I am perplexed by a use of ARIMA seen in a paper in the literature.
For a time interval T1, a periodic workload was generated for virtual machines.
For a training time interval T1, data was gathered about resource usage and power consumption. It is not stated whether any migration took place. This data was used to construct an ARIMA model.
For a (subsequent) testing time interval T2, the ARIMA model was put to use to predict resource usage and power consumption.
However, during T2, a migration was forced. Can an ARIMA model predict about an event unless it took place during the training interval T1?
Hi Etienne
So, if I understand correctly, there are two aspects to answering this. Firstly, I don't think the ARIMA would be able to predict the event.... especially also as ARIMA models are by their very nature continuous in their state-space, whereas a migration, if I understood correctly, is a binary flag (happened or not).
That being said, if after the migration, the dynamics of the system do not change (just a disruption), the ARIMA model should be eventually able to keep predicting states... it would be as if the initial conditions changed...
This is my understanding, although I must say I am not an expert in either ARIMA models or virtual machine technology.
Etienne, I do agree with Michael's answer. If the characteristics of the new system do not deviate too much from the previous one, the model might still provide a limited prediction error. However, from our previous discussions, especially regarding the power consumption, the situation might change a lot owing to the presence of other active VMs in the new system, for example. In the paper, do they provide any instance of potential deviation and recovery after a new training period?
Franco Davoli Unfortunately, the authors do not provide any detail about the construction of the ARIMA model. At least, I was unable to find any, even when I looked up earlier works that they referred to with regard to the ARIMA model. To answer your question directly, the best I can state is that they claim that their "historical data" allows them to predict the power consumption after a VM migration has taken place, both for the source and for the destination host.
They do emphasise that they use "static and periodic" workload patterns to accumulate the historical data but I find no reference to whether they use migrations in these workload patterns.
What bothers me most is that the predictions of host power consumption which they show, includes prediction (of the host's power consumption) even during the VM migration. I would dare say that unless migrations are also part of the periodic pattern, then it is not possible to make predictions on the ARIMA basis about host power consumption during a VM migration.