Create a new "general users friendly" mode

[freedomtan]

• basic/easy/layman/quick/fast/lite-bench/rapid mode?
• things to change
  • cool-down / waiting time shorten to 1 min,
  • reduce number of samples (and the minimum running time for each benchmark item)
  • total time to be 5 ~ 10 mins (same constraint for both premium and mid-end devices)
  • for this mode, maybe statistic meaningful is not that important? or we can use mean or median instead of 90 percentile.

[Mostelk]

Per discussion with Scott, Quick Bench
Target for the whole run to be less than 10 minutes
It is OK to target for Premium Tier first

[freedomtan]

• make it possible to run on oldest Pixel phones (Pixel 6) for > 64 iterations, 128 iterations.
• to change parameters, we can replace the https://github.com/mlcommons/mobile_app_open/blob/master/flutter/assets/tasks.pbtxt

[Mostelk]

@freedomtan @quic-pnoorzad @mohitmundhragithub @Mostelk come up with some suggestion
~ min time
~wait time
~min samples
https://github.com/mlcommons/mobile_app_open/blob/master/flutter/assets/tasks.pbtxt

@anhappdev should we start making a parallel Fast Bench Mode for this that does not overwrite the main Performance and Submission modes

[anhappdev]

@anhappdev should we start making a parallel Fast Bench Mode for this that does not overwrite the main Performance and Submission modes

Ok. I will start working on it.

[mohitmundhragithub]

@freedomtan @quic-pnoorzad @mohitmundhragithub @Mostelk come up with some suggestion ~ min time ~wait time ~min samples https://github.com/mlcommons/mobile_app_open/blob/master/flutter/assets/tasks.pbtxt

@anhappdev should we start making a parallel Fast Bench Mode for this that does not overwrite the main Performance and Submission modes

I think we can have 2 different tasks.pbtxt files. 1 for the quick bench and another for submission / performance. Would be easier to maintain the settings that way.

• If we want to have a quick benchmark, then probably we don't want to stress the system?
• So both, minimum time and minimum samples could be reduced I think. May be 20secs of min time and 100 minimum samples?
• With less stress on the system, wait time could also be reduced from current 5 mins to may be 1-2 mins? But may need more experimentation.
• Also, with lesser number of samples, i think, we could use mean / median instead of 90th %ile for the quick bench mode?