Page History
- 2015-Jun-18 17:14 tannenba
- 2015-Jun-18 17:10 tannenba
- 2015-Jun-18 17:08 tannenba
- 2015-Jun-18 16:49 johnkn
- 2015-Jun-18 15:56 johnkn
Since the purpose of the defrag daemon is to drain jobs on a p-slot so multi-core jobs can begin to match, it would be best to implement a policy where recently drained p-slots can insist on matching only multicore jobs for a period of time.
Unfortunately, there is no attribute that uniquely identifies a recently drained slot - but there are two candidate attributes that come close with some caveats.
EnteredCurrentState- resets to the current time when the slot leaves "Drained" state - but also when the STARTD is restarted or each time the negotiator matches the slot andNEGOTIATOR_INFORM_STARTDis true (which is the default prior to 8.3.7).ExpectedMachineGracefulDrainingCompletion- when this value is < time(), it represents the time at which the slot was fully drained - either because a drain state ran to completion or because the there were simply no jobs running on the slot or any children. when this value is > time() it represents the projected completion time of draining.
So a policy that won't match non-multicore jobs for 2 negotiation cycles on recently drained slots could be:
NEGOTIATOR_INFORM_STARTD = false IsntUnmatchedPSlot = PartitionableSlot=!=true || State=="Matched" IsMulticore = RequestCpus >= IfThenElse(Cpus<4,1,4) OnlyMulticoreInterval = $(NEGOTIATOR_INTERVAL:60)*2 StateTimer = (time() - EnteredCurrentState) OnlyMulticoreJobsAfterDrain = $(IsntUnmatchedPSlot) || $(IsMulticore) || $(StateTimer) > $INT(OnlyMulticoreInterval) START = $(START) && ( $(OnlyMulticoreJobsAfterDrain) )
The policy says that for 2 negotiation cycles after the STARTD starts up or leaves draining state, the slot should only match jobs that want at least 4 cores. Once the slot has less than 4 available Cpus remaining, it will match single-core jobs. We only want this portion of the START expression to be evaluated by the Negotiator, which is what the $(IsntUnmatchedPSlot) sub-expression does.
An alternate policy using ExpectedMachineGracefulDrainingCompletion would be:
IsMulticore = RequestCpus >= IfThenElse(Cpus<4,1,4) OnlyMulticoreInterval = $(NEGOTIATOR_INTERVAL:60)*2 DrainStateTimer = (time() - ExpectedMachineGracefulDrainingCompletion) OnlyMulticoreJobsAfterDrain = PartitionableSlot=!=true || $(IsMulticore) || $(DrainStateTimer) > $INT(OnlyMulticoreInterval) || $(DrainStateTimer) < 0 START = $(START) && ( $(OnlyMulticoreJobsAfterDrain) )
This policy will match single-core jobs only very slowly once the supply of multi-core jobs runs out, and it requires that the p-slot be fully drained in order to trigger the preference for multicore jobs. The first policy only requires that the drain state be canceled before the preference triggers.