search

ml-energy / zeus

Deep Learning Energy Measurement and Optimization
https://ml.energy/zeus
Apache License 2.0
200 stars 25 forks source link

`GlobalPowerLimitOptimizer` should support maximum slowdown factor #16

Closed jaywonchung closed 1 year ago

jaywonchung commented 1 year ago

Currently, Zeus solves the GPU time/energy multi-objective problem using linear scalarization (i.e., zeus_cost is a linear combination of time and energy). However, setting the weights for time and energy would be a challenge for most users.

In that sense, some may be more comfortable with specifying a maximum slowdown factor s, and GlobalPowerLimitOptimizer can choose the lowest power limit that satisfies that time constraint (i.e., epsilon-constraint method). This should also be very easy to do, as GlobalPowerLimitOptimizer already measures iteration time for each power limit.

jaywonchung commented 1 year ago

Completed by 07def08601305964a18574f3dfbb80e74891ec5a.