AFAIK usually you would use floats and not doubles on the GPU

Thats a lot of triangles and your GPU doesnt seem that powerful, so i would say your usage of VBO's isnt the problem, simply the amount of work youre trying to do.

Assuming your shader isnt full of ifs, what you should do is reduce the amount of triangles you draw. Dont draw objects outside the screen, draw far away stuff with lower detail (LoD)...

Unless you have the GL_ARB_vertex_attrib_64bit extension available you can safely assume that any GLdouble vertex input is going to be software emulated - in immediate mode by casting your glVertex3d parameters down to float, with VBOs by potentially running your entire per-vertex pipeline in software. Obviously that would result in VBOs being somewhat slower than immediate mode code, which is exactly what you've observed.

Double-precision support in hardware is still relatively new (requiring SM5 hardware IIRC), and the moral of the story is that even if the GL spec allows it for a particular call, don't always assume that it means your hardware supports it (especially if it's from an older part of GL that pre-dates hardware T&L).

How can you go faster still? If you're not already doing it, then consider interleaving your vertex struct. I.e. instead of:

GLfloat positions[ARBITRARY_NUMBER];
GLfloat normals[ARBITRARY_NUMBER];
GLfloat texcoords[ARBITRARY_NUMBER];

Use:

struct myVertex
{
    GLfloat position[3];
    GLfloat normal[3];
    GLfloat texcoord[2];
};

myVertex vertices[ARBITRARY_NUMBER];

This can be a win as it means that your GPU can now read all data for each vertex together, and without having to do any random jumping around in memory.

Another thing you can do, since you have an older GPU that only supports GL2.1, is to use GL_UNSIGNED_SHORT instead of GL_UNSIGNED_INT for your GL_ELEMENT_ARRAY_BUFFER, if you can get away with it. This is also coming back to the "don't always assume that hardware supports what the GL spec allows" point, and a GPU that old may not support 32-bit indices very well (or at all!)

Finally, and to cut down on buffer changes, you could consider packing all objects into a single VBO (rather than 150 * 2 separate VBOs) and using the parameters to your glDrawArrays/glDrawElements calls (or the offsets specified by your gl*Pointer calls) to select which object gets drawn. This may or may not be compatible with using 16-bit indices however, so you'll need to profile both approaches and see which works best for you.

you can safely assume that any GLdouble vertex input is going to be software emulated . . . with VBOs by potentially running your entire per-vertex pipeline in software

Perhaps I'm missing something, but it was my understanding that when transferring a static VBO from the client to the GPU, the actual data is copied over--so any doubles ought to have been already converted to floats in this case as well?

In the face of this, I might suspect something else may be afoot; for example, are you sure you're using VBOs, or are you using vertex arrays accidentally instead?

Perhaps I'm missing something, but it was my understanding that when transferring a static VBO from the client to the GPU, the actual data is copied over--so any doubles ought to have been already converted to floats in this case as well?

The data type is only specified at the gl*Pointer stage, by which point the VBO has already long been copied over. glBufferData itself is totally type agnostic; it doesn't care whether you've floats, doubles, plain text or even a JPEG in there - it just transfers a block of memory to the buffer object. gl*Pointer is what defines how that memory is interpreted.

I can see how 100x could be achievable. First of all the program has to be completely vertex-bound, let's say drawing millions of points with each individual point being in it's own glBegin/glEnd pair. Then, it needs to be doing virtually nothing else on the GPU - so all of those points are offscreen. Switch that to using a VBO and a single draw call, and you'll see huge performance increases of that magnitude.

For more realistic real-world programs that people actually write you're not going to see increases like that at all.