On Mon, 10 May 2004, Ian Stokes-Rees wrote:
> How do I specify a "normalised" time for a job? Is this handled
> automatically by LCG? What I mean is, if I have a job which takes, say,
> 2 hours on a P4 1.6GHz, it *may* take about 1.5 hours on a P4 2.4 GHz
> machine and 3 hours on a 1 GHz machine.
Nothing is handled automatically. The JDL has to specify requirements
explicitly, the broker has no knowledge of what different things mean.
The sites are supposed to publish a specint and specfloat rating, and you
can use those to scale your requested CPU time. However, there is a
potential problem with heterogeneous clusters given that the schema
only supports one value, which is probably likely to be set at something
like the minimum for the cluster.
> Also, how do I differentiate between CPU time requirements and wall
> clock time requirements?
The main purpose of the wallclock time limit is to kill jobs which get
stuck in a state which isn't consuming any cpu, so if your job is largely
cpu-bound you can probably ignore the wallclock limit. For I/O bound jobs
you would probably need to set requirements on both, although the
scheduling model we're using at the moment (generally one job per cpu)
isn't really suited for such jobs anyway.
Stephen
|