Process Accounting on a Linux Cluster

[Greg Lindahl]

> What about people who use MPI/PVM to start new processes?  According to
> the PBS Administrator's guide, these processes are not tracked by PBS'
> accounting facilities or controlled by PBS in any other way (from what I
> am reading in the PBS admin guide, you place a job in a queue and when
> its time comes, the *same* job is executed on all of the requested
> nodes).

PBS starts up 1 process on the "mother" and the resources reported are for
that process. For an mpi job, that means that mpirun and possibly (depends
on mpirun) the 0-th rank process gets accurately reported.

However, on a space-shared machine, we charge based on # of nodes and wall
clock. And that's accurate.

Some folks have looked into using the tm_ routines in PBS as a means of
starting mpi jobs. That results in more accurate accounting of cpu time,
memory, etc.

-- g