I'm not sure what exactly you mean by condition. Generally speaking, the number of observations required will vary with the number of predictors you include such that using more predictors requires a greater n. It also depends on the field, with some estimates as low as 4 x m where m is the number of predictors, to as high as 100 + m. Refer to this post on StackExchange: http://stats.stackexchange.com/questions/10079/rules-of-thumb-for-minimum-sample-size-for-multiple-regression
As for how much data to use in the model, you usually want to use as many observations as possible barring the crazy outliers, but you can reduce the number of predictors based upon their contribution to the model. The way I do this is to evaluate each candidate model (i.e., models containing different predictors) on the basis of Bayesian Information Criterion or Akaike Information Criterion, which penalize larger numbers of predictors.