This readme file describes the BEOSCYLD (Scyld Beowulf) port of PVM.
The purpose of this port is to implement a version of PVM for use with
the Scyld Beowulf cluster operating system. Since the slave nodes in a
Scyld Beowulf cluster do not by default contain a full installation,
the primary changes for this port were to provide the mechanisms for
getting jobs out to the slave nodes. There were two major problems to
solve. Number one, how do we get PVM slave daemons out to the slave
nodes? And number two, how can a slave daemon spawn a job when there
typically aren't any binaries on the slave node?
The first problem is typically solved using 'rsh' or 'ssh'. However
since Scyld slave nodes don't contain a full Linux installation, these
tools couldn't be used. The solution was to use 'bpsh'. The bproc shell
facility is used to migrate PVM slave daemons from the master node to
the appropriate slave nodes.
The solution to the second problem was implemented through the magic
of bproc. The traditional fork/exec sequence could not be used on a
slave node because in all likelihood the desired binary would not be
available locally. When a PVM slave daemon performs a task spawn, the
following sequence occurs: 1) fork is called to create a new process;
2) using bproc, the new process is moved back to the master node; 3)
the binary is located on the master by searching the DEFBINDIR path;
4) once located, bproc is again used to execmove the process back to
the slave node from which it came.
The master PVM daemon can be started on the master node using the PVM
console. You can use the console's add command to add slave nodes to
your virtual machine. The syntax for the command would be as follows:
"add .#", replacing the '#' character with the node's assigned number
in the cluster (as listed in beosetup). Alternately, you can start the
PVM console by specifying a hostfile name on the command line. This
file should contain the .# names of the slave nodes you want as part
of the virtual machine. As with standard PVM, this method automatically
spawns PVM slave daemons to the appropriate slave nodes in the cluster.
INSTALLATION & RUNTIME ISSUES:
This section describes the installation and runtime issues associated
with the RPM installation of the BEOSCYLD PVM port. If you're working
from a tarball, with the BEOSCYLD patches applied, you need to define
the following environment variables (as specified below) as explained
in the general PVM Readme file:
export PVM_RSH=`which bpsh`
export PVM_ROOT=<path where PVM resides>
The use of these variables and others is explained in more detail in
the pvm_intro man page. It's also probably a good idea to examine the
build time configuration file: $PVM_ROOT/conf/BEOSCYLD.def.
Assuming you're using the PVM tools from the RPM installation, the
following build / runtime information may prove useful.
1) PVM_ARCH was set to BEOSCYLD when building the binaries.
2) PVM_ROOT was set to /usr/share/pvm3 when building the binaries. You
can override this setting at runtime by creating a PVM_ROOT environment
3) RSHCOMMAND was set to /usr/bin/bpsh when building the binaries. You
can override this setting at runtime by creating a PVM_RSH environment
4) PVMDPATH was set to /usr/sbin/pvmd3 when building the binaries.
5) The default binary directory search path, DEFBINDIR, was set to the
following when building the binaries:
PVM_ROOT and PVM_ARCH are as defined above and $HOME will of course be
expanded appropriately at runtime.
6) All PVM tools and applications which make use of the PVM library, are
statically linked against /usr/lib/libbproc.a. This alleviates the need
to have the shared library version of libbproc out on the slave nodes to
execute PVM binaries.
This port was verified using the pvmtest33 "pvm test suite" tarball. All
tests were executed and all tests passed. The only item worth mentioning
is that if a PVM daemon running on a slave node crashes such that the
file /tmp/pvmd.uid does not get deleted, the daemon will not restart
on that slave node until the file is removed. This is not really a bug,
but a result of PVM's implementation. It's mentioned merely because it
is an easy thing to verify when things don't appear to be working
properly with respect to slave nodes. Note that the PVM debugger was
For questions or comments about this PVM port, contact the author by
sending an email to Joe Vitale at Scyld Computing Corporation,
firstname.lastname@example.org, or by contacting the Scyld Technical Support Team
directly at email@example.com.
THE AUTHOR MAKES NO REPRESENTATION ABOUT THE SUITABILITY OF THIS
SOFTWARE FOR ANY PURPOSE. IT IS PROVIDED "AS IS" WITHOUT EXPRESSED
OR IMPLIED WARRANTY.