If I understand the article properly, the partitioning of tasks into groups makes the difficult mathematical problem described in the recent piece of SCHED_DEADLINE soluble. In other words, we believe that a perfect admission criterion for the SCHED_DEADLINE algorithm exists, but we don't know how to compute it, so the compromise represented by IRMOS is a work-around. Or do we believe that the partitioning represented by IRMOS is intrinsically beneficial, as having fewer tasks in a group means that the total unpredictability per group is less than the total unpredictability over all tasks? Is the fact that partitioning into groups is beneficial is a natural consequence of the partitioning of the CPU into multiple cores?
From the nerd perspective, it's happy to see that these fascinating problems will be with us for a while!