prun [-IkOstv][-A account][-B base][-c cpus][-e mode][-i mode]
     [-m block|cyclic][-N nodes|all][-n procs][-o mode][-P project]
     [-p partition][-R request] program [args...]
prun [-dtv] -p root | -r
     [-B base][-H hostnamelist][-N nodes][-R request] program [args...]
prun -f procfile [-EItv][-e mode][-i mode][-o mode][-P project][-p partition]
prun -h

Run a parallel program
Flags:
	-A account	Specify account to use
	-B base	base node
	-c cpus	number of CPUs per process
	-d 		distribute rmsloader to /tmp on each node
	-E 		'explain' a procfile (only with -f option)
	-e mode		set stderr mode
	-f procfile	use procfile
	-H hostnamelist	run on listed hosts
	-h 		print this info
	-i mode		set stdin mode
	-I 		allocate CPUs immediately or fail
	-k 		forward Kerberos credentials
	-l tag		label stdout/stderr with a tag. The following
			tag characters have special meanings:
			   %g	the global rank of the process
			   %G	the number of processes
			   %h	the rank of the node
			   %H	the number of nodes
			   %l	the rank of the process on this node
			   %L	the number of processes on the node
			   %m	text base of the hostname
			   %n	numeric part of the hostname
			   %p	the pid of the process
			   %t	the current time (hh:mm:ss)
			   %@	the hostname of the node (same as "%m%-n")
			   %%	a single %
	-N nodes	number of nodes to allocate
	-n procs	number of processes to run
	-m block|cyclic	process to node mapping
	-o mode		set stdout mode
	-O 		allow resources to be overcommited
			ie. allow more processes than CPUs
	-P project	project to use
	-p partition	partition to use
	-R request	set resource request flag. request can be:
			    hwbcast=0|1
			    immediate=0|1
			    maxcpus=n
			    mincpus=n
			    pty=0|1
			    samecpus=0|1
			    rails=n
			    railmask=mask
	-r 		rsh mode (admin use only)
	-s 		print stats as job exits
	-t 		prefix output with process number (same as -l "%g ")
	-v 		enable reporting

Resources are allocated (and freed) for the job, unless previously
assigned using allocate. The default process distribution is block.

Multiple requests can be entered as a comma-separated list,
for example, 'prun -R hwbcast=1,immediate=1 ...'.

Stdio for a parallel program is controlled by three stdio modes one
each for stdin (-i), stdout (-o) and stderr (-e). These modes control
how input to prun is directed to the application processes and how
their output on stdout/stderr is processed by prun. The supported
stdio modes are:
	all	all processes send/receive
	none	no processes send/receive
	{rank}	only process {rank} can send/receive
	file	processes read/write from/to file.{rank}

The default modes are stdin=0, stdout=all, stderr=all, i.e. stdin is
directed to process 0 and output on stdout/stderr from all processes
is returned to prun.

'prun -p root ...' and 'prun -r ...' are equivalent in that -p root
implies rsh mode and -r implies -p root.  In both cases the partition,
will be root, rsh mode will be set and 1 process per node will run.
By default -N will be set to all nodes.

Procfile line format should be one of the following:
number			hostname[:path] executable arguments
lowrank;highrank;step	hostname[:path] executable arguments
number			index[:path] executable arguments
lowrank;highrank;step	index[:path] executable arguments