On Mon, Jul 3, 2023 at 9:13 AM David Rientjes <rientjes@xxxxxxxxxx> wrote: > > Thanks very much for looking at this, Jay! > > My colleague, Binder, has also been looking at opportunities to optimize > memory usage when using SLUB. We're preparing to deprecate SLAB > internally and shift toward SLUB since SLAB is scheduled for removal after > the next LTS kernel. > > Binder, do you have an evaluation with this patch similar to what Jay did? > > Also, tangentially: we are looking at other opportunities for reduction in > memory overhead when using SLUB. If you or anybody else are interested in > being involved in a working group with this shared goal, please let me > know. We could brainstorm, collaborate, and share data. I'm also interested in reducing SLUB memory overhead! I have some rough ideas, which should be evaluated further: 1. Lengthen or shrink number of cached objects per CPU based on list_lock contention. 2. Modify SLUB to enable linking objects from different slabs into the CPU freelist. Do you have any opinions, or are there any approaches you are already examining? -- Hyeonggon
