Dear Timur and Antoine,
As far as I know, Predictive Analytics can't override the settings on the server or on the database. It is meant to comply with these settings. So if you want to impose some limitations, I guess this is there that you need to do it. (I let Antoine confirm this regarding the very latest versions)
The answer could also be with the size of the data you are analyzing. By definition, training a model requires much resource as it tries to identify patterns in the data by exploring all possible combinations. This is why it is best practice to train models on random samples. You usually get as good a model in terms of statistical performance with a 5% or even 1% random sample. The rule of thumb is to make sure you have at least 1000 positive cases in your analytical data set (positive cases, not simply rows).
Finally, the mdel training is performed by the Predictive Analytics server, while the data preparation is done in-database. So if both servers are installed on the same machine, you could consider separate boxes to lighten the workload on the database server.
Hope this helps,
Serge