[slurm-dev] --cpu_bind=none still binds jobs to CPUs

[Phil Sharfstein]

With TaskPlugin=task/affinity (using hwloc), when I launch a job with 
--cpu_bind=none, the job still gets launched explicitly bound to CPUs.  Looking 
at src/plugins/task/affinity/dist_tasks.c, it seems like this behavior is 
by-design in lllp_distribution().  However, I really wanted --cpu_bind=none to 
make the TaskPlugin work like task/none (and just let the OS take care of 
resource management).

I made the following change to lllp_distribution() in 
src/plugins/task/affinity/dist_tasks.c to keep jobs from launching bound to 
CPUs when --cpu_bind=none:

*** dist_tasks.c_orig   2012-07-11 09:44:48.298620129 -0700
--- dist_tasks.c        2012-07-11 09:46:14.318395335 -0700
***************
*** 304,309 ****
--- 304,313 ----
                                    CPU_BIND_LDMASK | CPU_BIND_LDRANK |
                                    CPU_BIND_LDMAP;
  
+       if (req->cpu_bind_type & CPU_BIND_NONE) {
+             return;
+         }
+ 
        if (req->cpu_bind_type & bind_mode) {
                /* Explicit step binding specified by user */
                char *avail_mask = _alloc_mask(req,

This change seems to work without causing any problems.
Two questions, though-

Does this change break anything that I might need later (i.e. MPI)?

What is the reasoning behind changing cpu_bind_type CPU_BIND_NONE to 
CPU_BIND_MASK and launching with explicit CPU binding in lllp_distribution()?

Thanks,
Phil