Number of samples and instances in the training data

I'm trying to determine the ideal number of samples and instances of data that I should collect. I'm not sure how the number of samples and instances in each sample influences the training data. Is it a large number of samples good? Then, should I try to collect as many instances of each category as possible?

And I will be using the SVM algorithm.

Thanks for your help, and any clarification. And, I'm also not sure if I'm confusing definitions (samples vs instances).