Hello List,

I think I have found the culprit. 
Tickets are distributed according to STN_stt (short term target), which in turn 
is basically the inverse of the the recorded (and decayed) usage. Since 
recorded usage may be zero, the denominator is actually max (SGE_MIN_USAGE * 
ltt), where for a simple share tree with just a "default" user and nothing 
else, ltt (long_term_entitlement) is just 1/#users. 
Unless I am mistaken, usage values (as used for the share tree) are never 
normalized/scaled, so we get actual CPU-seconds here, which can easily be of 
the order of 10**8. However, SGE_MIN_USAGE is #defined to 1.0, so while calling 
max() does prevent a real devision by zero, it will allow very large values of 
STN_stt.

That said, I still do not understand why the limiting effects of 
compensation_factor (i.e. stt := min (stt, compensation_factor * ltt) ) prevent 
this.

A.

On Dec 21, 2011, at 10:34 , Esztermann, Ansgar wrote:

> The problem at hand is this: sometimes, a user with no recorded usage will 
> submit jobs with very demanding resource requests, which consequently sit 
> around in qw for quite some time. They will also keep all available share 
> tickets to themselves, leaving none for other users' jobs. This is OK for 
> user A (no recorded usage means top priority), but users B and C might have 
> very different usage, and should therefore be assigned very different 
> priorities. However, there are no tickets left, so everyone but user A is 
> treated as an equal, upsetting fairness until those few special nodes that 
> can run A's jobs become free. Once A accumulates usage, share tickets are 
> gradually re-distributed to other users, and things work fine; until user X 
> becomes the new A, that is.

-- 
Ansgar Esztermann
DV-Systemadministration
Max-Planck-Institut für biophysikalische Chemie, Abteilung 105

Attachment: smime.p7s
Description: S/MIME cryptographic signature

_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users

Reply via email to