condorports
-----------

This is a version of portbuild designed to exploit the features of the HTCondor
batch scheduler instead of a homegrown scheduler, in particular the abilities
of the Condor DAG manager.

Features:
- Heterogeneous clusters: support builds for multiple architectures using a
  mixed architecture cluster
- No shared file system required: distfiles and packages are transferred via
  condor file transfer to the build node, requiring only static read-only copies
  of the base system and the ports tree[s] on build nodes
- Flocking support: if you have multiple Condor clusters, jobs submitted on one
  can transparently "flock" to another, letting the package builds exploit the
  resources of geographically distributed computing clusters
- Many builds for many architectures, OS releases, and ports trees can
  continue in parallel
- Uses standard HPC software for scheduling with high performance, reliability,
  and security with only a light layer on top (155 lines of shell scripts)

Submit node prerequisites:
--------------------------
- sysutils/condor, configured to allow job submission
- Copies of the relevant ports trees, at the same paths as they are on the
  build nodes (in /buildshare/ports/XXX)
- python
- A lot of disk space (> 100 GB)

Build node prerequisites:
-------------------------
- sysutils/condor, configured using condor_config.local in this directory
  (adapted for the system in question)
- security/sudo, with sudoers file in this directory
- A lot of local disk space in /scratch
- A copy of all relevant worlds in /buildshare/releases/<releasenames>
- A copy of all relevant ports trees in /buildshare/ports/XXX
- NB: /buildshare can be read-only and node-private
- The prepbuildjail and reapbuildjail scripts installed in /pkgscripts

Using:
------

The basic procedure is:
cd condorports
./startbuild.sh 355755 /path/to/output amd64 10.0-RELEASE </scratch/horde>
<Eat one to several meals depending on cluster size>

This will configure and submit a Condor DAG for the ports tree in
/buildshare/ports/355755, the amd64 architecture, and the 10.0-RELEASE world,
using the common distfiles directory in /scratch/horde (on the submit node).
The last argument is optional. If unspecified, it will instead use a
build-private directory in /path/to/output.

As the build runs, log files and built packages will appear under
/path/to/output. The build queue can be inspected using condor_q -wide.

TODO:
----
Not all features are fully implemented yet. In particular:
- Need to advertise release and ports directory availability in the machine
  classad (via STARTD_CRON) so that jobs are scheduled only on hosts that
  have acquired the appropriate resources. This is especially important in
  the flocking case.