First, I cannot imagine how could a vertex have a normal,...
Welcome to the exciting field of computer graphics. The first thing that might lead to your confusion is to think about a vertex too much as a point in space in a geometrical way. ---> A 3d point can't have normal since it does not have a face.
But the way you should think about it is more like a checkpoint that contains all information of an arbitrary 3d curve at a certain position.
... and second, when calculate the reflection lights, shouldn't I use the normal of a face?
What happens if you want to render a perfect sphere that can only be represented by a limited set of triangles? It will look like the one in the middle in this question. The flat surfaces are clearly visible, but that is not what you want. You can improve this by drastically increasing the triangle count, but that comes with a huge performance impact and memory footprint.
The solution here is using vertex normals. To understand, why this is a solution, you need to know a little bit about how the render pipeline works. The most important parts for your understanding are vertex processing, rasterization, and the pixel/fragment shader.
Vertex processing is the first programable (in modern APIs) step in the pipeline and is mostly used to reposition your vertices so that your 3d world is displayed correctly on a 2d display. Therefore you apply some mathematical corrections (perspective for example) to your vertex positions. You can do a lot more, but that is not important here. What is important is, that the vertex processor does not know to which triangle a vertex belongs. He sees it in total isolation. Therefore, he can not calculate any face normals, so you have to provide it in some way, for example by adding it as vertex attribute.
After some optional pipeline steps, you get to the rasterizer. Its job is to turn your triangles into actual pixels on your screen. The information he gets are three vertices that form a triangle. Now here is, where the magic happens. For each pixel, he interpolates the properties of the three vertices and passes them to the final step, the pixel shader (also programable in modern APIs). So if your vertices had normals, you will have an interpolated normal for each pixel of your triangle. This enables you to render perfectly smooth surfaces (or at least fake them) since you can perform the lighting calculation for each pixel with individual normals in the pixel shader.
This is why you need vertex normals. The vertices are checkpoints that hold information about your surfaces at specific points. You connect these points to triangles. The information in between is interpolated by the rasterizer. So you essentially describe a whole triangle-shaped field of normals with its three edge values.
However, even if you want to render flat surfaces, you might still need vertex normals. As I said before, the vertex processor does not know anything about triangles, so you can not calculate any surface normals there. So you have to pass them in somehow. There are other ways, but attaching the normals to the vertices is usually the best and easiest solution.
Additional note: You do not necessarily need to perform the lighting calculation in the pixel shader. You can also do this in the vertex stage. So you need vertex normals again ;). This is called Gouraud Shading and gives you a performance boost since you do not need to perform the light calculations per pixel. However, the visual quality of curved surfaces suffers from that. For flat surfaces, you should not see any differences.