| |||||||||||||
LC-Asia: A big LITTLE MP updateLC-Asia: A big LITTLE MP updatePosted Mar 8, 2013 13:38 UTC (Fri) by k3ninho (subscriber, #50375)Parent article: LC-Asia: A big LITTLE MP update
>[T]he problem is likely to be solved by scaling the scheduler's load calculations by a constant value associated with each processor. Processes running on a CPU that is ten times faster than another will accumulate load ten times more quickly.
Faster isn't the goal here. Efficiency is the goal, so a scaling based on the relative efficiency of the 'big' or 'LITTLE' core should fit better with improving battery life.
For example, a core which runs at 600MHz and consumes 0.3 watts with a bogoMIPS of 1 might have a measurement of 200 mega-ops-per-joule and you'd compare it to a core that runs at 1600MHz and consumes 12 watts with a bogoMIPS of 4 (so has 533 1/3 mega-ops per joule). In a race-to-idle scenario, it's obvious where to schedule the work, but in a long-runnning work scenario there may be certain workloads which you wouldn't leave on the LITTLE core because they break into blocks that can win the race-to-idle on the big core. Calculating where that line is will depend on the cost of moving the work between cores, but for a first approximation: how much time do you get on the big core before you've used a second's worth of energy in the LITTLE core? (Assuming it's no cost to start and stop cores, I think that's 1/40 second, so work that fits in intervals of < 1/40 second is unexpectedly better off on the big core.
Note: if these numbers aren't quite right, please do tell. :-)
K3n.
(Log in to post comments)
LC-Asia: A big LITTLE MP update Posted Mar 10, 2013 18:25 UTC (Sun) by iabervon (subscriber, #722) [Link]
For efficiency, you want to direct the tasks that take the most cycles to the most efficient processors. Since you control which processor the task will run on in the future, your model should estimate how much energy it will take for each scheduling choice, so you're guessing the future cycle count for the task and multiplying by the energy cost for each processor; if you look at the past energy usage of the task, you're assuming that your past scheduling decisions are a good predictor of something in the future, but the point of the model is to make better decisions than your past ones were.
|
|||||||||||||