Nodes are first
enlarged using the
sep attribute. If
true , overlaps are
retained. If the value is
"scale", overlaps are removed by uniformly scaling
in x and y. If the value converts to
"false", and it is available, Prism, a
proximity graph-based algorithm, is used to remove node overlaps. This can
also be invoked explicitly with
overlap=prism. This technique starts with a
small scaling up, controlled by the
attribute, which can remove a significant portion of the overlap. The prism
option also accepts an optional non-negative integer suffix. This can be used
to control the number of attempts made at overlap removal. By default,
overlap="prism" is equivalent to
overlap="prism0" causes only the scaling phase to be run.
If Prism is not available, or the version of Graphviz is earlier than 2.28,
"overlap=false" uses a Voronoi-based technique. This can always be invoked
overlap="scalexy", x and y are separately scaled to remove overlaps.
overlap="compress", the layout will be scaled down as much as
possible without introducing any overlaps, obviously assuming there are none
to begin with.
**N.B.**The remaining allowed values of
overlap correspond to algorithms which, at present, can produce bad aspect ratios. In addition, we deprecate the use of the
If the value is
"vpsc", overlap removal is done as a quadratic optimization
to minimize node displacement while removing node overlaps.
If the value is
"orthoyx", overlaps are moved by optimizing two
constraint problems, one for the x axis and one for the y. The suffix
indicates which axis is processed first. If the value is “ortho”, the
technique is similar to “orthoxy” except a heuristic is used to reduce the
bias between the two passes. If the value is
"ortho_yx", the technique is
the same as
"ortho", except the roles of x and y are reversed. The values
"portho_yx" are similar to the
previous four, except only pseudo-orthogonal ordering is enforced.
If the layout is done by neato with
mode=“ipsep”, then one can use
overlap=ipsep. In this case, the overlap removal constraints are
incorporated into the layout algorithm itself. N.B. At present, this only
supports one level of clustering.
sfdp, the layouts assume
overlap="true" as the default.
Fdp first uses a number of passes using a built-in, force-directed technique
to try to remove overlaps. Thus,
overlap with an integer
prefix followed by a colon, specifying the number of tries. If there is no
prefix, no initial tries will be performed. If there is nothing following a
colon, none of the above methods will be attempted. By default,
overlap="9:prism". Note that
overlap="0:" all turn off all overlap removal.
Except for the Voronoi and prism methods, all of these transforms preserve
the orthogonal ordering of the original layout. That is, if the x coordinates
of two nodes are originally the same, they will remain the same, and if the x
coordinate of one node is originally less than the x coordinate of another,
this relation will still hold in the transformed layout. The similar
properties hold for the y coordinates. This is not quite true for the
"porth*" cases. For these, orthogonal ordering is only preserved among nodes
related by an edge.