This is the multi-page printable view of this section. Click here to print.

Return to the regular view of this page.

FAQ

If you cannot find the answer to your question here, ask it in the Graphviz forum.


General

Where can I see a list of all the attributes that control dot or neato?

See Graph Attributes. There is also information on command-line usage and output formats.

Where can I discuss Graphviz?

Post questions and comments in the Graphviz forum

I'm trying to make a layout larger. How?

There are various ways to increase the size of a layout. In doing this, one has to decide if the sizes of the nodes and text should be increased as well.

One approach is to adjust individual parameters such as fontsize, nodesep and ranksep. For example,

digraph G {
  graph [fontsize=24]
  edge [fontsize=24]
  node [fontsize=24]
  ranksep = 1.5
  nodesep = .25
  edge [style="setlinewidth(3)"]
  a -> b -> c
}

If you do this, make sure you are not fighting a conflicting graph size setting, like size="6,6", which will then scale everything back down.

If you are using fdp or neato, increasing the edge len will tend to expand the layout.

graph G {
  layout="neato"
  edge [len=3]
  a -- { b c d }
}

For twopi and circo, there are other parameters such as ranksep which can be used. See the graph attributes.

You can also use the ratio attribute. If you set the size attribute to the desired drawing size, and then set ratio=fill, node positions are scaled separately in x and y until the drawing fills the specified size. Note that node sizes stay the same. If, instead, you set ratio=expand, the layout is uniformly scaled up in x and y until at least one dimension fits size.

If you specify the size attribute but end it with an exclamation mark (!), the final drawing will be scaled up uniformly in x and y until at least one dimension fits size. Note that everything is scaled up, including text and node sizes.

If you're using PostScript, you can just scale up the output by manually adding a command such as 2 2 scale where the PostScript environment is set up. Make sure to adjust the BoundingBox too if your tools look at this header.

How can I join or merge certain edge routes in dot?

You can try running dot -Gconcentrate=true or you can introduce your own virtual nodes drawn as tiny circles where you want to split or join edges:

digraph G {
  yourvirtualnode [shape=circle,width=.01,height=.01,label=""]
  a -> yourvirtualnode [arrowhead=none]
  yourvirtualnode -> {b;c}
}

How can I generate graph layouts in PDF?

If your version of Graphviz has cairo/pango support, you can just use the -Tpdf flag. Unfortunately, this does not handle embedded links.

If you need embedded links, or don't have cairo/pango, create PostScript output, then use an external converter from PostScript to PDF. For example, dot -Tps | epsf2pdf -o file.pdf. Note that URL tags are respected, to allow clickable PDF objects.

If your intention is to use the figure as PDF in some document preparation system, such as pdflatex, it is very important to use -Tps2 rather than -Tps. In general, if you really want PDF output, that is, you would like to have a -Tpdf flag, use -Tps2 before converting to PDF.

In the diagram below, the shaded nodes will contain bad output.

alt text

How can I make duplicate nodes?

Make unique nodes with duplicate labels.

digraph G {
  node001 [label = "A"]
  node002 [label = "A"]
  node001 -> node002
}

How can I set a graph or cluster label without its propagating to all sub-clusters?

Set the label at the end of the graph (before the closing brace), after all its contents have been defined. (We admit it seems desirable to define some special syntax for non-inherited attribute settings.)

How can I draw multiple parallel edges in neato?

When the splines attribute is false, which is the default, multiedges are drawn as a spindle of simply curved edges. There is no attempt to avoid intervening nodes.

When splines=true or polyline, multiedges are drawn as roughly parallel splines or polylines. This relies on there being no node overlaps.

An additional trick which is sometimes sufficient is to specify multiple colors for the edge using a color list. This will a produce set of tightly parallel splines, each in its specified color. Read about the color attribute for more information.

How can I symmetrize (balance) tree layouts?

When a tree node has an even number of children, it isn't necessarily centered above the two middle ones. If you know the order of the children, a simple hack is to introduce new, invisible middle nodes to re-balance the layout. The connecting edges should also be invisible. For example:

digraph G {
  a -> b0
  xb [label="",width=.1,style=invis]
  a -> xb [style=invis]
  a -> b1
  {rank=same b0 -> xb -> b1 [style=invis]}
  b0 -> c0
  xc [label="",width=.1,style=invis]
  b0 -> xc [style=invis]
  b0 -> c1
  {rank=same c0 -> xc -> c1 [style=invis]}
}

This trick really ought to be built into our solver (and made independent of the order of the children, and available for layouts other than trees, too).

How can I report a bug or issue I've found?

You can report or view Graphviz bugs and issues by visiting the Graphviz Issues page.

Clusters

How can I create edges between cluster boxes?

This only works in Graphviz version 1.7 and higher. To make edges between clusters, first set the graph attribute compound=true. Then, you can specify a cluster by name as a logical head or tail to an edge. This will cause the edge joining the two nodes to be clipped to the exterior of the box around the given cluster.

For example,

digraph G {
  compound=true; nodesep=1.0;
  subgraph cluster_A {
    a -> b; a -> c;
  }
  subgraph cluster_B {
    d -> e; f -> e;
  }
  a -> e [ ltail=cluster_A, lhead=cluster_B ];
}

has an edge going from cluster_A to cluster_B. If, instead, you say

a -> e [ltail=cluster_A];

this gives you an edge from cluster_A to node e. Or you could just specify an lhead attribute. The program warns if a cluster specified as a logical node is not defined. Also, if a cluster is specified as a logical head for an edge, the real head must be contained in the cluster, and the real tail must not be. A similar check is done for logical tails. In these cases, the edge is drawn between the real nodes as usual.

Clusters are hard to see.

Set bgcolor=grey (or some other color) in the cluster.

Output

How can I get high quality (antialiased) output?

The easiest thing is to use a vector-based output formats such as PDF, SVG or PostScript. In addition, if Graphviz has a cairo/pango backend, this will generate anti-aliased output.

Another approach is to make the layout in PostScript (option -Tps), then run through Ghostview with antialiasing enabled. The important command line options are:

  • -dTextAlphaBits=4
  • -dGraphicsAlphaBits=4 (4 is the highest level of antialiasing allowed - see the Ghostview documentation).

The full command line to render a raster could be something like:

$ gs -q -dNOPAUSE -dBATCH -dTextAlphaBits=4 -dGraphicsAlphaBits=4 -sDEVICE=png16m -sOutputFile=file.png file.ps

On Mac OS X, the pixelglow port uses Apple's Quartz renderer, which enables antialiasing. It also provides a beautiful document container for its user interface. (One downside is that you can't run Pixelglow Graphviz as a web server or other background process if your Mac has 3D graphics, because Quartz wants to get this resource to accelerate rendering.)

I can only get 11x17 output?

It's not us! It's probably your printer setup. If you don't believe this, run dot -Tps and look at the BoundingBox header. The coords are in 1/72ths of an inch.

How do I create special symbols and accents in labels?

Insert the symbol directly (such as with copy/paste) into your dot source, and save as UTF-8, e.g.:

graph G {
  yen [label="¥"]
}

If you can't save in UTF-8, try HTML entities, e.g. ¥ for the Yen currency symbol ¥. Example:

graph G {
  yen [label="¥"]
}

More generally, how do I use non-ASCII character sets?

The following applies to Graphviz 2.8 and later. (In older versions of Graphviz, you can sometimes get away with simply putting Latin-1 or other UTF-8 characters in the input stream, but the results are not always correct.)

Input: the general idea is to find the Unicode value for the glyph you want, and enter it within a text string "..." or HTML-like label <...>.

For example, the mathematical forall sign has the value 0x2200. There are several ways this can be inserted into a file. One is to write out the ASCII representation: &#<nnn>; where <nnn> is the decimal representation of the value. The decimal value of 0x2200 is 8704, so the character can be specified as &#8704; . Alternatively, Graphviz accepts UTF-8 encoded input. In the case of forall, its UTF-8 representation is 3 bytes whose decimal values are 226 136 128. For convenience, you would probably enter this using your favorite editor, tuned to your character set of choice. You can then use the iconv program to map the graph from your character set to UTF-8 or Latin-1.

We also accept the HTML symbolic names for Latin-1 characters as suggested in FaqSymbols. For example, the cent sign (unicode and Latin-1 value decimal 162) can be inserted as &cent;

Note that the graph file must always be a plain text document not a Word or other rich format file. Any characters not enclosed in "..." or <...> must be ordinary ASCII characters. In particular, all of the DOT keywords such as digraph or subgraph must be ASCII.

Because we cannot always guess the encoding, you should set the graph attribute charset to UTF-8, Latin1 (alias ISO-8859-1 or ISO-IR-100) or Big-5 for Traditional Chinese. This can be done in the graph file or on the command line. For example charset=Latin1.

Output: It is essential that a font which has the glyphs for your specified characters is available at final rendering time. The choice of this font depends on the target code generator. For the gd-based raster generators (PNG, GIF, etc.) you need a TrueType or Type-1 font file on the machine running the Graphviz program. If Graphviz is built with the fontconfig library, it will be used to find the specified font. Otherwise, Graphviz will look in various default directories for the font. The directories to be searched include those specified by the fontpath attribute, related environment or shell variables (see the fontpath entry), and known system font directories. The table points out that these glyphs are from the times.ttf font. With fontconfig, it's hard to specify this font. Times usually gets resolved to Adobe Type1 times, which doesn't have all the glyphs seen on that page.)

For PostScript, the input must be either the ASCII subset of UTF-8 or Latin-1. (We have looked for more general solutions, but it appears that UTF-8 and Unicode are handled differently for every kind of font type in PostScript, and we don't have time to hack this case-by-case.

For SVG output, we just pass the raw UTF-8 (or other encoding) straight through to the generated code.

How can I create custom shapes?

One approach is to use HTML-like labels possibly combined with embedded images use the IMG attribute.

As alluded to in the dot user's guide, if you want truly custom shape, there are several ways to incorporate this. At this point, they either must be in PostScript or image files, or you'll need to modify the source code. A serious problem is that you can't make custom shapes that work across all the drivers and the interactive front-ends such as Grappa. At least SVG has interactive renderers, and PostScript can be translated to PDF which also has some interactive features.

External image files

If using SVG (-Tsvg), PostScript (-Tps,-Tps2) or one of the raster formats (-Tgif, -Tpng, or -Tjpg), you can load certain images (e.g., pictures) by file name into nodes. For example:

   yournode [image="yourface.gif"];

indicates the contents of the node are given in the GIF file yourface.gif. The image attribute specifies which file to use. (There is also the deprecated shapefile attribute. This is similar to image but the node shape will always be a box.)

NOTE: In versions before 11 March 2006, in particular, 1.12 graphviz and earlier, it is necessary to also set the attribute shape=custom.

With -Tsvg, image must give the name of a file containing a GIF, PNG or JPEG bitmap file. Note that the file's contents are not copied into the SVG output, only the files name. Thus, for Graphviz SVG output to display correctly, the image file must be available to the SVG viewer.

With PostScript, image must give the name of a file containing encapsulated PostScript or a bitmap. The contents are copied into the output file. Note that encapsulated PostScript will only be copied in once. The restrictions on the contents of the image are the same as those specified below under External PostScript files.

For bitmap output, image is a file name containing a bitmap image. The file is opened and copied (and possibly scaled) into the output drawing.

This code is still preliminary, and we have noted some problems with color quantization in indexed color map management that we are trying to understand and correct. (You can use -Gtruecolor=1 to try a 32 bit internal canvas as an alternative, but we have observed fuzziness (lossiness?) in the images.)

When the software is used as a web server, access to image files is more restrictive. See SERVER_NAME.

External PostScript files

If using the PostScript driver (-Tps) you can import node shapes as external PostScript files such as EPS (Encapsulated PostScript). At a minimum, the external file must have a valid BoundingBox header and not do drastic things to the graphics state since we don't install a wrapper for example to inhibit showpage.

To import an external PostScript file, set the shape and shapefile attributes as shown here:

	somenode  [shape=epsf, shapefile="yourfile.ps" ];

An EPSF shape is always clipped to its bounding box.

The use of [shape=epsf, shapefile="yourfile.ps" ] is largely superseded by the mechanism described in the previous section, using [image="yourfile.ps" ].

External PostScript procedures

If using the PostScript driver (dot -Tps), you can define a PostScript procedure for shape drawing. The procedure must be able to draw variable-sized shapes. A file containing the definition can be loaded as a command line argument using the -l flag:

	$ dot  -Tps -l yourPS.ps  file.dot -o file.ps

In the graph file, invoke the shape like this:

	somenode [shape=yourshape]

In file.ps, for non-filled nodes, the procedure for yourshape will be called like this:

[ 54 36 0 36 0 0 54 0 54 36 ]  4 false yourshape

where the current color is the node's pencolor. The array contains the shape's bounding polygon, with the first point repeated at the end, followed by the number of points. At present, the shape is always a rectangle. From left to right, the points in the array always go counterclockwise, starting with the upper-right vertex. The boolean value after the number of vertices, here false, is the value of the node's fill attribute. The coordinates are absolute canvas coordinates.

For nodes with fill=true, the above invocation of yourshape will be preceded by

[ 54 36 0 36 0 0 54 0 54 36 ]  4 true yourshape

where the current color is the node's fillcolor.

NOTE: In versions before 23 September 2005, yourshape is only invoked once, with the node's fill value and with the color set to the node's pencolor.

For example, here are the contents of a plausible shape file, DFD.ps, which can be invoked by [shape=DFDbox]

	/xdef {exch def} bind def
	/DFDbox {
		10 dict begin
			/fflag xdef
			/sides xdef
			fflag   % if shape is filled
			{
				aload pop
				newpath
				moveto
				1 1 sides { pop lineto } for
				closepath fill
			}
			{
				aload pop
				% draw the sides
				newpath
				moveto
				1 1 sides {
					2 mod 0 ne
					{moveto} % even sides
					{lineto currentpoint stroke moveto} % odd sides
					ifelse
				} for
            }
			ifelse
		end
	} bind def

A custom shape of this kind is always clipped to its bounding box. It would not be hard to create a hook in the function user_shape() in shapes.c to determine clipping polygons other than rectangles (maybe) in case someone wants to try this and contribute the code for this.

Note that, by default, the bounding box is drawn around the contents and the node label is drawn. If you wish to eliminate these, set label="" and peripheries=0 in the node.

Driver-independent custom shapes

If not using PostScript, you'll need to roll up your sleeves and modify the source. None of the other code generators support custom node shapes directly. If the custom shape is to be high-level and driver-independent, then you can add shape-specific functions (methods) to shapes.c with a corresponding entry in the array Shapes[] that maps shape names to methods. The method interfaces are described in the comment header in this file. Methods must be defined to initialize a shape (typically to size it big enough to fit its text label), bind a port name to a coordinate, test if a point is inside an instance of the shape (for edge clipping), generate code for the shape via the functions provided in the gvrender_engine_t structure, and return a box path to reach ports on the interior of a node (if they can exist).

More information on the functions available via gvrender_engine_t and the Graphviz graphics model can be found in Section 5 of the Graphviz library manual.

Shapes that behave more or less like polygons can be bootstrapped from the basic polygon methods; see for example the invtri or tab shape. Such shapes use a polygon descriptor whose fields are listed below.

Field name Description Default
regular if a regular polygon FALSE
peripheries number of border peripheries 1
sides number of sides (1 for curves) 4
orientation angular rotation in degrees 0
distortion trapezoidal distortion 0
skew parallelogram distortion 0
option fancy options: ROUNDED, DIAGONALS, AUXLABELS 0

For shapes not derived from general polygons, see the record or epsf shape.

Driver-dependent custom shapes

To implement a driver-specific shape (such as GIF or PNG icons) you'll need to write a body for the driver function that implements user-defined shapes. This involves providing the library_shape function for the specific driver, if it does not already exist, and generating the graphics operations needed to display your shape using the driver's graphics functions. (The drivers supplied with Graphviz can be found in the plugins directory.)

A user shape function basically receives four arguments:

  • the custom shape name string
  • the absolute canvas coords of the shape bounding polygon
  • number of coords (currently, always 4)
  • fill flag

The rest is up to you, but contact us first just in case.

How can I use drawing layers (overlays)?

If the layers graph attribute is set, the graph is printed in a sequence of colored layers, or overlays. (This coloring overrides any other settings.) layers defines a list of layer names, each separated by a sequence of separator characters. The tokens can be any identifier or natural number, except the reserved word all. By default, the separator characters are colon, space and tab, though this can be overridden using the layersep graph attribute.

The layer attribute of a node, edge or cluster enables its appearance in given layers. Its value denotes a list of layers from the layers graph attribute. It is specified by a list of layer intervals, separated by a sequence of characters from the layerlistsep attribute. Each layer interval is written as a single layer name or two layer names separated by a sequence of separator characters from the layersep attribute. The keyword all means all possible layers. If all is used as part of a range, the range denotes all layers bounded on one side by the other token. Thus, the layer pvt:all of the edge node2 -> node3 in the example below corresponds to the layers pvt, test, new, and ofc. The all:pvt,new,ofc corresponds to the layers local, pvt, new, and ofc.

As an example, the graph:

digraph G {
	layers="local:pvt:test:new:ofc";

	node1  [layer="pvt"];
	node2  [layer="all"];
	node3  [layer="pvt:ofc"];		/* pvt, test, new, and ofc */
	node2 -> node3  [layer="pvt:all"];	/* same as pvt:ofc */
	node2 -> node4 [layer=3];		/* same as test */
}

produces the 5 layers shown below:

Layer 1 Layer 2 Layer 3
Layer 4 Layer 5

In a layered graph, if a given node (or edge) does not have a layer assignment, but incident edges (nodes, resp.) do, its layer specification is inferred from these. For example, in the examples above, node4 only appears on layer 3, because a layer assignment was given for its connecting edge. Note, however, that if a node or edge without a layer attribute is incident with an edge or node without a layer attribute (or such a node has no edges), then the node or edge appears on all layers.

To change the default so that nodes and edges with no layer attribute appear on all layers, insert

	node [layer=all];
	edge [layer=all];

at the beginning of the graph file.

The graph can have a layerselect attribute, which specifies which layers should be emitted. The value uses the same concrete syntax as the layer attribute.

At present, the output of multiple layers into a single output file is only available in PostScript. However, the layerselect attribute can be used to select a single layer for output in any format.

The color sequence for layers is set in the array layercolorseq (in PostScript at least). The first index is 1 and each element is a three-element color coordinate array. Custom layer colors can be created by setting the value of this array.

TODO

  • simply change default colors per layer, thus allowing user override on individual nodes or edges when desired.
  • turn off layer coloring entirely and just use the colors inherent in the drawing.
  • forcing nodes/edges in a given subgraph to take certain attributes. Might require a hook in the libgraph parser. It would be pretty easy to support the semantics: for every node/edge in this subgraph, assign it the default attributes that differ from the defaults of the parent of this graph. What you need to avoid is the problem exposed in the following example:
  subgraph sub0 {
    node [color=red];
    a; b; c;
  }
  subgraph sub1 {
    node [shape=diamond];
    a; b; c;
  }

We don't want to reset a,b,c to color=black just because that is default in sub1.

How do I get font and color changes in record labels or other labels?

This is not possible in record shapes. However, you can do this using HTML-like labels.

In -Tplain format, splines do not touch the nodes (arrowheads are missing).

Edges are specified as the main spline and, if necessary, arrowheads which actually abut the node. If the arrowheads are not given, drawing the edge spline will leave a gap between the edge and the node. This is a bug which has now solidified into a feature. A workaround is to set edge [dir=none]. Since the edges have no arrowheads, the spline specification will go all the way to both nodes.

Record nodes are drawn differently in dot and neato when rankdir=LR.

It's true. dot -Grankdir=LR rotates record nodes so that their top level fields are still listed across levels. rankdir=LR has no effect in neato. One workaround is to use HTML-like labels (they don't rotate; the downside is that you have to write in XML). In general, we recommend replacing record nodes with the much more general HTML-like labels. Another workaround is to enclose record labels in { } to rotate/unrotate the record contents. See also, How To Avoid Foolish Consistency by Scott Berkun (Microsoft Corp.)

How can I print a big graph on multiple pages?

The page attribute, if set, tells Graphviz to print the graph as an array of pages of the given size. Thus, the graph

digraph G {
  page="8.5,11";
  ...
}

will be emitted as 8.5 by 11 inch pages. When printed, the pages can be tiled to make a drawing of the entire graph. At present, the feature only works with PostScript output.

Alternatively, there are various tools and viewers which will take a large picture and allow you to extract page-size pieces, which can then be printed. See also the viewport attribute.

When I have a red edge it shows up as a solid red in PNG and GIF formats, but has a black border when rendered to JPEG.

This is an artifact of JPEG's lossy compression algorithm. JPEG isn't very good for line drawings. Consider using PNG.

Sometimes in dotty, right mouse click shows the global menu but none of the items can be selected.

Check that the NUMLOCK key is off. It's a known bug.

Why does dotty report a syntax error on a legal dot file?

The dotty tool is deprecated, but this entry is retained to aid anyone still using it.

Typically this error is reported as:

>> graph parser: syntax error near line 14
>> context: >>> <<< digraph G {
>> dotty.lefty: giving up on dot
>> dotty.lefty: graph that causes dot
>> dotty.lefty: to fail has been saved in file dottybug.dot

Probably there is a command in your shell environment (such as .alias or .profile) that does output even for non-interactive shells. When this occurs, those characters go in the pipe to the dot parser and cause this problem. An easy check is whether other users have the same problem.

How can I get rid of the little circles on edges ("edge handles") in dotty?

The dotty tool is deprecated, but this entry is retained to aid anyone still using it.

Edit the file dotty.lefty and change the line that says: 'edgehandles' = 1; to 'edgehandles' = 0; it's around line 110.

I already have all the coordinates for the nodes and edges of my graph and just want to use dot, neato, or dotty to render it. How?

Put the graph with layout attributes into a dot file. Then run neato -n2. For example:

$ neato -n2 -Tgif file.dot -o file.gif

Note that if an edge does not have a pos attribute defined, neato will perform whatever edge routing it would normally do. All of the usual backend attributes (size, overlap, page, etc.) are available.

I already have all the coordinates for the nodes, and I want dot or neato to route the edges.

Run neato -n. This will add the necessary edge information.

I already have all the coordinates for the nodes and edges of my graph and just want to use dotty to render it. How?

If you want to do a layout first, use -Txdot for the output format. Dotty uses the layout information provided there.

Same as above, but I have only node coords, not edges.

Run neato -Txdot -n. This will add the necessary edge information.

How can I make client-side image maps?

Use the -Tcmapx command line option. See here for more details.

Why aren't my server-side maps being recognized? I've checked the HTML!

Make sure that your server has map files enabled. For example, if running apache, check that httpd.conf has a line like the following:

AddHandler imap-file map

and that it is not commented out!

I've installed Debian Graphviz and it works fine on the command line, but when I execute a Perl/CGI script through Apache, no output is generated.

For example, the code system("/usr/local/bin/dot -Tpng /tmp/tree.dot -o /tmp/tree.png"); produces no file /tmp/tree.png.

As best as we can tell, dot dies with no stdout or stderr messages on Debian systems when run from an Apache cgi program with no HOME set. The workaround is to provide a HOME directory in the Apache userid's environment.

Someone has also suggested using the Perl module for Graphviz.

Greg Brauer from Dreamworks pointed out a different possibility: The problem turned out to be that I was not closing the file descriptor on my temporary dot file before running dot. Graphviz would end up getting a newly-created empty file with nothing in it before the write buffer to the file was flushed. Dot would happily run on that and generate an empty output file without any warnings.

How can I get 3D output?

The Graphviz authors have qualms about the gratuitous use of 3D.

Nonetheless, dot -Tvrml generates VRML files. There's no Z coordinate layout - you specify Z coords yourself in the z attribute of nodes, and the Z coordinates of edges are interpolated. If someone contributes a driver for a newer, more useful format (OpenGL Performer scene graphs? Open Scene Graphs? Java3D programs?) we'd like to try it.

neato internally supports layouts in higher dimensions through the dim and dimen attributes, e.g. neato -Gdim=7. Graphviz output handles 2D and 3D, but there's no way to get higher-dimensional output unless you invoke neato as a library and inspect ND_pos(n)[i] where n is a pointer to the relevant node.

Problems

How can I avoid node overlaps in neato?

Use the graph attribute overlap.

How can I avoid node-edge overlaps in neato?

Use the overlap attribute to leave room among the nodes, then use -Gsplines=true.

$ neato -Goverlap=... -Gsplines=true -Gsep=.1

The sep argument is the node-edge separation as a ratio of a node's bounding box. That is, sep=.1 means each node is treated as though it is 1.1 times larger than it is. The actual value may require some tinkering. (Don't ask why this isn't just a constant!) Note that this option really slows down neato, so should be used sparingly and only with modest-sized graphs.

cause a crash?

This usually happens when the Graphviz library is built using one version of the stdio library, and the user's program is compiled using another. If the FILE structure of stdio is different, the call to agread() will cause a crash. This is mainly a problem on Windows where we just provide a binary release built with one version of Visual Studio and stdio changes depending on the version of Visual Studio. It can also occur if the user tries to use cygwin or something similar which may also use an incompatible stdio.

The simplest solution is to read in the entire graph into memory, and pass a pointer to that memory to agmemread().

If that is not possible (e.g., the file is too large), you need to tell agread() to use a reader compatible with the stream you give it. By default, agread assumes you are passing it a FILE* generated by the version of stdio that was used to compile Graphviz, and uses that version of fgets to read the stream. To pass in your own reader, use the third argument:

Agdisc_t mydisc;
Agodisc_t myiodisc;

mydisc.mem = NULL;  // use system default
mydisc.id = NULL;   // use system default
mydisc.io = &myiodisc;
myiodisc.afread = reader; 
myiodisc.putstr = NULL;  // only need to set if calling gvRender()
myiodisc.flush = NULL;   // only need to set if calling gvRender()

The reader function has type

int (*reader) (void *chan, char *buf, int bufsize);

and should act like the read() system call. That is, it reads from the stream chan, storing the bytes in buf, whose size is bufsize, and returns the number of bytes it read.

For Unix-like stdio, one might use

static int reader(void *chan, char *buf, int bufsize)
{
    return fread(buf, 1, bufsize, (FILE*)chan); 
}

Then, to read the graph, use:

FILE* fp = fopen ("mygraph.gv","r");
Agraph_t* g = agread (fp, &mydisc);

Neato runs forever on a certain example.

First, how big is your graph? Neato is a quadratic algorithm, roughly equivalent to statistical multidimensional scaling. If you feed it a graph with thousands of nodes and edges, it can easily take hours or days. The first thing to check is to run neato -v to get a trace of the output. If the numbers you see are generally getting smaller, the layout is just taking a long time. You can set certain parameters, such as epsilon or maxiter to shorten the layout time, at the expense of layout quality. But if your graph is big, who's going to notice?

Alternatively, use sfdp to layout the graph (though at present, sfdp does not honor edge lengths).

If you use mode=KK with neato or an early version, it is possible the optimization may cycle. If you see the numbers repeating, or fluctuating up and down, then neato is cycling, especially if your graph is small. This should never happen by default for versions later than 1.13. If it does, please report it as a bug.

If you are using an earlier version of neato, or you used mode=KK, cycling is indeed possible. This cycling is very sensitive to the initial layout. By using the start attribute, for example,

$ neato -Gstart=3
$ neato -Gstart=rand

the cycling will most likely disappear. Or you can employ the parameters used for large graphs to stop the layout earlier:

$ neato -Gepsilon=.01
$ neato -Gmaxiter=500

Note that, if you have a large graph, the generation of edges as splines is a cubic algorithm, so you would do well to avoid using splines=true. (This comment applies to circo, fdp and twopi as well.)

Edge label placement in dot is bad, or the layout is very convoluted.

By default, edge labels in dot are modeled as dummy nodes. This guarantees adequate space for the labels, but for complicated graphs, this can dramatically distort the layout. In this case, it might be worthwhile to replace edge labels with xlabels. In this case, the graph is laid out as if there are no edge labels, with the labels added after the edges have been routed. This can prevent a contorted drawing at the expense of possibly have edge labels overlapping.

Dot runs forever on a certain example.

Try dot -v to observe its progress.

Note that it's possible to make graphs whose layout or even parsing is quadratic in the input size. For example, in dot,

digraph G {
  a -> b -> c -> .... -> x -> y -> z
  a -> z
  b -> z
  c -> z
  /* and so on... */
  x -> z
}

The total edge length (therefore the layout time) of this as a ranked graph is quadratic in the number of nodes. You probably won't encounter the following, but it is also possible to construct graphs whose parsing takes quadratic time in the number of attributes, by appending attributes to nodes and edges after the graph has been loaded. For example:

digraph G {
  /* really big graph goes here...with N+1 nodes */
  n0 -> n1 -> ... -> nN;

  n0 [attr0="whatever",
  attr1="something else",
  /* and so on with many more attributes */
  attrM="something again"]
}

When an attribute first appears, each object is visited with possible cost proportional to the number of previously declared attributes. Thus, the running time for the above would be cN*O(M) for some constant c. If there is any concern about this, the graph should specify the attributes first before declaring nodes or edges. In practice, this problem is negligible.

Twopi runs forever on a certain example.

If your graph is large (many thousands of edges), and you set splines=true, it takes a lot of cycles to fit all those splines!

Neato has unnecessary edge crossings, or has missed an obvious chance to make a much nicer layout.

Neato and all similar virtual physical model algorithms rely on heuristic solutions of optimization problems. The better the solution, the longer it takes to find. Unfortunately, it is also possible for these heuristics to get stuck in local minima. Also, it is heavily influenced by the initial position of the nodes. It is quite possible that if you run neato again, but with a different random seed value, or more iterations, you'll get a better layout. For example:

$ neato -Gstart=5 file.dot -Tps -o file.ps
$ neato -Gepsilon=.0000001 file.dot -Tps -o file.ps

In the default stress majorization mode of neato, using -Gstart=self can help produce a better initial layout.

Note that there are no guarantees that neato will produce a planar layout of a planar graph, or expose all or most of a graph's symmetries.

Webdot doesn't work.

We assume you're using Apache and have TCL installed. If you don't, it's probably better to just use the webdot perl script.

To debug webdot, first test whether tclsh can load the Tcldot shared library. Try:

$ tclsh
% load $prefix/lib/graphviz/tcl/libtcldot.so.0
%

where $prefix is the installation prefix for graphviz; usually /usr or /usr/local.

Then test whether webdot runs from a shell command. (With webdot we provide a helper script scaffold.tcl or scaffold.sh that sets up an environment like the one Apache provides.) For example

$ scaffold.tcl >out.gif
can't read "LIBTCLDOT": no such variable

while executing

"file mtime $LIBTCLDOT"

invoked from within

"set t1 [file mtime $LIBTCLDOT]"

(file "cgi-bin/webdot" line 67)
invoked from within

"source cgi-bin/webdot
"

(file "scaffold.tcl" line 22)

The above is a strong clue that webdot is not configured properly.

Finally, test whether webdot runs as a cgi-bin program. It may help to examine the cgi-bin environment using a simple cgi-bin tcl script like:

#!/bin/env tclsh
puts "Content-type: text/plain"
puts ""
foreach e [lsort [array names env]] {puts "$e: $env($e)"}

Save this script as .../cgi-bin/test.tcl, make it executable, then look at: http://localhost/cgi-bin/test.tcl

Also, if you see something like:

WebDot Error: 
Response Code = 403

This usually means that webdot ran successfully, but was not able to fetch the remote graph from the URL you gave as an argument. The reason is probably that your server is behind a firewall that blocks the webdot server, so it cannot get the graph file. You can either change firewall permissions, put the graph on a different server, or install webdot locally so you don't need a remote server to fetch your graph data.

It would be nice if someone hacked webdot to take the contents of a graph as a cgi-bin argument, so it wouldn't need permission to fetch a graph remotely. This is left as an exercise for the Open Source Community.

I have "Font not found" errors, or text labels missing in webdot.

Firstly, recent versions of graphviz will use fontconfig if it is available on your platform. With fontconfig, this error should not occur, so you may want to see if an upgrade to graphviz is available, or if a rebuild will add fontconfig support.

If fontconfig is not available then graphviz tries to resolve fontnames to fontpaths itself, and uses DOTFONTPATH (or GDFONTPATH) to indicate where it should look.

For copyright reasons, Graphviz doesn't come with its own fonts. On a Windows machine, it knows to search in C:\Windows\Fonts. On a Unix machine, you need to set up a directory that contains Truetype fonts. You can get a copy of some fonts here.

The default DOTFONTPATH is:

#define DEFAULT_FONTPATH "/usr/X11R6/lib/X11/fonts/TrueType:/usr/X11R6/lib/X11/fonts/truetype:/usr/X11R6/lib/X11/fonts/TTF:/usr/share/fonts/TrueType:/usr/share/fonts/truetype:/usr/openwin/lib/X11/fonts/TrueType:/usr/X11R6/lib/X11/fonts/Type1"

If your fonts are somewhere else, then you must set that directory in the webdot script, or recompile Graphviz with the correct DEFAULT_FONTPATH (or set fontpath="/your/font/directory" in every graph you lay out, but that's pretty clumsy.)

You can also try to comment the line #set SIGNATURE "Graph by WebDot" in /var/www/cgi-bin/webdot.

What is the coordinate transformation between the graph bb and a .png image?

  • The bb is expanded by 4 graph-units in all directions (pad) to allow for finite line widths.
  • Then it is zoomed and/or rotated according to -Gviewport, -Gsize, -Glandscape, -Gorientation options. At the default scaling of 1:1, one graph unit = 1 point (1/72 inch).
  • Then it is paginated, if requested by -Gpage and if the output format supports it. Not the -Tpng renderer, yet.
  • Then a margin is added, -Gmargin, in absolute units (inches). The top/bottom margin can be set independently of the left/right margin.
  • Then it is converted to device units, according to -Gdpi, or a dpi value that is given by the output device, or a default that is provided by each render. There are separate dpi values for x and y to allow for non-square pixels. Some renderers invert the Y axis and need an offset to place the origin in the top left corner. The default dpi for -Tpng is 96dpi (approximating the resolution of most computer monitors) so this is where the scaling by 96/72 (4/3) comes from.

At the renderer api, plugins have a choice of coordinate representation:

  • coordinates in graph-units, and composite transformation data consisting of: scaling, rotation, and translation. (used by svg, cairo, ps, renderers)
  • coordinates pre-transformed into device units.

1 - Font FAQ

Before we launch into the gory details, we would like to explain why fonts are a hard problem. The naming and rendering of text fonts in Graphviz (and other programs) is complicated. There are several reasons:

  • Graphviz runs on a wide range of systems: Linux and other Unix variants, Microsoft Windows, and Mac.
  • Graphviz has a wide range of output formats: raster-oriented formats like PNG and GIF; path-based ones like Postscript, PDF and SVG; some idiosyncractic legacy formats, like troff PIC.
  • Often, output will be downloaded and displayed on a computer or other device, different than the one where the layout was created.
  • Graphviz layouts should be identical in size and appearance, regardless of the output format.
  • Graphviz can run on external libraries that help with naming and rendering text fonts, but they are not required, and stripped-down Graphviz tools can be built without them. In fact, Graphviz may have to run on systems with no font files installed.
  • There are several major font file formats to be supported.
  • Non-Western, international character sets should be supported.
  • Graphviz should provide a good set of standard fonts.
  • It should be easy to specify standard fonts.
  • Users should be able to load their own custom fonts.
  • Output should be small to download quickly.
  • Output should allow the best rendering possible in a given format.
  • Output files should be easy to postprocess, for example, retaining the objects of the original graph if possible.
  • It is very helpful to work around known bugs or missing features in support libraries and popular external tools.

This is a tall order. Some of the goals conflict. Generally our approach has been to define defaults that favor convenience and good looking output, and give the user options to override the defaults.

Overview

In the following, we will assume a "standard" version of Graphviz with the full set of support libraries (fontconfig, gd, Cairo and Pango), running on a desktop system or server with a standard installation of font files.

The graphviz layout engines (dot, neato, etc) create layouts with nodes sized to enclose the text labels. This requires knowing the size of the text blocks, which in turn requires knowing the metrics of the font glyphs and their composition into words, taking into account wordspacing, kerning, hinting, etc. So the overall process is: font specification, then text layout, followed by Graphviz output (and final rendering on the target display or device, which may or may not be by a Graphviz tool).

A font is usually selected by family name (fontname) and other properties (see below: "Font selection"). Then fontconfig matches the request to a system font.

Note: in older versions of Graphviz, fontname was simply a file name. This required exact file name matching (with a little bit of helpful name mangling under the hood, e.g. translating Times-Roman to Times, or Helvetica to Arial on Windows systems (and yes, we know there is a difference). Under fontconfig, fontnames are family names, which fontconfig matches to the closest font it finds. This always "succeeds", but unfortunately produces surprising results if fontconfig's idea of "close" doesn't match yours. This can happen when you specify a custom (or just nonexistent) font, like Steve-North-Handwriting, and fontconfig silently falls back to something safe like a typewriter font.

Text layout is performed by pango, which accepts text and computes a layout with metrics that determine node sizes.

Though line drawing is provided by cairo for many output formats (and likely more in the future), for raster output formats, font rendering is passed though cairo to freetype. Freetype is also called if gd is used for drawing. (gd can also be requested explicitly, e.g. dot -Tpng:gd, or by default when Graphviz is built without cairo). Freetype provides antialiasing, hinting, kerning, and other low-level font features.

Font metrics are obtained from the fonts installed on the system running Graphviz. Results are guaranteed when Graphviz outputs raster formats, because freetype immediately renders the fonts into pixels. On the other hand, with path-based formats like Postscript (-Tps) and SVG (-Tsvg), final rendering may be done on a different platform altogether, with different font files installed. Clearly, Your Mileage May Vary. In the case of Postscript, the driver in Graphviz passes the expected metrics of the text block down to the renderer, and asks it to make a final stretch (or squeeze) to force the text to fit the metrics that were in effect at layout time. In Graphviz SVG, there is only a hope and a prayer that the SVG rendering program's fonts match the ones fontconfig and freetype used when Graphviz was run. (More about this later.)

Default fonts and PostScript fonts

The default font in graphviz is, and always has been, Times-Roman.

Graphviz has historically supported some "standard" Postscript fonts, initially, Times-Roman, Helvetica, Courier and Symbol. This list was later enlarged by Adobe to include 35 fonts, which are:

AvantGarde-Book AvantGarde-BookOblique AvantGarde-Demi
AvantGarde-DemiOblique Bookman-Demi Bookman-DemiItalic
Bookman-Light Bookman-LightItalic Courier Courier-Bold
Courier-BoldOblique Courier-Oblique Helvetica
Helvetica-Bold Helvetica-BoldOblique Helvetica-Narrow
Helvetica-Narrow-Bold Helvetica-Narrow-BoldOblique
Helvetica-Narrow-Oblique Helvetica-Oblique NewCenturySchlbk-Bold
NewCenturySchlbk-BoldItalic NewCenturySchlbk-Italic
NewCenturySchlbk-Roman Palatino-Bold Palatino-BoldItalic
Palatino-Italic Palatino-Roman Symbol Times-Bold Times-BoldItalic
Times-Italic Times-Roman ZapfChancery-MediumItalic ZapfDingbats

Unfortunately, fontconfig doesn't recognize PostScript-style font names directly, so Graphviz makes custom mappings from its list of PostScipt names into fontconfig family names for use in all cairo and gd based renderers. In -Tps output, these fonts are used without name translation.

Font selection

The fontname attribute in .gv graphs is a fontconfig style specification.

From: http://www.fontconfig.org/fontconfig-user.html

Fontconfig provides a textual representation for patterns that the library can both accept and generate. The representation is in three parts, first a family name list, second list of point sizes, and finally a list of additional properties:

<families>-<point sizes>:<name1>=<values1>:<name2>=<values2>...

Values in a list are separated with commas. The name needn't include either a family or point size; they can be elided. In addition, there are symbolic constants that simultaneously indicate both a name and a value. Here are some examples:

Name Meaning
Times-12 12 point Times Roman
Times-12:bold 12 point Times Bold
Courier:italic Courier Italic in the default size
Monospace:matrix=1 .1 0 1 The users preferred monospace font with artificial obliquing

Graphviz currently has a separate attribute for specififying fontsize.

TODO: We should allow the fontconfig style specification. Times-20 does not currently result in a 20pt font. This is probably because of special treatment of - for postscript font names.

TODO: We seem to have a bug with use of : in fontnames, probably because of special treatment for filenames in Windows. In fontnames, use <space> instead of : to separate values. -Nfontname="Courier:italic" doesn't produce an italic font in graphviz-2.16.1, but: -Nfontname="Courier italic" works, but -Nfontname="Monospace matrix=1 .1 0 1" doesn't.

Font management with fontconfig

How can I tell what fonts are available?

$ fc-list

How can I tell what fonts dot is using?

$ dot foo.gv -Tpng -o foo.png -v 2>&1 | grep font

How can I add a custom font?

In the current version of Graphviz with fontconfig, Cairo and Pango, this cannot be done by simply putting a file in the current directory or setting the DOTFONTPATH path variable. Your custom font must be explicitly installed by fontconfig tools.

For a single font, e.g., foo.ttf:

$ mkdir -p ~/.fonts
$ cp foo.ttf ~/.fonts/

One can run fc-cache to speed up the use of fontconfig.

$ fc-cache

For Windows users, one can go to the C:\windows\fonts folder and use File -> Install New Font from the pull-down menus to install the font.

For a new font directory, e.g., /Library/Fonts, add a new <dir> element

<dir>/Library/Fonts</dir>

to a .conf file. Note that the file must have a correct xml structure as specified by the fontconfig fonts.dtd. Possible choices for the .conf file are local.conf in the same directory as the system-wide fonts.conf file, or .fonts.conf in your home directory.

How can I ... font?

See: http://www.fontconfig.org/fontconfig-user.html

Can I specify a font by filename instead of by familyname?

Sorry, the answer is no. (The reason is that for this to work, Graphviz has to intercept the font lookup before fontconfig is called, and this can't be done when fonts are being looked up by Pango.)

Some versions of fontconfig appear to recognize pathnames and attempt to use that, but this isn't always the case.

How can I be sure that a specific font is selected?

Provide enough specification in the fontname, and test it with fc-match to ensure that your desired font is selected. (Note, this will not ensure that the same font is used in -Tps or -Tsvg renderings where we rely on the fonts available on the final printer or computer.)

Note the downside, as mentioned previously, is that Graphviz cannot do much to warn you when fontconfig didn't find a very good match, because fontconfig just cheerfully falls back to some standard font. It would be really nice if the fontconfig developers could provide a metric reflecting the quality of the font match in their API.

What about SVG fonts?

Graphviz has a native SVG driver that we wrote (which is the default), and cairo's SVG driver (which you get with -Tsvg:cairo).

Graphviz' native SVG driver generates Windows compliant names like Times New Roman or Arial by default. The names work in a lot of situations (like Firefox running on Windows), but are not guaranteed to be portable. If you set -Gfontnames=ps, you get Postscript names like Times-Roman. If you set -Gfontnames=svg you are guaranteed to get rock solid standards compliant SVG. The SVG standard says that the legal generic font names are Serif, Sans-Serif, and Monospace (plus Cursive and Fantasy which we don't use in Graphviz). We generate those names. The bad news is that various downstream renderers and editors may resolve the generic font names differently, so it's not quite clear how your SVG will look. Many W3C examples show how to use CSS (Cascading Style Sheets) to get around this problem by giving a list of font family names in order of lookup precedence, but some downstream processors (like the inkscape editor in Linux) don't implement CSS, so we're up a tree here.

The cairo SVG driver solves this in an effective though brute force way: it simply encodes embeds the needed fonts as lines and curves in the target SVG. For small examples, -Tsvg:cairo is about 10 times bigger than -Tsvg, but maybe it's worth it for correctness. The other problem is that such SVG is much much slower to render, no doubt because it bypasses any system font rendering services, and does it the old fashioned way.

What about Postscript fonts?

TODO: say something here. What about non-ASCII like Latin1? What about loading your own fonts via -L like in the old days with the weird outline font example.

"What if" issues for nonstandard Graphviz builds

The following only apply if you build your own version of Graphviz by configuring and compiling the source code to build your own custom executable. If you don't know what this means, it definitely does not mean you.

No freetype

When graphviz is built on systems without freetype, then only the gd renderer will be available for bitmap outputs, and the only available fonts are a small set of builtin bitmap fonts. The poor quality of these fonts will be evident, also, dot ... -v 2>&1 | grep font will say that the font is <internal>. This may actually be desirable for installing minimal graphviz programs on a server where fonts may not even be installed.

No fontconfig

If graphviz is built on systems without fontconfig (e.g. Redhat-7) then the fontname attribute will be interpreted as a font file name. The system directories will be searched for this, or the directories can be specified with the GDFONTPATH environment variable (or DOTFONTPATH for historical reasons). Graphviz will use gd and freetype to obtain metrics and render text. No pango/cairo renderers will be available without fontconfig support.

Disabling fontconfig

Pango/cairo depends on fontconfig, so to disable fontconfig you also have to disable pango/cairo. The easiest way to do this temporarily is to edit /usr/lib/graphviz/config and remove the entire libpango block.

Note that any changes to this file will be lost the next time graphviz is updated, or dot -c is run with installer privileges.

With pango disabled, graphviz will use gd which, even if it was built with fontconfig support, will still allow fontnames to be given as filenames.

You can also disable cairopango at build time with configure script options.

No gd

Cairopango works without gd. We are moving graphviz to the pango/cairo libraries, but gd still offers some features that are hard to replace, such as JPEGs, GIFs and paletted color bitmap outputs. However, font support is fully functional without gd so long as pango, cairo, fontconfig, freetype are available.

No pango/cairo

Without pango/cairo, some of the key renderers are only available with gd, which produces lower quality (but smaller) output.

Looking forward, we expect to depend more on pango for things like: line wrapping, multiple fonts per label, bidirectional text and other internationalization features.

No gd and no cairopango

This is basically the original Graphviz without any external fonts. It cannot render any raster formats, so it's mainly good for Postscript. It relies on a few internal font tables