I don't know what a parallellizable piece or a parallellizable manifold is, but the key for interpolation is not to compute with the germs but to compute with lots of values. (i.e., don't just localize, but compute in the residue ring). If you do it right, your complicated linear algebra computation now takes place over a field where coefficient swell is orders of magnitude less (and what's more, if it's a field where that's still an issue, you can use the trick again)
As an example, if you have a 2x2 matrix with polynomials in two variables x,y, of degree (n,m). Then you know that the determinant is a polynomial of degree (2n,2m). So, if you now compute the determinant for values of (x,y) in a 2n+1,2m+1 grid, then you can interpolate the determinant back. This can be considerably cheaper than computing the determinant directly over the polynomial ring.