At a previous job we made an argument for moving away from cfs and to look at only full core allocation, often pinning with NUMA. The speed up was noticeable, s

At a previous job we made an argument for moving away from cfs and to look at only full core allocation, often pinning with NUMA. The speed up was noticeable, since it removed the cfs overhead and memory access was now local. We then got stuck in discussions around partial core allocation. We didn’t have that many jobs configured to use less than a full core, but it did impact our container packing.