Floating-point error in demand

dirksammel commented 6 days ago

We see in our logs that the demand sometimes gets larger (or smaller) than the configured value due to some floating-point errors, for example:

Sep 11 14:46:36 arc3.bfg.uni-freiburg.de docker-COBalD-Tardis-atlprd[223200]: nemo_tardis_c40m100: 2024-09-11 12:46:36 demand = 1000.0000000000001 [demand=1000.0000000000001, supply=880.0, utilisation=0.78, allocation=0.78]

In consequence, a new drone is started and an old one begins to drain. If the draining takes some time and the floating-point issue happens again during this time, more drones than anticipated will run in parallel.

giffels commented 6 days ago

Hi @dirksammel, are you using the Standardiser/Limiter decorator to limit the demand or how do you set the anticipated number of drones running in parallel?

dirksammel commented 5 days ago

Yes, we're using the Limiter:

pipeline:
  - !RelativeSupplyController
    low_utilisation: 0.75
    high_allocation: 0.75
    high_scale: 1.5
    low_scale: 0.9
  - !Limiter
    minimum: 40
    maximum: 1000
  - !Logger
    name: 'changes'
  - !TardisPoolFactory
    configuration: 'tardis.yml'

MatterMiners / cobald

Floating-point error in demand #139