What does primary decomposition of (sub) modules mean geometrically? I want to know how I should visualize modules in algebraic geometry.  The way we visualize rings, via their spectra, automatically (or by the beauty of its design) depicts primary decomposition of ideals: the primary components of an ideal $I \triangleleft A$ cut out "primary subschemes" (irreducible and embedded components) whose union is $Z(I)=Spec(A/I)$.  (See, for example, Eisenbud and Harris, The Geometry of Schemes, II.3.3, pp. 66-70).  This aspect of scheme theory is essential to what makes it "geometric."
By this standard, I think however we visualize modules should allow us to depict primary decomposition of submodules; otherwise I would say it's not a very good visualization.  
If we're happy taking quotients, WLOG we can just look at primary decompositions of $0$.  So let $M$ be a finitely generated module over a Noetherian ring $A$, and  $0=N_1\cap\cdots\cap N_n$ be a primary decomposition of $0$ in $M$, with primes $P_i$ co-associated to the primary modules $N_i$, i.e. associated to the coprimary modules $M/N_i$.

How can one visualize the modules $M,N_1,\ldots,N_n$ in relation to $Spec(A)$ in a way that meaningfully depicts:
  (1) the primary decomposition of $0$ in $M$ (in particular that the $N_i$ are primary in $M$), and
  (2) the relationship of the modules $N_i$ to their co-associated primes, say { $P_i$ } $ = Ass(M/N_i) \subseteq Spec(A)$?

Some useful background results to make sense of the above (all rings and modules are Noetherian):


*

*The primes $P_i$ co-associated to $N_i$ are precisely the associated primes of $M$ (see R. Ash, Comutative Algebra, Theorem 1.3.9)

*A module $Q$ is coprimary iff it has exactly one associated prime $P$, and then $P=\sqrt{ann Q}$.  (see R. Ash, Comutative Algebra, Corollary 1.3.11)
 A: I will assume that everything in sight is Noetherian and finitely generated.  Then the primary decomposition amounts to a description of the geometric support of the module $M$ if you view it as a sheaf on $\text{Spec}(A)$.  The scheme $\text{Spec}(A/\text{Ann}(M))$ is a subscheme of $\text{Spec}(A)$ that, by definition, supports $M$.  If $N_i$ is minimal, then $\text{Spec}(A/P_i)$ is an irreducible component of $\text{Spec}(A/\text{Ann}(M))$.  The module $M/N_i$ is also a quotient of $M \otimes (A/P_i^n)$ for $n$ large enough (and in a reduced decomposition I think they are equal), so it is part of $M$ on that irreducible component of its support.  If $N_i$ is not minimal, then $\text{Spec}(A/P_i)$ is an irreducible scheme inside of a component of $\text{Spec}(A/\text{Ann}(M))$.
A: Visualizing embedded primes:
In P^2,  a one dimensional scheme cannot have embedded points unless its ideal has more than one generator, by the unmixedness theorem of Macaulay.  So imagine we have two polynomials that define a one dimensional scheme in P^2.  We will imagine this scheme as a limit of zero dimensional schemes.  First take two quadratic polynomials, one of which is a product of two linear factors, i.e. take one pair of lines meeting at p, and another irreducible conic.  In general the irreducible conic C meets each of the lines twice, away from p.  Thus the two qudratic polynomials define a zero dimensional scheme of 4 points. 
Now hold fixed the two intersections of C with one of the lines L, and let the two intersections of C with the other line M approach p, i.e. let C become tangent to M at p.  When this occurs, the conic C now contains three distinct points of L, hence C has become reducible and contains L.  Now the scheme defined by intersecting L+M with C has become one dimensional, reducible, and consists set theoretically only of the line L.  I claim the point p is an embedded point of the component L of the scheme defined by L+M and C.
This is easy algebraically, since the ideal of the given scheme is (xy,(x(x-y)) = (x^2, xy) which is the intersection of the primary ideals (x) and (x^2, xy, y^2), with associated primes (x) and (x,y).  Hence (x,y) is an embedded prime.  I.e. the origin is an embedded point on the y axis for this scheme.  This also helps explain the apparent failure of Bezout's theorem for this intersection of two conics apparently not having degree 4.
In general, in P^n, a scheme S with embedded subschemes must be defined by intersecting more hypersurfaces than the codimension of S.  Thus such an S can always be viewed as a limit of lower dimensional schemes.  It seems to me that embedded subschemes should arise when these lower dimensional schemes are reducible and some lower dimensional component comes to lie on a larger dimensional component of the limit.  I do not know if this intuition is the only possibility, and since the world is wide, probably not.
