Ticket 287

The customer is having problems with SLURM v2.5.4 when using a decay halftime of 1 minute (60 sec).

In  the file plugins/priority/multifactor/priority_multifactor.c, at line 1029, the real_decay variable is evaluated using pow function:

real_decay = pow(decay_factor, (double)run_delta);

When the decay_factor is evaluated at line 980:
decay_factor = 1 - (0.693 / decay_hl);

For a decay halftime of 1 minute (decay_hl=60), 
decay_factor = (double)1 - (double)(0.693/(double)60)= 0.98845

For large value of run_delta ( .i.e. run_delta > 61000), the pow function returns a numerical result  out of range.

Then we get the error: problem applying decay in the slurmctld.log file.
This means that NO decay will be applied EVER.
The decay is VERY important in ensuring that proper Fairshare is applied to each job.
The customer proposes the following bug fix:
/ * In order to prevent this, we can find the largest run_delta possible. In other words:
  * if run_delta * log(decay_factor) <= DBL_MIN_10_EXP * log(10), we have a problem because run_delta is too large
  * and causes an underflow of the double. However, since decay_factor is always in the range [0,1], the log of it will always be negative, so we need to check if 
  * run_delta is larger than (DBL_MIN_10_EXP/log(decay_factor)).
  * Reminder, DBL_MIN_10_EXP (-337 on some systems) from float.h provides the exponent for the smallest possible value
  * of a base 10 exponent to produce a valid double. */

Replace line 1029
                                real_decay = pow(decay_factor, (double)run_delta);

by tle following lines of code:

                if(run_delta >= (DBL_MIN_10_EXP/log(decay_factor)))
                {
                        real_decay = pow(10, DBL_MIN_10_EXP); // This sets real_decay to the smallest representable double.     
                }
                else
                {
                        real_decay = pow(decay_factor, (double)run_delta);
                }