Add your comments in the Discussion Forum.
...
Cray says that things have changed and we should now be able to poll at a different interval. This way, the job's ALPS reservation being released can be discovered sooner, and the next job can use those resources sooner. The best way for PBS to handle this, will be to put the control in the PBS administrator's hands. 2 new mom tunables will allow the PBS administrator to individually adjust the base interval value, and the amount of potential jitter added to the total interval time. Total interval time is determined by adding the value for alps_release_wait_time + the randomly generated value based off alps_release_jitter. The minimum wait time interval is implementation dependent and may be different for different versions of ALPS and PBS Pro. The supplied value may be adjusted (rounded or truncated) based on the available resolution.
Tunable 1 - alps_release_wait_time
- This sets the base wait time in seconds to wait between ALPS release reservation requests
- It is a floating point number.
- Remember, there is an existing mom tunable alps_release_timeout which defaults to 600 seconds (10 min). That is the point at which PBS gives up trying to contact ALPS, and no more ALPS release reservation requests will be sent to ALPS.
- Set alps_release_wait_time in the mom_priv/config file
- If it is not set in the mom_priv/config file, the default value of alps_release_wait_time is 0.5 4 sec
Tunable 2 - alps_release_jitter
...
- Based on this value, PBS will randomly generate how much time to add as jitter. The jitter amount is randomly generated and can range from 0 to alps_release_jitter.
- alps_release_jitter is a floating point.
- Set alps_release_jitter in the mom_priv/config file
- If it is not set in the mom_priv/config file, the default value of of alps_release_jitter is 4 0.12 sec
...
...