A problem about AI or MI

[ChengjunJia]

Hi, Yuliang. I have a question about the maxStage parameters. From the public HPCC key results, I found

Besides, HPCC’s maxStage=0, which is the same as the simulation setting in our paper. The maxStage=5 in the HPCC paper is due to a typo for the simulation.

while in the paper, it is claimed that

maxStage controls a simple tradeoff between steady-state stability and the speed to reclaim free bandwidth.

From the theoretical model and ns3 simulation, I believe that maxStage=0 (which means that MI is directly used) is indeed better. But I notice that Google's Swift also used AI and removed the MI stage of Timely. I wonder if MI is too radical, causing some other problems, in the testbed or in practice. We have no testbed experience, but I guess that there would be many jitters and uncertainty for a large-scale testbed, which makes MI not suitable? Or MI is indeed better and it is not adopted just due to inertia? Can you share your opinion or some experiences?

Thanks a lot Best wishes

[liyuliang001]

Hi, You are right. In theory maxStage=0 is best. The reason we set maxStage in testbed was due to some HW bug I believe.

HPCC's goal is precise control, that implies some quick and large movement (otherwise it cannot be precise :)). And you are right, not using MI may be more stable under unexpected cases, but I believe in most (if not all) those cases, problem's root cause is not MI, but MI just helps us find out other bugs that is the culprit. Not using MI may mask such bugs.