One thing I would note - gblinear can often outperform (performance and speed) tree or dart models at regression tasks. For classification I would start with trees, but linear boosters may be the best choice for many quantitative output models.