMPM has the highest accuracy of all the methods
because it’s sensitive to abrupt changes in the number
of requests and successfully filters small changes
in vibration (such as small changes in size or speed),
which is suitable for emergency situation.