[Beowulf] scheduler policy design

[Tim Cutts]

On 25 Apr 2007, at 4:12 pm, Bogdan Costescu wrote:

> ... or would supply some meaningless values...
>
> For example, if the queue is limited to 24h maximum runtime and  
> they would set 24h as their requirement, thinking that it would  
> certainly cover what they want and the job would not run anyway  
> longer than that, as it would be killed by the scheduler. Even when  
> they run sets of identical jobs, for which they could very well  
> predict the needed runtime after the first 1 or a few jobs have  
> finished... And the most painful part is that quite of lot of these  
> users also feel very strongly that they are doing nothing wrong:  
> they set a value which is less-or-equal than the maximum allowed,  
> "so it's got to be right, no ?"
> (obviously, if the queue limit would be set to something different  
> (say 12h) they would simply adjust to the new value ;-)).

With our queueing system that forces the job into a lower priority  
queue.  If they want their job to run with higher priority, they have  
to be more accurate.  So indeed, they're not doing anything wrong,  
but they're pushing themselves down the priority list in favour of  
people who are being a bit smarter.

For sites which charge for machine time, of course, that's a further  
selection pressure.  Many systems I've spoken with people about  
charge you for the time you ask for, not what you use (and obviously  
don't let you use more than what you asked for); so the above  
approach by users wastes a lot of their money.

Tim