This page tries to give a short commented overview on the options one would need to enable in MAUI configuration to enable fairshare with some reasonable algorithms.
The configuration file we use here at T2 Estonia is the following (the full file can be downloaded from here):
# define fairshare properties FSPOLICY DEDICATEDPS FSINTERVAL 24:00:00 FSDEPTH 4 FSDECAY 0.75 FSWEIGHT 1 FSUSERWEIGHT 250 FSGROUPWEIGHT 250
FSINTERVAL gives the interval which is taken into account as a single unit. In our configuration this is one day. FSDEPTH shows how many intervals are taken into account and FSDECAY shows the decay factor as we go further and further back in time. So today is taken in 100%, yesterday 75%, the day before is 56% etc.
FSWEIGHT shows the relative weight of fairshare compared to other weighting systems (like queue time etc). Usually all of them are at 1. FSUSERWEIGHT and FSGROUPWEIGHT say which are the respective weights depending on what user is considered and what group it has. The use of 250 here for both means that both are equally important and 250x more important than queue time for example. This could also be done by setting the user and group weights to 1 and the fsweight to 250. It's up to your taste :)
## GROUP SHARES GROUPCFG[cms] FSTARGET=79 GROUPCFG[balticgrid] FSTARGET=10 GROUPCFG[ops] FSTARGET=1 PRIORITY=500000 GROUPCFG[biit] FSTARGET=5 GROUPCFG[lhcb] FSTARGET=5
This section provides the different targets for different groups. So CMS gets 79% of the share, balticgrid 10%, biit 5%, LHCb 5% and ops ca 1%. It's always good to have the numbers add up to 100 so it's easier to understand the actual share. In this case OPS is also given absolute priority of 500000 to assure that whatever the share it gets the absolute priority.
QUEUETIMEWEIGHT 1 # Enable high priority of some users (SGM and PRD) USERCFG[cmssgm] PRIORITY=500000 USERCFG[cmsprd] FSTARGET=40 USERCFG[default] FSTARGET=1 USERCFG[lhcbsgm] PRIORITY=500000 USERCFG[balticgridsgm] PRIORITY=500000
This is where one sets first the priority from queuing which adds one priority point once per minute. It's good to try to find a balance between the queue time and fairshare so that everybody is happy, but basically it means only slight modification of either fsweight and queuetimeweight or the user and group weights.
The other part here is setting the user priorities and fairshare weights. The current configuration sets for SGM jobs absolute high priority and in addition sets cmsprd user to get 40% of user share while others should get around 1%. This basically has effect only on cmsprd as others get as they come and although they get penalized once they exceed the 1% target, cmsprd also gets penalized once it exceeds 40% and from there on the actual queue time dominates in the job selection for running.