High Performance Computing: How Dotkit Works

How Dotkit Works

Background | Approach | Finding Packages | Dotkits | Site Nodes
Search Customization | Installations | Use-Usage

Dotkit Overview | Using Dotkit | Creating Dotkit Packages | Dotkit Commands | How Dotkit Works


If you read many existing shell startup files, you will observe that much of the conditional code—"if/else" or "case" and "switch" statements—uses the machine architecture or operating system type as its selector. A small section from a (much longer) .cshrc file reads:

if ($HostType == "Solaris") then
  set path = ( $path /usr/local/cvs/bin )
  set path = ( $path ~/bin/solaris )
else if ($HostType == "Linux") then
  setenv PGI /usr/local/pgi
  set path = ( $path /usr/local/intel/compiler60/ia32/bin )

The process that set the value of $HostType above is not shown, but it is quite useful to settle on a convention for this process. As a general statement, the more widely the convention is understood and utilized, the more useful it becomes. Dotkit assumes such a convention and uses an environment variable named SYS_TYPE to express the value of the current system type.

Continuing the example from above, a step in the direction of Dotkit is to divide your shell startup file into several files, one master, plus an additional file for each SYS_TYPE. The conditional code in the master file can then be reduced to one check:

if ( -e $HOME/.cshrc.$SYS_TYPE ) then
  source $HOME/.cshrc.$SYS_TYPE

Each of the approaches to writing dotfiles shown above is workable. (After all, we've been doing it this way for decades.) Some of the problems we have encountered might include:

  • If/else and case statements tend to get longer as time goes by. They rarely get shorter, even though some parts are no longer relevant.
  • The dotfiles require maintenance by (or on behalf of) each user whenever a new SYS_TYPE appears, or other things change network-wide. In part because they are complicated, dotfiles proliferate by sharing among users. However, sharing code between users usually requires each recipient to start a text editor and carry out a merge. The historical variations between your and my dotfiles may be trivial or they may not. Even if they are trivial, the code has to be read and understood (or not) over and over.
  • Sharing code across shell types (csh to bash, say) requires translation. It is vexing to move a carefully tuned set of startup files to a new environment. If/else constructions are brittle with respect to changes. For example, in the case above, for HostType "Linux," it doesn't matter that PGI was set prior to adding the Intel compiler to path. However, that ordering is imposed by the code, whether the code author intended it to matter or not.



As mentioned already, one of the assumptions Dotkit starts with is the notion of a received SYS_TYPE to guide basic decisions about the environment. The value of SYS_TYPE ("Solaris" or "Linux" in the above example) might be set outside Dotkit as part of the standard login procedure, or it can be set by Dotkit itself, using a short script (provided) that applies some simple heuristic logic to the output of uname(1). Determining an appropriate value for SYS_TYPE on a given machine, or deciding when two machines differ enough to merit different values of SYS_TYPE, can itself be a knotty problem, but we'll assume that has been done.

Dotkit breaks out the shell code needed to set up a given code system or other facility into a short package file or dotkit. It is similar to the Modules system in this respect.

One difference from Modules is that Dotkit integrates site location, shell, and SYS_TYPE information into the data structure of dotkits. It then uses polymorphism or, more simply, name overloading, to select a particular piece of code (dotkit) that satisfies the site, shell, USER, and SYS_TYPE requirements for that request. The data structure (and the problem space, not incidentally) is arranged such that the majority of package files have only one instance (e.g., if a given package does the same thing across multiple shells and SYS_TYPEs, only one—shell and systype-independent—instance of the dotkit file needs to be written).

How Dotkit Finds Package Files

Given a request such as use foo, Dotkit searches a list of directories for a file named foo.dk, taking the first such file found if there are multiple instances. (The search algorithm is slightly modified for the special case of use Super.)

The list can be depicted as a two-level tree structure, where each node in the upper level tree is itself the root of a smaller tree of subnodes. Every subtree has the same structure as all others.

The first node searched in the upper tree is $HOME/.kits, if it exists. The last node searched is $DK_ROOT, which is guaranteed to exist. In between, the entries in $DK_NODE, if they exist, (two are shown in the diagram below) are the intermediate nodes:

                            /      \
                           /        \
                    $HOME/.kits  $DK_NODE[2]
                Tree of Nodes Searched by Dotkit

As previously stated, each node in the tree above is the root of a subtree, laid out as shown below:

                            / \
                           /   \
                        *sh   $SYS_TYPE
                 Directory Layout of One Node

That is, the node directory itself, signified by (.), has subdirectories corresponding to each of the supported shell types—bash, csh, ksh, and tcsh—and it has subdirectories corresponding to each possible SYS_TYPE in the network at hand. Each *sh subdirectory may itself have SYS_TYPE subdirectories. Empty nodes and subdirectories do not need to be created or searched.

Dotkit traverses the logical tree under $DK_ROOT left to right, in post-order: Visit the children, then visit the node itself. This means that the $HOME/.kits node is visited first and will always win the race to find a given package file. The post-order traverse continues in the subnodes, so that within $HOME/.kits, the *sh/$SYS_TYPE directory is visited first, and $HOME/.kits/. (the node directory itself) is visited last.

Three characteristics of the Dotkit search process might be highlighted here:

  1. It dynamically adapts to the USER, shell, system type, and site.
  2. As discussed further below, it defines an ordering that can be utilized to inherit general characteristics into more-specific dotkits.
  3. The tree is under user control, and can be dynamically modified to suit project purposes.


The Scope of a Dotkit

In practice, the *sh subdirectories are usually empty, except in $DK_ROOT itself. (Dotkit is implemented as a set of shell scripts, so what better place to store each shell's code than $DK_ROOT/*sh?)

In a logical sense, the dotkits stored in the $DK_ROOT node describe software systems that are universal, available everywhere, independent of site or user or system type. For example, the package "bin.dk" is stored at $DK_ROOT/. It simply adds /bin to PATH. This action is useful on any UNIX or UNIX-like system.

The $DK_ROOT/$SYS_TYPE directory stores dotkits that should apply to every system of type $SYS_TYPE, wherever they are. For example, the dotkit $DK_ROOT/redhat_9_ia32/x11.dk adds /usr/X11R6/bin to PATH, and /usr/X11R6/man to MANPATH on every system running Red Hat Linux, version 9. There is also a $DK_ROOT/x11.dk file, which adds /usr/bin/X11 to PATH. The first variant is the one found if your current $SYS_TYPE happens to match "redhat_9_ia32"; otherwise, the second is used.

The $HOME/.kits/ node stores your personal dotkits. You can create your own set of new dotkits as needed, categorizing them according to SYS_TYPE or *sh if necessary, or you can override system dotkits by reusing their file name and creating your own new actions. No one but you has direct access to your personal dotkits. However, if you want to share, you can place your unmodified dotkit in a node that is visible to other Dotkit users, or they can copy your file, unchanged, to their own personal stash of kits.

The nodes listed in $DK_NODE, if any, are usually used for site-specific or project-specific dotkits. (A "site" is typically that portion of a network under a single administrative domain. Or it may simply be the set of machines that have common NFS access to $DK_ROOT.) These node(s) are searched after your personal node, but before the universal node.


More about Site Nodes

By convention, a site node is a directory whose name begins with the string "site=." It has the internal structure shown in the second diagram above, and it can be located anywhere in the file system. It is often convenient, but not required, to locate site nodes as physical subdirectories of $DK_ROOT.

A site node identifies a set of dotkits that belong to and describe the software available at that site. For example, site=olbullit identifies dotkits for my personal machine, which I manage as a stand-alone host. On that machine, my DK_NODE setting is simple:


Like olbullit, many or most sites need have only a single entry in DK_NODE.

As a more demanding example, in my working environment, I have accounts at two primary sites. One is a division-wide network (OAX), and the other is a Lab-wide network (OCF). They are managed separately, and have separate $DK_ROOT directories. Each has some software that is unique and some that is the same as the other network. It has therefore been useful to divide their dotkits into three sets: one for OAX alone, one for OCF alone, and one for dotkits that are the same at both sites. Consider the proper setting for DK_NODE on my local OAX network:


In words, this says that for any site-specific dotkit, we look first in the OAX node. (We're on the OAX network, so it takes precedence.) If there is no match, look next in the OAX+OCF node, where dotkits that work for either network can be found.

On the OCF network, the setting for DK_NODE becomes:


Site=ocf takes the place of site=oax. DK_NODE makes it easy to customize Dotkit to a site with minimal configuration effort.

Adding site nodes to DK_NODE is a task usually assigned to the Dotkit maintainer at that site. In addition to site nodes, individual users can add entries to DK_NODE that make project-specific or other collections of dotkits available to the project contributors.

If your network is large, it is also possible to divide up the contents of a site node into subsites, each containing a useful category of software available at your site, such as compilers, parallel libraries, etc.

A subsite is structured exactly like a site, conventionally named subsite=FOO. One dotkit in the primary site node is associated with each subsite, and modifies DK_NODE to add (use) or remove (unuse) the subsite.


Customizing the Search within Nodes

DK_NODE gives you and your site maintainers the ability to define the nodes on the Dotkit search tree (the first diagram shown above). If you or your site has special requirements, it is also possible to modify the tree internal to each node by changing the value of the environment variable $DK_SUBNODE, set in $DK_ROOT/$_dk_shell/.dk_init.

For example, if your network is organized as many small client workstations around a couple of large application servers, it might make sense to use the $HOSTNAME instead of or in addition to $SYS_TYPE as a selector inside each node. In this way, the view of available applications (dotkits) would vary according to whether you were logged into a client or an application server.

Small Dotkit installations

The discussion above describes how to configure Dotkit for a large multi-network situation. Dotkit is also well-suited to a small network or a single machine. It is distributed as a single small tar file. Installation consists of choosing a location for DK_ROOT, which may well be your HOME directory, and unpacking the tar file at that spot. At this point, Dotkit is fully usable without having to compile any code or create any dotkits of your own. The universal dotkits under $DK_ROOT are all available with no setup. Further customization for yourself or your site can then be done as needed.

Use-usage and Other Miscellany

When you type use with no arguments (or with just the -a or -v options or use -l with or without file arguments), Dotkit arranges to run a shell script named use-usage, located at $DK_ROOT/etc/. This script looks for a file named $DK_ROOT/etc/DK_MAINTAINER, whose contents is the name of the person who will maintain Dotkit for your site. Another optional file named dk-news can also be placed at $DK_ROOT/etc/ to provide site-wide notice of updates, additions, or other changes to Dotkit at your location.

If you prefer, you can write your own usage program by defining DK_USEUSAGE to the path name of an executable program or script.

Also at $DK_ROOT/etc/ are a script named envdiff, a helper program used by the "envdiff" dotkit, a makefile for various maintenance operations, the systype script used to compute values of $SYS_TYPE, a test/ subdirectory with several Dotkit regression tests, and the where, alias, and rep scripts referenced by the dk_where, dk_alias, and dk_rep commands.